- Email: [email protected]
Soft sensor for real-time cement fineness estimation
ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Contents lists available at ScienceDirect
ISA Transactions journal homepage: www.elsevier.com/locate/isatrans
Soft sensor for real-time cement fineness estimation Darko Stanišić n, Nikola Jorgovanović, Nikola Popov, Velimir Čongradac Faculty of Technical Sciences, University of Novi Sad, Novi Sad, Serbia
art ic l e i nf o
a b s t r a c t
Article history: Received 10 August 2013 Received in revised form 14 August 2014 Accepted 27 September 2014 This paper was recommended for publication by Dr. Ahmad B. Rad.
This paper describes the design and implementation of soft sensors to estimate cement fineness. Soft sensors are mathematical models that use available data to provide real-time information on process variables when the information, for whatever reason, is not available by direct measurement. In this application, soft sensors are used to provide information on process variable normally provided by offline laboratory tests performed at large time intervals. Cement fineness is one of the crucial parameters that define the quality of produced cement. Providing real-time information on cement fineness using soft sensors can overcome limitations and problems that originate from a lack of information between two laboratory tests. The model inputs were selected from candidate process variables using an information theoretic approach. Models based on multi-layer perceptrons were developed, and their ability to estimate cement fineness of laboratory samples was analyzed. Models that had the best performance, and capacity to adopt changes in the cement grinding circuit were selected to implement soft sensors. Soft sensors were tested using data from a continuous cement production to demonstrate their use in real-time fineness estimation. Their performance was highly satisfactory, and the sensors proved to be capable of providing valuable information on cement grinding circuit performance. After successful off-line tests, soft sensors were implemented and installed in the control room of a cement factory. Results on the site confirm results obtained by tests conducted during soft sensor development. & 2014 ISA. Published by Elsevier Ltd. All rights reserved.
Keywords: Soft sensors Cement fineness Estimation Neural-network models Product quality
1. Introduction Fineness is one of the most important physical properties of cement. Cement fineness tests are performed in laboratory on samples taken from a grinding circuit. The testing procedure is defined by appropriate standards and is executed at uniform time intervals, typically every two hours [1]. There are numerous limitations due to such important information being available only periodically at extremely long time intervals. Any algorithm that automatically controls a grinding circuit, instead of cement fineness, must be related to another, alternative and easily measurable system parameter, which makes direct control of the most important system parameter impossible. This situation means it is always necessary to guide the system toward operating points that should provide, at a high level of confidence, the desired fineness, usually resulting in fineness far above the limit defined for the product. In this way, much more energy is consumed during the cement grinding process than necessary. The problem of excess energy consumption during the cement grinding process is clear when accounting for the amount of electrical energy consumed in cement n
Corresponding author. Tel.: þ 381 21 485 2452; fax: þ 381 21 458 873. E-mail addresses: [email protected] (D. Stanišić), [email protected] (N. Jorgovanović), [email protected] (N. Popov), [email protected] (V. Čongradac).
production, which is approximately 110 kW h/t, of which, approximately 40% is consumed for cement grinding [2,3]. Since there is no available real-time cement fineness measurement, the focus in energy consumption reduction in cement grinding circuit is directed to improvements in grinding technology [3]. Improved grinding media in traditional ball mills, new grinding mill designs, combined grinding systems [4], clinker pre-crushing [2] and use of highefficiency separators are energy saving measures commonly undertaken by cement manufacturers. Providing real-time information of cement fineness would offer new, complementary opportunities for reduction of energy consumption. In addition, other equally important problems originate from the lack of information about cement fineness. Because no information on cement fineness between two samples and related laboratory tests is available, when a laboratory test shows deviation from desired values, it is reasonable to assume that there is a significant amount of poor-quality cement produced between the two samples. That deviation from desired values can be significant and indicate that the process is shifted far away from normal operating conditions; in this situation, it is typically necessary to take radical corrective actions to guide the process to the desired performance. In addition, there is a possibility that even when the cement fineness of two consecutive samples are adequate, in several situations, a large portion of cement produced between the two samples may be of completely different, inadequate, quality. All
http://dx.doi.org/10.1016/j.isatra.2014.09.019 0019-0578/& 2014 ISA. Published by Elsevier Ltd. All rights reserved.
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
2
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
these problems lead to the conclusion that finding a method that provides information on cement fineness in real time can significantly and positively influence the two main indicators of cement grinding circuit performance, energy consumption and quality of cement. The objective of this paper is to obtain information on cement fineness by determining its correlation with measurable process variables. In this way, reliable real-time estimates of cement fineness can be obtained using a mathematical model to derive the fineness from available data. Such systems are typically called soft sensors or virtual sensors. Soft sensors are used in different types of applications. Soft sensors for measuring acid gases emission in sulfur recovery unit of oil refinery presented in [5] and particle size estimation in wet grinding circuits presented in [6] are applications where soft sensors are used to provide reliable process information when hardware sensors are removed for maintenance. When one hardware sensor is time shared by several parallel processes, soft sensors can be used during time intervals in which hardware sensor is unavailable [7]. Soft sensors working in parallel with hardware sensors can be used for fault detection and can also make it possible for a system in such situations to continue process control using estimated values from a soft sensor to replace measurements from a faulty hardware sensor. In [8] an example of such application in winding machine is presented. Situations when measurements from hardware sensors are not suitable for use in control algorithms because they have a large time delay due to their position in the production process or because of their time-consuming operation, can be solved by soft sensors [9–11]. In [9] this type of solution is used for monitoring of product quality in oil refinery, while in [10,11] chemical process is controlled by controlling the product composition estimated by soft sensor. In a production process where product quality is controlled with only a few sample products and the quality of every product cannot be assured, soft sensors provide a way to obtain information on the quality of each product [12] and assist quality management and process control [13]. Soft sensor applications similar to the cement fineness application address situations where important process variables are available only from off-line laboratory tests and are therefore unavailable for use in control algorithms. This is the case in fermentation process as is presented in [14,15], also this approach can be used for emission prediction and control for a gasoline engine [16]. In [17,18] soft sensors are utilized to provide information of process variables used in steam quality control since online measurement of process variables of interest is not always accurate or reliable and is therefore supplemented by lab analysis taken infrequently in a manual manner. Results obtained from soft sensors for online estimation of clinker quality [19] show that important information in cement production process can be provided by the use of soft sensors. Soft sensors have already demonstrated their ability to provide solutions in grinding circuits [6,7], where they were used to provide information on the particle size distribution in wet grinding circuits. The problem described in this paper has one significant difference compared with the aforementioned paper on wet grinding circuits. Whereas soft sensors used in the wet grinding circuits were developed using information from a hardware sensor, which in normal operation, provides real-time measurements of particle sizes, a soft sensor for cement fineness must be developed based on information provided from off-line laboratory tests performed every two hours. Soft sensors are typically developed using first-principle models when they are available and in a suitable form [20]; however, more often, gray or black box models are used. In the case of a cement grinding circuit, due to the complexity of the system, particularly the grinding process, a complete first-principle model suitable for soft sensor development is not available. Attempts to
develop gray box model of this process lead to exceptionally complex models whose parameters could only be determined by delicate and extremely time-consuming experiments with relatively large uncertainties associated with several parameters [21]. In addition, though gray box models based on grinding laws agree with laboratory experiments, their transposition to industrial mills gives poor results. However, for purposes of monitoring and system performance analysis, a great amount of historical data for various process variables is collected, which makes black box model parameter identification convenient. It is necessary to select black box model inputs from large number of candidate process variables. In this paper, selection of model inputs was based on an information theoretic approach. An algorithm called the information theoretic input variable subset selection (ITSS) [22] is used. This algorithm does not depend on any nonlinear function approximation methods; it solely depends on the measured system's input–output data. Information theory is used to analyze interdependency between the process output and inputs to determine, from the entire input data, a subset of input variables that contains most of the information necessary to predict the output. Theoretical knowledge regarding the process that is modeled should also be incorporated when using ITSS to ensure that all important variables to the model are included in the subset [22,23]. In this paper the soft sensor is proposed to estimate cement fineness and provide real-time information on a process variable available only from off-line laboratory tests. The design and implementation of the soft sensor is presented. Data collected during regular production in cement factory Lafarge BFC1 in Beočin, Serbia is used for soft sensor development. Soft sensor was developed using a black box model based on a multi-layer perceptron neural network. Model is trained using the data obtained by standard offline laboratory tests conducted every two hours. Their ability to provide real-time fineness estimation on every minute is tested. Soft sensors are implemented and installed in the cement factory Lafarge in Beočin, Serbia. The remainder of this paper is structured as follows. Section 2 provides a schematic description of the cement grinding circuit. In Section 3, analysis of the collected data is presented together with algorithms and procedures used for models structure selection. In Section 4, experimental results are presented and analyzed, and Section 5 presents the conclusions of this work.
2. Description of the grinding circuit The cement grinding circuit is the final stage in the cement production process. In this stage of production, clinker produced in the kiln together with additives that are used to attain the desired properties of the produced cement are grinded in the mill into a fine powder (cement). The material is grinded in the rotating ball mill where grinding is performed by steel balls. During grinding, the clinker particles should be reduced in size by a factor between 1000 and 10,000, and it has been proven to be more efficient to apply different grinding forces to different size ranges of the material, which is achieved by constructing mills with several grinding chambers, typically two, with different properties, such as chamber length, size of the ball charge and shape of the liners along the chamber walls. Fig. 1 shows a simplified scheme of the cement grinding circuit that indicates process variables of interest in this paper. Material flow is indicated with solid lines, whereas dotted lines represent air flow in the circuit. Clinker and additives are transported from 1
www.lafarge.rs
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
3
P T A
Product
F
P Separator
S Elevator
Clinker
Additive ... Additive 1 n
F
L F
F
I
Rejects L P
F
T
T
Ball Mill I Fig. 1. Simplified scheme of cement grinding circuit.
their bunkers via conveyor belts and are fed into the ball mill, which consists of two chambers. Flow rate of every component fed into ball mill is measured and controlled by dosing system to comply with defined recipes. The level of material in mill chambers is measured by the sensors based on vibration or noise measurement. Two motors, whose currents are measured and monitored, are turning the mill. Material continuously flows through the mill, and the grinded particles that exit the mill are transported by air slide to an elevator. A fan placed at the mill exit provides air ventilation necessary to prevent over-heating of the mill. Temperature of the material exiting the mill is measured by the appropriate sensor, as well as the air temperature and air pressure in the ventilation system on the mill output. The elevator is used to transport grinded material to a high-efficiency separator and the elevator current is measured to provide the information on the amount of material exiting the mill. The material coming from the elevator is dropped into a blowing air stream inside the separator where rotor provides centrifugal force that is used to separate coarse particles. The fine particles are removed from the separator via the air stream and transported to cyclones along with the flow of hot gas heated by the material. Inside the cyclones, the solid material is separated from the gas and taken to storage silos as the final product. After exiting the cyclones, the hot air is recirculated through a fan into the system and back to the separator. Separated coarse particles fall down to the bottom of the separator where they are discharged into the air slide. The coarse particles form rejects flow that is returned to the mill's inlet for an additional grinding cycle along with fresh feed. Several process variables related to the separator operation are measured. Those variables are air flow rate on separator inlet, provided by differential pressure measurement, separator rotor speed, air pressure on the separator inlet and air pressure and air temperature on the separator outlet.
3. Data analysis and model structure Because separating coarse particles from fine particles, which makes the final product, is performed in the separator by interacting drag forces provided by the air stream that enters the separator and centrifugal forces provided by rotor rotation, it is clear that air flow into the separator inlet and rotor speed are two essential variables that must be chosen as model inputs. Other variables that may provide additional insight about the material that enters the separator and the operating conditions of the separator should also be chosen. Material that enters the separator
comes from a ball mill, and its properties depend on the properties of the material that enters the mill and the grinding performances of the mill, which means that process variables from all parts of the grinding circuit are potential model inputs. Selecting appropriate additional variables for model inputs is a complex task. There are numerous measured variables that provide information on system performance; however, separating variables that should be used as model inputs from irrelevant variables that can be detrimental to model performance must be conducted carefully. In this paper, this is done in two steps. In the first step, the set of candidate variables is chosen based on an analysis of the grinding system functionality. All candidate variables are listed in Table 1. From this set of candidate variables, in the second step, inputs are selected using an information theoretic approach. 3.1. Data description and preprocessing The laboratory fineness test is done by sieving the product sample through a 35 μm sieve. The percentage of material that remains on the sieve defines the fineness of the product tested. Samples for the laboratory fineness tests are taken every two hours by an automatic sampler. The sampling procedure lasts 36 min, where 6 samples are taken. Every sample is taken during 1 min with a pause of 6 min between samples. Samples are then mixed together and fineness of joint sample is determined. This implies that the laboratory fineness is determined as an average fineness of the material that exited the separator during the 36 min of the sampling duration time. Therefore, the model inputs should be processed accordingly. The sampling time of the measured process variables is 1 min; however, their mean values over the fineness sampling duration time are used as model inputs. It is clear that all process variables should be delayed for the amount of time necessary for material to come from the point where the process variable is measured to the separator. It takes less than one minute for material to come from the mill output to the separator, and it is clear that process variables related to the mill output, as those linked with the separator can be processed without delay. The time necessary for material to come from the mill input to the mill output is significant, and variables related to the mill input should be delayed. An analysis of the rejects flow rate in situations when material is fed into empty mill showed that rejects are detected after 24–27 min from the start of the process. In normal operating conditions, mill chambers have significantly more material, which could additionally slow the flow of new material from the mill input. To account for any additional delay that could appear during normal operating conditions, a 30 min
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
4
denoted by mx, is done by
Table 1 Candidate input variables and their introduced delay. Variable description
Notation
Units
Introduced delay (min)
Clinker flow rate Gypsum flow rate Slag flow rate Limestone flow rate Fly ash flow rate Fill level of mill chamber 1 Fill level of mill chamber 2 Air pressure on mill output Air temperature on mill output Mill motor current Cement temperature on mill output Elevator current Separator rotor speed Differential pressure at separator inlet Air pressure at separator inlet Air pressure at separator outlet Temperature at separator outlet Rejects flow rate on separator Rejects flow rate on mill input
Clink_flw Gyps_flw Slag_flw Lims_flw FAsh_flw Mill_lvl1 Mill_lvl2 Mill_press Mill_airT Mill_curr Mill_cemT Elev_curr Sep_spd Sep_dpress Sep_pressIn Sep_pressOut Sep_outT Rtrn_sep Rtrn_mill
% % % % % % % bar 1C A 1C A rpm bar bar bar 1C t/h %
30 30 30 30 30 30 0 0 0 0–30 0 0 0 0 0 0 0 0 30
di ¼
xi mx
ð1Þ
σx
where σx is standard deviation of x. If data follows normal distribution, the probability that di 43 is 0.27% and in that case the sample xi can be considered as an outlier. That means that every input variable x needs to be in interval [mx 3σx, mx þ3σx]. This interval can be used as a starting point and in further analysis with plant experts, it can be modified to include working conditions that may occur rarely and are therefore excluded by “3σ edit rule”. Also the data for every variable should be analyzed by visual inspection to check how much it concurs with normal distribution and, if necessary, in cooperation with plant experts the modification of intervals obtained by “3σ edit rule” can be performed. All variables then need to be normalized, usually in [ 1, 1] interval. After all three types of outliers are removed from the initial data set with the presented procedure, the data set which contains 980 samples of laboratory tests with corresponding measured process variables is obtained. This data set is used during the soft sensor development presented in the paper. 3.2. Input selection
delay was defined for all process variables related to the mill input. The introduced delays on measured signal values are presented in Table 1. The mill motor current was not processed in the same way as the other variables. The mill motor current provides information on the mill operating conditions during a time period of approximately 30 min necessary for material to pass from the mill input to mill output. Because of this, in this paper, for every 1 min sample of other process variables, a mean value for the previous 30 min mill current samples is computed as information on the mill current for the material that is exiting the mill at that moment. For mill current values obtained in this way, a mean value over the fineness sampling duration time is then calculated in the same way that it is performed for the other process variables. The measured flow rates of the fresh material entering the mill are used to calculate the percentage of every material in the fresh mill feed. This percentage is thus considered a potential model input. The data collected from a cement grinding circuit by an acquisition system during 4 months of normal operation was used for the soft sensor development. The data were analyzed, and three types of outliers were identified and removed: (1) Data collected during production of products whose fineness is not measured by the standard test. (2) Data from system operation during mill startup. (3) Outliers related to faulty measurements by the sensors.
From the candidate variables listed in Table 1, model inputs can then be selected using the ITSS algorithm. To implement the ITSS algorithm on continuous variables, it is necessary to discretize and divide the domain space of every variable into a finite number of intervals. In this way, vector of input variables x can take one of M discrete values x1, x2, …, xM and dependent variable y can take one of S discrete values y1, y2, …, yS. If N is number of available samples in data set and Ni is number of occurrences of x ¼ xi in the data set, then probability that x can take value xi is pi ¼ pðx ¼ xi Þ ¼
Ni N
The entropy of information presented by variable set x is defined by M
HðxÞ ¼ ∑ pi lnðpi Þ
ð3Þ
i¼1
The probability that x will take value of xi and y will take value of yj is pij ¼ pðx ¼ xi ; y ¼ yj Þ ¼
N ij N
ð4Þ
where Nij is number of joint occurrences of xi and yj in the data set. The conditional entropy of variable y when variables x are given, expresses the information in dependent variable y that remains unexplained when input variables x are known and is defined by HðyjxÞ ¼ ∑pij ln
The first type of outliers is easy to detect, since the information on type of the product is stored in the acquisition system. The second type of outliers can be identified from information on the level of material in mill chambers and rejects flow rate, since mill startup is considered over and system is in normal operation when there is certain amount of material in the mill and rejects flow rate reaches certain level. The minimal amount of material in the mill and minimal amount of rejects during normal operation are technological parameters and can be obtained from plant experts. After removing the first two types of outliers, the remaining data can be processed to identify and remove all other outliers usually related to faulty measurements. It can be done by using “3σ edit rule” [24]. For this rule, normalization of distance between one sample xi and mean value of all available samples of variable x,
ð2Þ
i;j
pij pi
ð5Þ
The amount of information about y contained in variables x is shown by an asymmetric dependency coefficient that measures the dependency of output y on variables in x UðyjxÞ ¼
HðyÞ HðyjxÞ HðyÞ
ð6Þ
where H(y) is the entropy calculated for the dependent variable y. When there are no useful information about y contained in x, UðyjxÞ is 0. If information contained in x completely defines y and can be used to predict y exactly, UðyjxÞ is 1. Variable selection is performed by a forward selection procedure, where in each step, a new subset is generated by adding one input variable to the current subset. The new input variable added to x is the variable
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
f1 x1
f1
x2
yˆ f2
. . .
f1 . . .
xn
f1 Input layer
Hidden layer
Output layer
Fig. 2. Multi-layer perceptron neural network.
that gives the maximum value of UðyjxÞ. Input variables are added until UðyjxÞ value becomes close to 1. It is important not to divide domain space in too many intervals because random noise may be considered as information. Also, not dividing domain space in enough intervals can result in loss of accuracy. Analysis of collected data shows that measurement noise for all candidate variables is below 5% of domain space. For this reason the domain space of every input variable and output cement fineness was divided into 20 intervals.
5
developed one candidate model from the available data set to be used as a soft sensor. For this approach, from the available data set containing 980 samples, the first 550 samples were used for training and validation, where 70% of the samples were used for training and 30% for validation. Selecting the training and validation samples was performed randomly. The remaining 430 samples were used for testing. The second approach, referred in this paper as the retrained model approach, began with one candidate model defined using the first approach; however, every time a new sample is taken and laboratory results are available, the model is updated. The neural network is then retrained using the newest 550 data samples for training and validation. Because new data samples are obtained every two hours, on-line retraining of the neural network between two samples can be achieved without any problems, and using this approach does not have any negative influence on soft sensor implementation. Motivation for this approach was derived from the fact that cement grinding circuit characteristics change with time (changes in material, wear of grinding balls, etc.), and a soft sensor for cement fineness estimation must be able to adopt changes in the cement grinding circuit. In this way, the soft sensor is defined by a moving window of new data that includes only the current characteristics of the cement grinding circuit. This approach is based on a sliding-window policy commonly used in identification of time varying systems [27,28].
3.3. Neural network models In this paper, all models considered as candidates for soft sensors are static nonlinear models based on a multi-layer perceptron neural network shown in Fig. 2. For input vector x, model output is _ y ¼ f ðxÞ ð7Þ The neural network models had one hidden layer that used a sigmoidal activation function f 1 ðaÞ ¼
2 1 1 þ e 2a
ð8Þ
and an output layer that used a linear activation function. f 2 ðaÞ ¼ a
ð9Þ
The models were trained by the Levenberg–Marquardt back propagation algorithm [25,26] with early stopping. For single output neural network, model parameters are obtained by minimizing the least squares cost function N k J ¼ ∑ ðyk _ y Þ2
ð10Þ
k¼1
where yk is the desired target output of kth training sample and k _ y is the model output from the kth training sample. If vector of model parameters is denoted by ω, with Levenberg–Marquardt algorithm parameters are updated, at iteration i, by h i1 ~ ωði 1ÞÞ þ μ I ωðiÞ ¼ ωði 1Þ Hð ∇Jðωði 1ÞÞ ð11Þ i N
~ ωðiÞÞ ¼ ∑ Hð
k¼1
k y ek ¼ yk _
∂ek ∂ωðiÞ
∂ek ∂ωðiÞ
3.4. Soft sensor implementation After evaluating the ability of candidate models to estimate cement fineness of laboratory samples, the models with the best performances were selected as soft sensor candidates. Because the purpose of a soft sensor is to provide real-time estimates of cement fineness between two laboratory samples, the performance of the candidate models were tested on a data set that contains 98 h of continuous cement production. Fineness is determined every minute to demonstrate the soft sensor's capability to provide real-time information required to monitor the cement grinding circuit performance. In this paper, candidate neural network models for soft sensors were trained using a training and validation data set consisting of laboratory fineness measurements and the corresponding measured process variables processed as described in Section 3.1. However, input values were computed somewhat differently, to facilitate an effective real-time implementation. All model inputs that represent the process variables related to the mill output or separator were based on the current measured value. Model inputs that represent process variables related to the mill input were based on the value measured 30 min earlier. These measured process variables were then filtered with a low-pass filter and applied as model inputs to estimate the current fineness. For the mill motor current, following the analysis conducted in Section 3.1, the mean value of previous 30 min samples was calculated as a model input.
4. Results and discussion
T ð12Þ ð13Þ
where μi is positive. Levenberg–Marquardt algorithm is combination of Gauss–Newton algorithm and gradient descent method. For small values of μi, it behaves like Gauss–Newton algorithm and for large values of μi, it is equivalent to gradient descent method. Two model development approaches were considered. The first approach, referred in this paper as the single model approach,
4.1. Input selection During the input selection process, it is useful to incorporate knowledge regarding the process to ensure that all important process variables are included in the model. Two approaches were used during input selection and they resulted in two model structures with two different sets of inputs. The first structure was obtained by expending input variables selected by ITSS with additional, potentially significant, inputs on the basis of process knowledge. In the second approach, an initial selection of variables
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
6
important for the process was made based on process knowledge, which is then expanded by means of ITSS algorithm. For the first subset of input variables, the ITSS algorithm was applied on the available data set, and the cumulative information theoretic curve in Fig. 3 presents the results. The figure shows that according to the ITSS algorithm, most of the information (99.7%) in the input set is contained in the first six variables. These six variables are selected as model inputs. Taking into account operating principles of the separator, it is clear that the variables Sep_dpress and Sep_spd have significant influence on cement fineness; however, because Sep_dpress was not included in the subset of input variables selected by the ITSS algorithm, this variable was added to the subset as the seventh input. The second subset of input variables was chosen starting from an initial subset that contained Sep_dpress and Sep_spd variables and then applying the ITSS algorithm. The cumulative information theoretic curve in Fig. 4 presents the results. The curve shows that Sep_spd and Sep_dpress, when chosen alone, contained 30.2% of the information and that the first six variables contain 99.8% of information. First six variables from Fig. 3 are selected as the second subset of variables constituting the model inputs. Careful analysis of the first six variables defined by the ITSS algorithm for the two selected input subsets shows, despite the subsets being different, why they both contain the same amount of information on cement fineness. The mill motor current, separator
1 0.9 0.8
rotor speed and cement temperature at the mill output are included in both input subsets, and the remaining three variables in both input subsets are closely correlated with the corresponding variable in the other subset. When data set includes correlated variables, ITSS algorithm ranks one of them highly and the other lowly [29] and by beginning selection of second subset with initial subset of two variables, it leads to the selection of different correlated variables compared to the first subset. The correlated variables are presented in Table 2. It is logical that there is a correlation between fill levels in the mill chambers. The second set of variables from Table 2 provides information on the amount of material passing through the separator. The rejects flow rate provides information on the amount of material returning from the separator, and the elevator current provides information on the amount of material entering the separator. The correlation between clinker flow at the mill input and the differential pressure at the separator inlet is derived from the control strategy used during production. Fineness can be controlled by changing the differential pressure at the separator inlet and by changing the separator speed. In the control strategy used in the factory, the differential pressure is set to a constant value depending on cement type, and additional corrective actions are achieved by changing the separator speed. It is clear that in this way, the differential pressure at the separator inlet is correlated with the clinker flow because both are defined by the cement type. A close correlation between those two variables will disappear if the control strategy is changed. Selecting both variables for model inputs, such as it was performed for the first input subset by adding the differential pressure to the input subset after the ITSS algorithm, is a sensible approach for preserving information in case the control strategy changes.
U-value
0.7
4.2. Model performance
0.6 0.5
Because during input selection, two candidate subsets of input variables were adopted, the models using the two approaches presented in the previous section were developed for both input subsets. Model performance is evaluated by calculating mean square error (MSE)
0.4 0.3
Mill_airT
Elev_curr
Sep_spd
Mill_lvl2
Mill_cemT
Rtrn_sep
Mill_curr
0.1
Clink_flw
0.2
Fig. 3. Cumulative information theoretic curve according to the ITSS algorithm used in selecting the first subset of input variables.
0.9 0.8 0.7 U-value
MSE ¼
i¼1
ð14Þ
n
produced for test data set. 4.2.1. Single model approach results Using this approach, two models were created, model 1, based on the first subset of input variables selected with the ITSS algorithm, and model 2, based on the second subset of input variables selected with the ITSS algorithm. The results obtained from the two models are presented in Figs. 5 and 6. The results show that the performances of these two models using the test data set are significantly different. The performance of model 1 on the test set is far from satisfactory, whereas the performance of model 2 is much better, as seen from the mean square errors presented in Table 3.
1
0.6 0.5 0.4 0.3
Table 2 Correlated variables in the selected subsets of input variables.
Mill_airT
Mill_lvl2
Mill_lvl1
Mill_cemT
Elev_curr
Mill_curr
Sep_dpress
Sep_spd
0.2 0.1
n i y Þ2 ∑ ðyi _
Fig. 4. Cumulative information theoretic curve according to the ITSS algorithm used in selecting the second subset of input variables.
Input subset 1 variable
Input subset 2 variable
Mill_lvl2 Rtrn_sep Clink_flw
Mill_lvl1 Elev_curr Sep_dpress
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
16
14
14
fineness[%]
12 10 8 6 4
laboratory
fineness - 35µm sieve residue [%]
16
estimation 0
50
100
150
200 250 samples
300
350
Fig. 5. Fineness determined in laboratory and estimated by neural network model 1 on the test data set.
8 6
laboratory estimation 0
50
100
150
200 250 samples
300
350
400
16
14 12 10 8 6
laboratory estimation 0
50
100
150
200
250
300
350
400
fineness - 35µm sieve residue [%]
fineness - 35µm sieve residue [%]
10
Fig. 7. Fineness determined in laboratory and estimated by neural network model 3 on the test data set.
16
4
12
4
400
7
14 12 10 8 6 4
laboratory estimation 0
50
100
150
samples Fig. 6. Fineness determined in laboratory and estimated by neural network model 2 on the test data set.
Table 3 Mean square error on the test data set for models based on the single model approach. Model
MSE
1 2
2.52 1.25
Model 1 performed well for first 120 samples but then began to significantly deteriorate. Even when the model performance began to deteriorate, there were still sets of samples where the model performed well; however, by the end of test data set, the model performance was so poor that it effectively lost prediction capability. Model 2 performed much better; however, upon closer examination, the results shows that its performance also deteriorated over time, although not as dramatically as for model 1. It is obvious that the model performance decreased with every new sample due to their inability to adopt the changes in the grinding circuit.
200 250 samples
300
350
400
Fig. 8. Fineness determined in laboratory and estimated by neural network model 4 on the test data set.
Table 4 Mean square error on test data set for models based on the retrained model approach. Model
MSE
3 4
0.95 1.01
Figs. 7 and 8 show that the retrained neural network models based on the retrained model approach performed satisfactory for both subsets of input variables. Table 4 gives the MSE computed for the two models on the test data set. For both models, the MSE is significantly smaller compared to the MSE of the models based on the single model approach; in addition, model performance did not deteriorate with time. Therefore, this approach was adopted for soft sensor implementation.
4.3. Soft sensor performance 4.2.2. Results of using the retrained model approach Using this approach, two models were created, model 3, based on the first subset of input variables selected with the ITSS algorithm, and model 4, based on the second subset of input variables selected with the ITSS algorithm. The results obtained for these two models are presented in Figs. 7 and 8.
The retrained neural network models showed satisfactory performance on the test data set consisting of laboratory samples taken every two hours, and therefore, both models, model 3 and model 4, were considered for soft sensor implementation. Soft sensors were first tested off-line on data collected in cement factory during normal production. Following the results of off-line
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
8
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Fig. 9. Information on product fineness provided by laboratory and by the soft sensor based on neural network model 3 – off-line test. (a) complete test data; (b) 85th–89th hour of test data.
tests, soft sensors were implemented and their on-line performance is tested in the cement factory.
4.3.1. Results of off-line tests Soft sensors were tested on the data set that represents continuous cement production. Fineness was determined every minute, and the results presented in Figs. 9 and 10 show that the developed soft sensors exhibited satisfactory performance. From the results obtained between the 74th and 76th hour of the test, and even more between the 86th and 88th hour, it is clear that in the cases when significant changes in cement fineness occurred, information from the soft sensors is provided much earlier than when obtained from the laboratory tests. Detailed results from 86th to 88th hour presented in Figs. 9 and 10 illustrate this ability of soft sensor. They show that soft sensors detected that cement fineness started to change significantly after laboratory test conducted in 86th hour and how fineness after 30 min reached the level that was detected by laboratory test only 90 min later, in 88th hour. Irregular operating conditions of the cement grinding circuit can also be identified from behavior exhibited by the soft sensor, as it can be seen from test results when in the 55th hour of the test a problem occurred in the mill dosing system and the dosing system had been turned off for approximately 20 min. The behavior of the soft sensor based on model 3 was more influenced by this irregular situation because one of the model inputs, the
Fig. 10. Information on product fineness provided by laboratory and by the soft sensor based on neural network model 4 – off-line test. (a) complete test data; (b) 85th–89th hour of test data.
clinker flow rate, is related to the dosing system, whereas this is not the case with the soft sensor based on model 4. This situation shows that instead of choosing one model for soft sensor of the two presented in this paper, it is reasonable to use both models for implementation. This would provide redundancy that can be useful in situations when irregular operating conditions influence only one of the soft sensors.
4.3.2. Performance of implemented soft sensors Both sensors, the first one based on model 3 and the second one based on model 4, are currently implemented in cement factory to provide the estimation of cement fineness on-line. One change is made during implementation compared to the approach used in off-line tests of soft sensors performance. To reduce the influence of measurement noise on estimations, for every input, instead of one measurement taken every minute, mean value of measurements taken every 10 s in the previous minute is used. The on-line performance of soft sensors for period lasting 4 days, three months after installation is presented in Figs. 11 and 12. As it can be seen, results obtained during on-line operation fully confirm results obtained during off-line tests and measurement noise influence in on-line fineness estimation is reduced. Fineness estimations provided by soft sensors are currently used by plant operators in monitoring of cement grinding process. Laboratory tests of fineness conducted after soft sensor installation show that all produced cement is of adequate fineness. By obtaining timely information on changes in grinding process performance
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
9
5. Conclusions
Fig. 11. On-line performance of implemented soft sensor based on neural network model 3.
Fig. 12. On-line performance of implemented soft sensor based on neural network model 4.
from soft sensors, operators are able to undertake appropriate corrective actions and completely eliminate situations that lead to production of cement with inadequate fineness. Use of soft sensors in this way significantly contributes to stable performance of grinding circuit, but the main benefits can be achieved by including soft sensors in automatic control of grinding circuit. The simplest approach would be to implement PI control of fineness by manipulating separator rotor speed or air flow rate and using soft sensor to provide feedback information. This approach should provide the possibility to guide the production process closer to the product specification limits regarding fineness and further reduce energy consumption. However, this simple approach is not considering the influence of separator performance on mill operation through rejects flow rate. The research in development of multivariable controller for cement grinding circuit that, besides fineness, includes control of main indicators of mill grinding performance, primarily mill fill level, would probably result in more efficient production.
Soft sensors that provide real-time information on immeasurable process variables or process variables available only from offline laboratory tests are extremely useful tools that can be used for monitoring and control purposes. The implementation of soft sensors is an extensive task, and the quality of the result is highly dependent on the process analysis conducted in cooperation with plant experts and process operators. Soft sensors presented in this paper could determine the cement fineness in a cement grinding circuit that is normally available only from off-line laboratory tests. Real-time information on cement fineness provided by the soft sensor can help reduce energy consumption of the cement grinding circuit and maintain a uniform quality of cement. Two types of static nonlinear models were considered as candidates for soft sensors, and their ability to estimate cement fineness of laboratory samples was analyzed. As it can be seen from Figs. 5 and 6, the performance of models developed using one training and validation set deteriorated over time, whereas Figs. 7 and 8 show that models based on a retrained neural network had significantly improved performance. Soft sensors based on the retrained neural network model were developed and tested on data from continuous cement production. The models were trained using processed input data to correspond with the sampling procedure used in the laboratory fineness measurements, where fineness was estimated in real time using the current measured values for inputs. The soft sensor off-line tests showed fully satisfactory performance and, as Figs. 9 and 10 demonstrate, the estimations provided by soft sensors concur with fineness laboratory tests. Also, when rapid changes in system operation occur, soft sensors indicate them much earlier than the laboratory tests. On-site performance of soft sensors presented on Fig. 11 and 12 confirms results obtained during off-line tests. This indicates that soft sensors based on this approach can supply significant information on real-time performance of a cement grinding circuit. The soft sensors currently implemented in the cement factory are used to monitor the production process, avoiding the long delay introduced by laboratory tests. The quality of the soft sensor performance suggests the possibility of using their real-time estimates in a control loop. The further research should be directed towards designing the appropriate control algorithm for cement fineness control based on soft sensor estimation.
Acknowledgment The authors would like to thank Lafarge, Beočin, Serbia, for help and support, which made this study possible.
References [1] Bhatty JI, Miller FM, Kosmatka SH. Innovations in Portland Cement Manufacturing. Skokie: Portland Cement Association; 2004. [2] Jankovic A, Valery W, Davis E. Cement grinding optimisation. Miner Eng 2004;17:1075–81. [3] Madloola NA, Saidura R, Hossaina MS, Rahimb NA. A critical review on energy use and savings in the cement industries. Renew Sustain Energy Rev 2011;15:2042–60. [4] Hosten C, Fidan B. An industrial comparative study of cement clinker grinding systems regarding the specific energy consumption and cement properties. Powder Technol 2012;221:183–8. [5] Fortuna L, Rizzo A, Sinatra M, Xibilia MG. Soft analyzers for a sulfur recovery unit. Control Eng Pract 2003;11:1491–500. [6] Sbarbaro D, Ascencio P, Espinoza P, Mujica F, Cortes G. Adaptive soft-sensors for on-line particle size estimation in wet grinding circuits. Control Eng Pract 2008;16:171–8. [7] Casali A, Gonzalez G, Tones F, Vellebuona G, Castelli L, Gimenez P. Particle size distribution soft-sensor for a grinding circuit. Powder Technol 1998;99:15–21.
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
10
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
[8] Ponsart J-C, Theilliol D, Aubrun C. Virtual sensors design for active fault tolerant control system applied to a winding machine. Control Eng Pract 2010;18:1037–44. [9] Fortuna L, Graziania S, Xibilia MG. Soft sensors for product quality monitoring in debutanizer distillation columns. Control Eng Pract 2005;13:499–508. [10] Rani A, Singh V, Gupta JRP. Development of soft sensor for neural network based control of distillation column. ISA Trans 2013;52:438–49. [11] Vijaya Raghavana SR, Radhakrishnan TK, Srinivasan K. Soft sensor based composition estimation and controller design for an ideal reactive distillation column. ISA Trans 2011;50:61–70. [12] Kang P, Lee H-j, Cho S, Kim D, Park J, Park C-K. A virtual metrology system for semiconductor manufacturing. Expert Syst Appl 2009;36:12554–61. [13] Kang P, Kim D, Lee H -j, Doh S, Cho S. Virtual metrology for run-to-run control in semiconductor manufacturing. Expert Syst Appl 2011;38:2508–22. [14] de Assis AJ, Filho RM. Soft sensors development for on-line bioreactor state estimation. Comput Chem Eng 2000;24:1099–103. [15] Liu G, Zhou D, Xu H, Mei C. Model optimization of SVM for a fermentation soft sensor. Expert Syst Appl 2010;37:2708–13. [16] Yap WK, Karri V. ANN virtual sensors for emissions prediction and control. Appl Energy 2011;88:4505–16. [17] Xie L, Zhao Y, Aziz D, Jin X, Geng L, Goberdhansingh E, Qi F, Huang B. Soft sensors for online steam quality measurements of OTSGs. J Process Control 2013;23:990–1000. [18] Deng J, Xie L, Chen L, Khatibisepehr S, Huang B, Xu F, Espejo A. Development and industrial application of soft sensors with on-line Bayesian model updating strategy. J Process Control 2013;23:317–25.
[19] Pani AK, Vadlamudi VK, Mohanta HK. Development and comparison of neural network based soft sensors for online estimation of cement clinker quality. ISA Trans 2013;52:19–29. [20] Etien E. Modeling and simulation of soft sensor design for real-time speed estimation, measurement and control of induction motor. ISA Trans 2013;52: 358–64. [21] Boulvin M, Vande Wouwer A, Lepore R, Renotte C, Remy M. Modeling and Control of Cement Grinding Processes. IEEE Trans Control Syst Technol 2003;11:715–25. [22] Sridhar DV, Bartlett EB, Seagrave RC. Information theoretic subset selection for neural network models. Comput Chem Eng 1998;22:613–26. [23] Papadokonstantakis S, Machefer S, Schnitzlein K, Lygeros AI. Variable selection and data pre-processing in NN modelling of complex chemical processes. Comput Chem Eng 2005;29:1647–59. [24] Fortuna L, Graziani S, Rizzo A, Xibilia MG. Soft Sensors for Monitoring and Contorl of Industrial Processes. London: Springer-Verlag; 2007. [25] Marquardt DW. An algorithm for least-squares estimation of nonlinear parameters. J Soc Ind Appl Math 1963;2:431–41. [26] Dreyfus G. Neural Networks: Methodology and Applications. Berlin: SpringerVerlag; 2005. [27] Dias FM, Antunes A, Vieira J, Mota A. A sliding window solution for the on-line implementation of the Levenberg–Marquardt algorithm. Engineering Applications of Artificial Intelligence 2006;19:1–7. [28] Ferreira PM, Ruano AE. Online sliding-window methods for process model adaptation. IEEE Trans Instrum Meas 2009;58:3012–20. [29] Papadokonstantakis S, Lygeros A, Jacobsson SP. Comparison of recent methods for inference of variable influence in neural networks. Neural Netw 2006;19:500–13.
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
Contents lists available at ScienceDirect
ISA Transactions journal homepage: www.elsevier.com/locate/isatrans
Soft sensor for real-time cement fineness estimation Darko Stanišić n, Nikola Jorgovanović, Nikola Popov, Velimir Čongradac Faculty of Technical Sciences, University of Novi Sad, Novi Sad, Serbia
art ic l e i nf o
a b s t r a c t
Article history: Received 10 August 2013 Received in revised form 14 August 2014 Accepted 27 September 2014 This paper was recommended for publication by Dr. Ahmad B. Rad.
This paper describes the design and implementation of soft sensors to estimate cement fineness. Soft sensors are mathematical models that use available data to provide real-time information on process variables when the information, for whatever reason, is not available by direct measurement. In this application, soft sensors are used to provide information on process variable normally provided by offline laboratory tests performed at large time intervals. Cement fineness is one of the crucial parameters that define the quality of produced cement. Providing real-time information on cement fineness using soft sensors can overcome limitations and problems that originate from a lack of information between two laboratory tests. The model inputs were selected from candidate process variables using an information theoretic approach. Models based on multi-layer perceptrons were developed, and their ability to estimate cement fineness of laboratory samples was analyzed. Models that had the best performance, and capacity to adopt changes in the cement grinding circuit were selected to implement soft sensors. Soft sensors were tested using data from a continuous cement production to demonstrate their use in real-time fineness estimation. Their performance was highly satisfactory, and the sensors proved to be capable of providing valuable information on cement grinding circuit performance. After successful off-line tests, soft sensors were implemented and installed in the control room of a cement factory. Results on the site confirm results obtained by tests conducted during soft sensor development. & 2014 ISA. Published by Elsevier Ltd. All rights reserved.
Keywords: Soft sensors Cement fineness Estimation Neural-network models Product quality
1. Introduction Fineness is one of the most important physical properties of cement. Cement fineness tests are performed in laboratory on samples taken from a grinding circuit. The testing procedure is defined by appropriate standards and is executed at uniform time intervals, typically every two hours [1]. There are numerous limitations due to such important information being available only periodically at extremely long time intervals. Any algorithm that automatically controls a grinding circuit, instead of cement fineness, must be related to another, alternative and easily measurable system parameter, which makes direct control of the most important system parameter impossible. This situation means it is always necessary to guide the system toward operating points that should provide, at a high level of confidence, the desired fineness, usually resulting in fineness far above the limit defined for the product. In this way, much more energy is consumed during the cement grinding process than necessary. The problem of excess energy consumption during the cement grinding process is clear when accounting for the amount of electrical energy consumed in cement n
Corresponding author. Tel.: þ 381 21 485 2452; fax: þ 381 21 458 873. E-mail addresses: [email protected] (D. Stanišić), [email protected] (N. Jorgovanović), [email protected] (N. Popov), [email protected] (V. Čongradac).
production, which is approximately 110 kW h/t, of which, approximately 40% is consumed for cement grinding [2,3]. Since there is no available real-time cement fineness measurement, the focus in energy consumption reduction in cement grinding circuit is directed to improvements in grinding technology [3]. Improved grinding media in traditional ball mills, new grinding mill designs, combined grinding systems [4], clinker pre-crushing [2] and use of highefficiency separators are energy saving measures commonly undertaken by cement manufacturers. Providing real-time information of cement fineness would offer new, complementary opportunities for reduction of energy consumption. In addition, other equally important problems originate from the lack of information about cement fineness. Because no information on cement fineness between two samples and related laboratory tests is available, when a laboratory test shows deviation from desired values, it is reasonable to assume that there is a significant amount of poor-quality cement produced between the two samples. That deviation from desired values can be significant and indicate that the process is shifted far away from normal operating conditions; in this situation, it is typically necessary to take radical corrective actions to guide the process to the desired performance. In addition, there is a possibility that even when the cement fineness of two consecutive samples are adequate, in several situations, a large portion of cement produced between the two samples may be of completely different, inadequate, quality. All
http://dx.doi.org/10.1016/j.isatra.2014.09.019 0019-0578/& 2014 ISA. Published by Elsevier Ltd. All rights reserved.
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
2
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
these problems lead to the conclusion that finding a method that provides information on cement fineness in real time can significantly and positively influence the two main indicators of cement grinding circuit performance, energy consumption and quality of cement. The objective of this paper is to obtain information on cement fineness by determining its correlation with measurable process variables. In this way, reliable real-time estimates of cement fineness can be obtained using a mathematical model to derive the fineness from available data. Such systems are typically called soft sensors or virtual sensors. Soft sensors are used in different types of applications. Soft sensors for measuring acid gases emission in sulfur recovery unit of oil refinery presented in [5] and particle size estimation in wet grinding circuits presented in [6] are applications where soft sensors are used to provide reliable process information when hardware sensors are removed for maintenance. When one hardware sensor is time shared by several parallel processes, soft sensors can be used during time intervals in which hardware sensor is unavailable [7]. Soft sensors working in parallel with hardware sensors can be used for fault detection and can also make it possible for a system in such situations to continue process control using estimated values from a soft sensor to replace measurements from a faulty hardware sensor. In [8] an example of such application in winding machine is presented. Situations when measurements from hardware sensors are not suitable for use in control algorithms because they have a large time delay due to their position in the production process or because of their time-consuming operation, can be solved by soft sensors [9–11]. In [9] this type of solution is used for monitoring of product quality in oil refinery, while in [10,11] chemical process is controlled by controlling the product composition estimated by soft sensor. In a production process where product quality is controlled with only a few sample products and the quality of every product cannot be assured, soft sensors provide a way to obtain information on the quality of each product [12] and assist quality management and process control [13]. Soft sensor applications similar to the cement fineness application address situations where important process variables are available only from off-line laboratory tests and are therefore unavailable for use in control algorithms. This is the case in fermentation process as is presented in [14,15], also this approach can be used for emission prediction and control for a gasoline engine [16]. In [17,18] soft sensors are utilized to provide information of process variables used in steam quality control since online measurement of process variables of interest is not always accurate or reliable and is therefore supplemented by lab analysis taken infrequently in a manual manner. Results obtained from soft sensors for online estimation of clinker quality [19] show that important information in cement production process can be provided by the use of soft sensors. Soft sensors have already demonstrated their ability to provide solutions in grinding circuits [6,7], where they were used to provide information on the particle size distribution in wet grinding circuits. The problem described in this paper has one significant difference compared with the aforementioned paper on wet grinding circuits. Whereas soft sensors used in the wet grinding circuits were developed using information from a hardware sensor, which in normal operation, provides real-time measurements of particle sizes, a soft sensor for cement fineness must be developed based on information provided from off-line laboratory tests performed every two hours. Soft sensors are typically developed using first-principle models when they are available and in a suitable form [20]; however, more often, gray or black box models are used. In the case of a cement grinding circuit, due to the complexity of the system, particularly the grinding process, a complete first-principle model suitable for soft sensor development is not available. Attempts to
develop gray box model of this process lead to exceptionally complex models whose parameters could only be determined by delicate and extremely time-consuming experiments with relatively large uncertainties associated with several parameters [21]. In addition, though gray box models based on grinding laws agree with laboratory experiments, their transposition to industrial mills gives poor results. However, for purposes of monitoring and system performance analysis, a great amount of historical data for various process variables is collected, which makes black box model parameter identification convenient. It is necessary to select black box model inputs from large number of candidate process variables. In this paper, selection of model inputs was based on an information theoretic approach. An algorithm called the information theoretic input variable subset selection (ITSS) [22] is used. This algorithm does not depend on any nonlinear function approximation methods; it solely depends on the measured system's input–output data. Information theory is used to analyze interdependency between the process output and inputs to determine, from the entire input data, a subset of input variables that contains most of the information necessary to predict the output. Theoretical knowledge regarding the process that is modeled should also be incorporated when using ITSS to ensure that all important variables to the model are included in the subset [22,23]. In this paper the soft sensor is proposed to estimate cement fineness and provide real-time information on a process variable available only from off-line laboratory tests. The design and implementation of the soft sensor is presented. Data collected during regular production in cement factory Lafarge BFC1 in Beočin, Serbia is used for soft sensor development. Soft sensor was developed using a black box model based on a multi-layer perceptron neural network. Model is trained using the data obtained by standard offline laboratory tests conducted every two hours. Their ability to provide real-time fineness estimation on every minute is tested. Soft sensors are implemented and installed in the cement factory Lafarge in Beočin, Serbia. The remainder of this paper is structured as follows. Section 2 provides a schematic description of the cement grinding circuit. In Section 3, analysis of the collected data is presented together with algorithms and procedures used for models structure selection. In Section 4, experimental results are presented and analyzed, and Section 5 presents the conclusions of this work.
2. Description of the grinding circuit The cement grinding circuit is the final stage in the cement production process. In this stage of production, clinker produced in the kiln together with additives that are used to attain the desired properties of the produced cement are grinded in the mill into a fine powder (cement). The material is grinded in the rotating ball mill where grinding is performed by steel balls. During grinding, the clinker particles should be reduced in size by a factor between 1000 and 10,000, and it has been proven to be more efficient to apply different grinding forces to different size ranges of the material, which is achieved by constructing mills with several grinding chambers, typically two, with different properties, such as chamber length, size of the ball charge and shape of the liners along the chamber walls. Fig. 1 shows a simplified scheme of the cement grinding circuit that indicates process variables of interest in this paper. Material flow is indicated with solid lines, whereas dotted lines represent air flow in the circuit. Clinker and additives are transported from 1
www.lafarge.rs
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
3
P T A
Product
F
P Separator
S Elevator
Clinker
Additive ... Additive 1 n
F
L F
F
I
Rejects L P
F
T
T
Ball Mill I Fig. 1. Simplified scheme of cement grinding circuit.
their bunkers via conveyor belts and are fed into the ball mill, which consists of two chambers. Flow rate of every component fed into ball mill is measured and controlled by dosing system to comply with defined recipes. The level of material in mill chambers is measured by the sensors based on vibration or noise measurement. Two motors, whose currents are measured and monitored, are turning the mill. Material continuously flows through the mill, and the grinded particles that exit the mill are transported by air slide to an elevator. A fan placed at the mill exit provides air ventilation necessary to prevent over-heating of the mill. Temperature of the material exiting the mill is measured by the appropriate sensor, as well as the air temperature and air pressure in the ventilation system on the mill output. The elevator is used to transport grinded material to a high-efficiency separator and the elevator current is measured to provide the information on the amount of material exiting the mill. The material coming from the elevator is dropped into a blowing air stream inside the separator where rotor provides centrifugal force that is used to separate coarse particles. The fine particles are removed from the separator via the air stream and transported to cyclones along with the flow of hot gas heated by the material. Inside the cyclones, the solid material is separated from the gas and taken to storage silos as the final product. After exiting the cyclones, the hot air is recirculated through a fan into the system and back to the separator. Separated coarse particles fall down to the bottom of the separator where they are discharged into the air slide. The coarse particles form rejects flow that is returned to the mill's inlet for an additional grinding cycle along with fresh feed. Several process variables related to the separator operation are measured. Those variables are air flow rate on separator inlet, provided by differential pressure measurement, separator rotor speed, air pressure on the separator inlet and air pressure and air temperature on the separator outlet.
3. Data analysis and model structure Because separating coarse particles from fine particles, which makes the final product, is performed in the separator by interacting drag forces provided by the air stream that enters the separator and centrifugal forces provided by rotor rotation, it is clear that air flow into the separator inlet and rotor speed are two essential variables that must be chosen as model inputs. Other variables that may provide additional insight about the material that enters the separator and the operating conditions of the separator should also be chosen. Material that enters the separator
comes from a ball mill, and its properties depend on the properties of the material that enters the mill and the grinding performances of the mill, which means that process variables from all parts of the grinding circuit are potential model inputs. Selecting appropriate additional variables for model inputs is a complex task. There are numerous measured variables that provide information on system performance; however, separating variables that should be used as model inputs from irrelevant variables that can be detrimental to model performance must be conducted carefully. In this paper, this is done in two steps. In the first step, the set of candidate variables is chosen based on an analysis of the grinding system functionality. All candidate variables are listed in Table 1. From this set of candidate variables, in the second step, inputs are selected using an information theoretic approach. 3.1. Data description and preprocessing The laboratory fineness test is done by sieving the product sample through a 35 μm sieve. The percentage of material that remains on the sieve defines the fineness of the product tested. Samples for the laboratory fineness tests are taken every two hours by an automatic sampler. The sampling procedure lasts 36 min, where 6 samples are taken. Every sample is taken during 1 min with a pause of 6 min between samples. Samples are then mixed together and fineness of joint sample is determined. This implies that the laboratory fineness is determined as an average fineness of the material that exited the separator during the 36 min of the sampling duration time. Therefore, the model inputs should be processed accordingly. The sampling time of the measured process variables is 1 min; however, their mean values over the fineness sampling duration time are used as model inputs. It is clear that all process variables should be delayed for the amount of time necessary for material to come from the point where the process variable is measured to the separator. It takes less than one minute for material to come from the mill output to the separator, and it is clear that process variables related to the mill output, as those linked with the separator can be processed without delay. The time necessary for material to come from the mill input to the mill output is significant, and variables related to the mill input should be delayed. An analysis of the rejects flow rate in situations when material is fed into empty mill showed that rejects are detected after 24–27 min from the start of the process. In normal operating conditions, mill chambers have significantly more material, which could additionally slow the flow of new material from the mill input. To account for any additional delay that could appear during normal operating conditions, a 30 min
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
4
denoted by mx, is done by
Table 1 Candidate input variables and their introduced delay. Variable description
Notation
Units
Introduced delay (min)
Clinker flow rate Gypsum flow rate Slag flow rate Limestone flow rate Fly ash flow rate Fill level of mill chamber 1 Fill level of mill chamber 2 Air pressure on mill output Air temperature on mill output Mill motor current Cement temperature on mill output Elevator current Separator rotor speed Differential pressure at separator inlet Air pressure at separator inlet Air pressure at separator outlet Temperature at separator outlet Rejects flow rate on separator Rejects flow rate on mill input
Clink_flw Gyps_flw Slag_flw Lims_flw FAsh_flw Mill_lvl1 Mill_lvl2 Mill_press Mill_airT Mill_curr Mill_cemT Elev_curr Sep_spd Sep_dpress Sep_pressIn Sep_pressOut Sep_outT Rtrn_sep Rtrn_mill
% % % % % % % bar 1C A 1C A rpm bar bar bar 1C t/h %
30 30 30 30 30 30 0 0 0 0–30 0 0 0 0 0 0 0 0 30
di ¼
xi mx
ð1Þ
σx
where σx is standard deviation of x. If data follows normal distribution, the probability that di 43 is 0.27% and in that case the sample xi can be considered as an outlier. That means that every input variable x needs to be in interval [mx 3σx, mx þ3σx]. This interval can be used as a starting point and in further analysis with plant experts, it can be modified to include working conditions that may occur rarely and are therefore excluded by “3σ edit rule”. Also the data for every variable should be analyzed by visual inspection to check how much it concurs with normal distribution and, if necessary, in cooperation with plant experts the modification of intervals obtained by “3σ edit rule” can be performed. All variables then need to be normalized, usually in [ 1, 1] interval. After all three types of outliers are removed from the initial data set with the presented procedure, the data set which contains 980 samples of laboratory tests with corresponding measured process variables is obtained. This data set is used during the soft sensor development presented in the paper. 3.2. Input selection
delay was defined for all process variables related to the mill input. The introduced delays on measured signal values are presented in Table 1. The mill motor current was not processed in the same way as the other variables. The mill motor current provides information on the mill operating conditions during a time period of approximately 30 min necessary for material to pass from the mill input to mill output. Because of this, in this paper, for every 1 min sample of other process variables, a mean value for the previous 30 min mill current samples is computed as information on the mill current for the material that is exiting the mill at that moment. For mill current values obtained in this way, a mean value over the fineness sampling duration time is then calculated in the same way that it is performed for the other process variables. The measured flow rates of the fresh material entering the mill are used to calculate the percentage of every material in the fresh mill feed. This percentage is thus considered a potential model input. The data collected from a cement grinding circuit by an acquisition system during 4 months of normal operation was used for the soft sensor development. The data were analyzed, and three types of outliers were identified and removed: (1) Data collected during production of products whose fineness is not measured by the standard test. (2) Data from system operation during mill startup. (3) Outliers related to faulty measurements by the sensors.
From the candidate variables listed in Table 1, model inputs can then be selected using the ITSS algorithm. To implement the ITSS algorithm on continuous variables, it is necessary to discretize and divide the domain space of every variable into a finite number of intervals. In this way, vector of input variables x can take one of M discrete values x1, x2, …, xM and dependent variable y can take one of S discrete values y1, y2, …, yS. If N is number of available samples in data set and Ni is number of occurrences of x ¼ xi in the data set, then probability that x can take value xi is pi ¼ pðx ¼ xi Þ ¼
Ni N
The entropy of information presented by variable set x is defined by M
HðxÞ ¼ ∑ pi lnðpi Þ
ð3Þ
i¼1
The probability that x will take value of xi and y will take value of yj is pij ¼ pðx ¼ xi ; y ¼ yj Þ ¼
N ij N
ð4Þ
where Nij is number of joint occurrences of xi and yj in the data set. The conditional entropy of variable y when variables x are given, expresses the information in dependent variable y that remains unexplained when input variables x are known and is defined by HðyjxÞ ¼ ∑pij ln
The first type of outliers is easy to detect, since the information on type of the product is stored in the acquisition system. The second type of outliers can be identified from information on the level of material in mill chambers and rejects flow rate, since mill startup is considered over and system is in normal operation when there is certain amount of material in the mill and rejects flow rate reaches certain level. The minimal amount of material in the mill and minimal amount of rejects during normal operation are technological parameters and can be obtained from plant experts. After removing the first two types of outliers, the remaining data can be processed to identify and remove all other outliers usually related to faulty measurements. It can be done by using “3σ edit rule” [24]. For this rule, normalization of distance between one sample xi and mean value of all available samples of variable x,
ð2Þ
i;j
pij pi
ð5Þ
The amount of information about y contained in variables x is shown by an asymmetric dependency coefficient that measures the dependency of output y on variables in x UðyjxÞ ¼
HðyÞ HðyjxÞ HðyÞ
ð6Þ
where H(y) is the entropy calculated for the dependent variable y. When there are no useful information about y contained in x, UðyjxÞ is 0. If information contained in x completely defines y and can be used to predict y exactly, UðyjxÞ is 1. Variable selection is performed by a forward selection procedure, where in each step, a new subset is generated by adding one input variable to the current subset. The new input variable added to x is the variable
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
f1 x1
f1
x2
yˆ f2
. . .
f1 . . .
xn
f1 Input layer
Hidden layer
Output layer
Fig. 2. Multi-layer perceptron neural network.
that gives the maximum value of UðyjxÞ. Input variables are added until UðyjxÞ value becomes close to 1. It is important not to divide domain space in too many intervals because random noise may be considered as information. Also, not dividing domain space in enough intervals can result in loss of accuracy. Analysis of collected data shows that measurement noise for all candidate variables is below 5% of domain space. For this reason the domain space of every input variable and output cement fineness was divided into 20 intervals.
5
developed one candidate model from the available data set to be used as a soft sensor. For this approach, from the available data set containing 980 samples, the first 550 samples were used for training and validation, where 70% of the samples were used for training and 30% for validation. Selecting the training and validation samples was performed randomly. The remaining 430 samples were used for testing. The second approach, referred in this paper as the retrained model approach, began with one candidate model defined using the first approach; however, every time a new sample is taken and laboratory results are available, the model is updated. The neural network is then retrained using the newest 550 data samples for training and validation. Because new data samples are obtained every two hours, on-line retraining of the neural network between two samples can be achieved without any problems, and using this approach does not have any negative influence on soft sensor implementation. Motivation for this approach was derived from the fact that cement grinding circuit characteristics change with time (changes in material, wear of grinding balls, etc.), and a soft sensor for cement fineness estimation must be able to adopt changes in the cement grinding circuit. In this way, the soft sensor is defined by a moving window of new data that includes only the current characteristics of the cement grinding circuit. This approach is based on a sliding-window policy commonly used in identification of time varying systems [27,28].
3.3. Neural network models In this paper, all models considered as candidates for soft sensors are static nonlinear models based on a multi-layer perceptron neural network shown in Fig. 2. For input vector x, model output is _ y ¼ f ðxÞ ð7Þ The neural network models had one hidden layer that used a sigmoidal activation function f 1 ðaÞ ¼
2 1 1 þ e 2a
ð8Þ
and an output layer that used a linear activation function. f 2 ðaÞ ¼ a
ð9Þ
The models were trained by the Levenberg–Marquardt back propagation algorithm [25,26] with early stopping. For single output neural network, model parameters are obtained by minimizing the least squares cost function N k J ¼ ∑ ðyk _ y Þ2
ð10Þ
k¼1
where yk is the desired target output of kth training sample and k _ y is the model output from the kth training sample. If vector of model parameters is denoted by ω, with Levenberg–Marquardt algorithm parameters are updated, at iteration i, by h i1 ~ ωði 1ÞÞ þ μ I ωðiÞ ¼ ωði 1Þ Hð ∇Jðωði 1ÞÞ ð11Þ i N
~ ωðiÞÞ ¼ ∑ Hð
k¼1
k y ek ¼ yk _
∂ek ∂ωðiÞ
∂ek ∂ωðiÞ
3.4. Soft sensor implementation After evaluating the ability of candidate models to estimate cement fineness of laboratory samples, the models with the best performances were selected as soft sensor candidates. Because the purpose of a soft sensor is to provide real-time estimates of cement fineness between two laboratory samples, the performance of the candidate models were tested on a data set that contains 98 h of continuous cement production. Fineness is determined every minute to demonstrate the soft sensor's capability to provide real-time information required to monitor the cement grinding circuit performance. In this paper, candidate neural network models for soft sensors were trained using a training and validation data set consisting of laboratory fineness measurements and the corresponding measured process variables processed as described in Section 3.1. However, input values were computed somewhat differently, to facilitate an effective real-time implementation. All model inputs that represent the process variables related to the mill output or separator were based on the current measured value. Model inputs that represent process variables related to the mill input were based on the value measured 30 min earlier. These measured process variables were then filtered with a low-pass filter and applied as model inputs to estimate the current fineness. For the mill motor current, following the analysis conducted in Section 3.1, the mean value of previous 30 min samples was calculated as a model input.
4. Results and discussion
T ð12Þ ð13Þ
where μi is positive. Levenberg–Marquardt algorithm is combination of Gauss–Newton algorithm and gradient descent method. For small values of μi, it behaves like Gauss–Newton algorithm and for large values of μi, it is equivalent to gradient descent method. Two model development approaches were considered. The first approach, referred in this paper as the single model approach,
4.1. Input selection During the input selection process, it is useful to incorporate knowledge regarding the process to ensure that all important process variables are included in the model. Two approaches were used during input selection and they resulted in two model structures with two different sets of inputs. The first structure was obtained by expending input variables selected by ITSS with additional, potentially significant, inputs on the basis of process knowledge. In the second approach, an initial selection of variables
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
6
important for the process was made based on process knowledge, which is then expanded by means of ITSS algorithm. For the first subset of input variables, the ITSS algorithm was applied on the available data set, and the cumulative information theoretic curve in Fig. 3 presents the results. The figure shows that according to the ITSS algorithm, most of the information (99.7%) in the input set is contained in the first six variables. These six variables are selected as model inputs. Taking into account operating principles of the separator, it is clear that the variables Sep_dpress and Sep_spd have significant influence on cement fineness; however, because Sep_dpress was not included in the subset of input variables selected by the ITSS algorithm, this variable was added to the subset as the seventh input. The second subset of input variables was chosen starting from an initial subset that contained Sep_dpress and Sep_spd variables and then applying the ITSS algorithm. The cumulative information theoretic curve in Fig. 4 presents the results. The curve shows that Sep_spd and Sep_dpress, when chosen alone, contained 30.2% of the information and that the first six variables contain 99.8% of information. First six variables from Fig. 3 are selected as the second subset of variables constituting the model inputs. Careful analysis of the first six variables defined by the ITSS algorithm for the two selected input subsets shows, despite the subsets being different, why they both contain the same amount of information on cement fineness. The mill motor current, separator
1 0.9 0.8
rotor speed and cement temperature at the mill output are included in both input subsets, and the remaining three variables in both input subsets are closely correlated with the corresponding variable in the other subset. When data set includes correlated variables, ITSS algorithm ranks one of them highly and the other lowly [29] and by beginning selection of second subset with initial subset of two variables, it leads to the selection of different correlated variables compared to the first subset. The correlated variables are presented in Table 2. It is logical that there is a correlation between fill levels in the mill chambers. The second set of variables from Table 2 provides information on the amount of material passing through the separator. The rejects flow rate provides information on the amount of material returning from the separator, and the elevator current provides information on the amount of material entering the separator. The correlation between clinker flow at the mill input and the differential pressure at the separator inlet is derived from the control strategy used during production. Fineness can be controlled by changing the differential pressure at the separator inlet and by changing the separator speed. In the control strategy used in the factory, the differential pressure is set to a constant value depending on cement type, and additional corrective actions are achieved by changing the separator speed. It is clear that in this way, the differential pressure at the separator inlet is correlated with the clinker flow because both are defined by the cement type. A close correlation between those two variables will disappear if the control strategy is changed. Selecting both variables for model inputs, such as it was performed for the first input subset by adding the differential pressure to the input subset after the ITSS algorithm, is a sensible approach for preserving information in case the control strategy changes.
U-value
0.7
4.2. Model performance
0.6 0.5
Because during input selection, two candidate subsets of input variables were adopted, the models using the two approaches presented in the previous section were developed for both input subsets. Model performance is evaluated by calculating mean square error (MSE)
0.4 0.3
Mill_airT
Elev_curr
Sep_spd
Mill_lvl2
Mill_cemT
Rtrn_sep
Mill_curr
0.1
Clink_flw
0.2
Fig. 3. Cumulative information theoretic curve according to the ITSS algorithm used in selecting the first subset of input variables.
0.9 0.8 0.7 U-value
MSE ¼
i¼1
ð14Þ
n
produced for test data set. 4.2.1. Single model approach results Using this approach, two models were created, model 1, based on the first subset of input variables selected with the ITSS algorithm, and model 2, based on the second subset of input variables selected with the ITSS algorithm. The results obtained from the two models are presented in Figs. 5 and 6. The results show that the performances of these two models using the test data set are significantly different. The performance of model 1 on the test set is far from satisfactory, whereas the performance of model 2 is much better, as seen from the mean square errors presented in Table 3.
1
0.6 0.5 0.4 0.3
Table 2 Correlated variables in the selected subsets of input variables.
Mill_airT
Mill_lvl2
Mill_lvl1
Mill_cemT
Elev_curr
Mill_curr
Sep_dpress
Sep_spd
0.2 0.1
n i y Þ2 ∑ ðyi _
Fig. 4. Cumulative information theoretic curve according to the ITSS algorithm used in selecting the second subset of input variables.
Input subset 1 variable
Input subset 2 variable
Mill_lvl2 Rtrn_sep Clink_flw
Mill_lvl1 Elev_curr Sep_dpress
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
16
14
14
fineness[%]
12 10 8 6 4
laboratory
fineness - 35µm sieve residue [%]
16
estimation 0
50
100
150
200 250 samples
300
350
Fig. 5. Fineness determined in laboratory and estimated by neural network model 1 on the test data set.
8 6
laboratory estimation 0
50
100
150
200 250 samples
300
350
400
16
14 12 10 8 6
laboratory estimation 0
50
100
150
200
250
300
350
400
fineness - 35µm sieve residue [%]
fineness - 35µm sieve residue [%]
10
Fig. 7. Fineness determined in laboratory and estimated by neural network model 3 on the test data set.
16
4
12
4
400
7
14 12 10 8 6 4
laboratory estimation 0
50
100
150
samples Fig. 6. Fineness determined in laboratory and estimated by neural network model 2 on the test data set.
Table 3 Mean square error on the test data set for models based on the single model approach. Model
MSE
1 2
2.52 1.25
Model 1 performed well for first 120 samples but then began to significantly deteriorate. Even when the model performance began to deteriorate, there were still sets of samples where the model performed well; however, by the end of test data set, the model performance was so poor that it effectively lost prediction capability. Model 2 performed much better; however, upon closer examination, the results shows that its performance also deteriorated over time, although not as dramatically as for model 1. It is obvious that the model performance decreased with every new sample due to their inability to adopt the changes in the grinding circuit.
200 250 samples
300
350
400
Fig. 8. Fineness determined in laboratory and estimated by neural network model 4 on the test data set.
Table 4 Mean square error on test data set for models based on the retrained model approach. Model
MSE
3 4
0.95 1.01
Figs. 7 and 8 show that the retrained neural network models based on the retrained model approach performed satisfactory for both subsets of input variables. Table 4 gives the MSE computed for the two models on the test data set. For both models, the MSE is significantly smaller compared to the MSE of the models based on the single model approach; in addition, model performance did not deteriorate with time. Therefore, this approach was adopted for soft sensor implementation.
4.3. Soft sensor performance 4.2.2. Results of using the retrained model approach Using this approach, two models were created, model 3, based on the first subset of input variables selected with the ITSS algorithm, and model 4, based on the second subset of input variables selected with the ITSS algorithm. The results obtained for these two models are presented in Figs. 7 and 8.
The retrained neural network models showed satisfactory performance on the test data set consisting of laboratory samples taken every two hours, and therefore, both models, model 3 and model 4, were considered for soft sensor implementation. Soft sensors were first tested off-line on data collected in cement factory during normal production. Following the results of off-line
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
8
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Fig. 9. Information on product fineness provided by laboratory and by the soft sensor based on neural network model 3 – off-line test. (a) complete test data; (b) 85th–89th hour of test data.
tests, soft sensors were implemented and their on-line performance is tested in the cement factory.
4.3.1. Results of off-line tests Soft sensors were tested on the data set that represents continuous cement production. Fineness was determined every minute, and the results presented in Figs. 9 and 10 show that the developed soft sensors exhibited satisfactory performance. From the results obtained between the 74th and 76th hour of the test, and even more between the 86th and 88th hour, it is clear that in the cases when significant changes in cement fineness occurred, information from the soft sensors is provided much earlier than when obtained from the laboratory tests. Detailed results from 86th to 88th hour presented in Figs. 9 and 10 illustrate this ability of soft sensor. They show that soft sensors detected that cement fineness started to change significantly after laboratory test conducted in 86th hour and how fineness after 30 min reached the level that was detected by laboratory test only 90 min later, in 88th hour. Irregular operating conditions of the cement grinding circuit can also be identified from behavior exhibited by the soft sensor, as it can be seen from test results when in the 55th hour of the test a problem occurred in the mill dosing system and the dosing system had been turned off for approximately 20 min. The behavior of the soft sensor based on model 3 was more influenced by this irregular situation because one of the model inputs, the
Fig. 10. Information on product fineness provided by laboratory and by the soft sensor based on neural network model 4 – off-line test. (a) complete test data; (b) 85th–89th hour of test data.
clinker flow rate, is related to the dosing system, whereas this is not the case with the soft sensor based on model 4. This situation shows that instead of choosing one model for soft sensor of the two presented in this paper, it is reasonable to use both models for implementation. This would provide redundancy that can be useful in situations when irregular operating conditions influence only one of the soft sensors.
4.3.2. Performance of implemented soft sensors Both sensors, the first one based on model 3 and the second one based on model 4, are currently implemented in cement factory to provide the estimation of cement fineness on-line. One change is made during implementation compared to the approach used in off-line tests of soft sensors performance. To reduce the influence of measurement noise on estimations, for every input, instead of one measurement taken every minute, mean value of measurements taken every 10 s in the previous minute is used. The on-line performance of soft sensors for period lasting 4 days, three months after installation is presented in Figs. 11 and 12. As it can be seen, results obtained during on-line operation fully confirm results obtained during off-line tests and measurement noise influence in on-line fineness estimation is reduced. Fineness estimations provided by soft sensors are currently used by plant operators in monitoring of cement grinding process. Laboratory tests of fineness conducted after soft sensor installation show that all produced cement is of adequate fineness. By obtaining timely information on changes in grinding process performance
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
9
5. Conclusions
Fig. 11. On-line performance of implemented soft sensor based on neural network model 3.
Fig. 12. On-line performance of implemented soft sensor based on neural network model 4.
from soft sensors, operators are able to undertake appropriate corrective actions and completely eliminate situations that lead to production of cement with inadequate fineness. Use of soft sensors in this way significantly contributes to stable performance of grinding circuit, but the main benefits can be achieved by including soft sensors in automatic control of grinding circuit. The simplest approach would be to implement PI control of fineness by manipulating separator rotor speed or air flow rate and using soft sensor to provide feedback information. This approach should provide the possibility to guide the production process closer to the product specification limits regarding fineness and further reduce energy consumption. However, this simple approach is not considering the influence of separator performance on mill operation through rejects flow rate. The research in development of multivariable controller for cement grinding circuit that, besides fineness, includes control of main indicators of mill grinding performance, primarily mill fill level, would probably result in more efficient production.
Soft sensors that provide real-time information on immeasurable process variables or process variables available only from offline laboratory tests are extremely useful tools that can be used for monitoring and control purposes. The implementation of soft sensors is an extensive task, and the quality of the result is highly dependent on the process analysis conducted in cooperation with plant experts and process operators. Soft sensors presented in this paper could determine the cement fineness in a cement grinding circuit that is normally available only from off-line laboratory tests. Real-time information on cement fineness provided by the soft sensor can help reduce energy consumption of the cement grinding circuit and maintain a uniform quality of cement. Two types of static nonlinear models were considered as candidates for soft sensors, and their ability to estimate cement fineness of laboratory samples was analyzed. As it can be seen from Figs. 5 and 6, the performance of models developed using one training and validation set deteriorated over time, whereas Figs. 7 and 8 show that models based on a retrained neural network had significantly improved performance. Soft sensors based on the retrained neural network model were developed and tested on data from continuous cement production. The models were trained using processed input data to correspond with the sampling procedure used in the laboratory fineness measurements, where fineness was estimated in real time using the current measured values for inputs. The soft sensor off-line tests showed fully satisfactory performance and, as Figs. 9 and 10 demonstrate, the estimations provided by soft sensors concur with fineness laboratory tests. Also, when rapid changes in system operation occur, soft sensors indicate them much earlier than the laboratory tests. On-site performance of soft sensors presented on Fig. 11 and 12 confirms results obtained during off-line tests. This indicates that soft sensors based on this approach can supply significant information on real-time performance of a cement grinding circuit. The soft sensors currently implemented in the cement factory are used to monitor the production process, avoiding the long delay introduced by laboratory tests. The quality of the soft sensor performance suggests the possibility of using their real-time estimates in a control loop. The further research should be directed towards designing the appropriate control algorithm for cement fineness control based on soft sensor estimation.
Acknowledgment The authors would like to thank Lafarge, Beočin, Serbia, for help and support, which made this study possible.
References [1] Bhatty JI, Miller FM, Kosmatka SH. Innovations in Portland Cement Manufacturing. Skokie: Portland Cement Association; 2004. [2] Jankovic A, Valery W, Davis E. Cement grinding optimisation. Miner Eng 2004;17:1075–81. [3] Madloola NA, Saidura R, Hossaina MS, Rahimb NA. A critical review on energy use and savings in the cement industries. Renew Sustain Energy Rev 2011;15:2042–60. [4] Hosten C, Fidan B. An industrial comparative study of cement clinker grinding systems regarding the specific energy consumption and cement properties. Powder Technol 2012;221:183–8. [5] Fortuna L, Rizzo A, Sinatra M, Xibilia MG. Soft analyzers for a sulfur recovery unit. Control Eng Pract 2003;11:1491–500. [6] Sbarbaro D, Ascencio P, Espinoza P, Mujica F, Cortes G. Adaptive soft-sensors for on-line particle size estimation in wet grinding circuits. Control Eng Pract 2008;16:171–8. [7] Casali A, Gonzalez G, Tones F, Vellebuona G, Castelli L, Gimenez P. Particle size distribution soft-sensor for a grinding circuit. Powder Technol 1998;99:15–21.
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i
10
D. Stanišić et al. / ISA Transactions ∎ (∎∎∎∎) ∎∎∎–∎∎∎
[8] Ponsart J-C, Theilliol D, Aubrun C. Virtual sensors design for active fault tolerant control system applied to a winding machine. Control Eng Pract 2010;18:1037–44. [9] Fortuna L, Graziania S, Xibilia MG. Soft sensors for product quality monitoring in debutanizer distillation columns. Control Eng Pract 2005;13:499–508. [10] Rani A, Singh V, Gupta JRP. Development of soft sensor for neural network based control of distillation column. ISA Trans 2013;52:438–49. [11] Vijaya Raghavana SR, Radhakrishnan TK, Srinivasan K. Soft sensor based composition estimation and controller design for an ideal reactive distillation column. ISA Trans 2011;50:61–70. [12] Kang P, Lee H-j, Cho S, Kim D, Park J, Park C-K. A virtual metrology system for semiconductor manufacturing. Expert Syst Appl 2009;36:12554–61. [13] Kang P, Kim D, Lee H -j, Doh S, Cho S. Virtual metrology for run-to-run control in semiconductor manufacturing. Expert Syst Appl 2011;38:2508–22. [14] de Assis AJ, Filho RM. Soft sensors development for on-line bioreactor state estimation. Comput Chem Eng 2000;24:1099–103. [15] Liu G, Zhou D, Xu H, Mei C. Model optimization of SVM for a fermentation soft sensor. Expert Syst Appl 2010;37:2708–13. [16] Yap WK, Karri V. ANN virtual sensors for emissions prediction and control. Appl Energy 2011;88:4505–16. [17] Xie L, Zhao Y, Aziz D, Jin X, Geng L, Goberdhansingh E, Qi F, Huang B. Soft sensors for online steam quality measurements of OTSGs. J Process Control 2013;23:990–1000. [18] Deng J, Xie L, Chen L, Khatibisepehr S, Huang B, Xu F, Espejo A. Development and industrial application of soft sensors with on-line Bayesian model updating strategy. J Process Control 2013;23:317–25.
[19] Pani AK, Vadlamudi VK, Mohanta HK. Development and comparison of neural network based soft sensors for online estimation of cement clinker quality. ISA Trans 2013;52:19–29. [20] Etien E. Modeling and simulation of soft sensor design for real-time speed estimation, measurement and control of induction motor. ISA Trans 2013;52: 358–64. [21] Boulvin M, Vande Wouwer A, Lepore R, Renotte C, Remy M. Modeling and Control of Cement Grinding Processes. IEEE Trans Control Syst Technol 2003;11:715–25. [22] Sridhar DV, Bartlett EB, Seagrave RC. Information theoretic subset selection for neural network models. Comput Chem Eng 1998;22:613–26. [23] Papadokonstantakis S, Machefer S, Schnitzlein K, Lygeros AI. Variable selection and data pre-processing in NN modelling of complex chemical processes. Comput Chem Eng 2005;29:1647–59. [24] Fortuna L, Graziani S, Rizzo A, Xibilia MG. Soft Sensors for Monitoring and Contorl of Industrial Processes. London: Springer-Verlag; 2007. [25] Marquardt DW. An algorithm for least-squares estimation of nonlinear parameters. J Soc Ind Appl Math 1963;2:431–41. [26] Dreyfus G. Neural Networks: Methodology and Applications. Berlin: SpringerVerlag; 2005. [27] Dias FM, Antunes A, Vieira J, Mota A. A sliding window solution for the on-line implementation of the Levenberg–Marquardt algorithm. Engineering Applications of Artificial Intelligence 2006;19:1–7. [28] Ferreira PM, Ruano AE. Online sliding-window methods for process model adaptation. IEEE Trans Instrum Meas 2009;58:3012–20. [29] Papadokonstantakis S, Lygeros A, Jacobsson SP. Comparison of recent methods for inference of variable influence in neural networks. Neural Netw 2006;19:500–13.
Please cite this article as: Stanišić D, et al. Soft sensor for real-time cement fineness estimation. ISA Transactions (2014), http://dx.doi. org/10.1016/j.isatra.2014.09.019i