- Email: [email protected]
Soft Sensors for Industrial Problems
4th International Conference on Advances in Control and Optimization of Dynamical Systems 4th International International Conference on Advances Advances in in Control Control and and 4th Conference on Optimization of Dynamical Systems February 1-5, 2016. NIT Tiruchirappalli, India 4th International Conference on Advances in Control Available onlineand at www.sciencedirect.com Optimization of of Dynamical Dynamical Systems Systems Optimization February 1-5, 2016. NIT Tiruchirappalli, India Optimization of Dynamical Systems February February 1-5, 1-5, 2016. 2016. NIT NIT Tiruchirappalli, Tiruchirappalli, India India February 1-5, 2016. NIT Tiruchirappalli, India
ScienceDirect
IFAC-PapersOnLine (2016)Sensors 100–105 Developing ANN based Virtual 49-1 / Soft for Industrial Problems Developing ANN based Virtual / Soft Sensors for Industrial Problems Developing ANN based Virtual / Soft Sensors for Industrial Problems Developing ANN basedShivaram VirtualKamat* / Soft Sensors for Industrial Problems KP Madhavan**
Shivaram Shivaram Kamat* Kamat* KP KP Madhavan** Madhavan** Shivaram Kamat* Madhavan** KP Shivaram Kamat* Madhavan** *SMIEEE, Engineering & Industrial Services, Tata Consultancy Services Limited, Pune, 411006 KP *SMIEEE, Engineering & Industrial Services, Tata Consultancy Services India (Tel: 20-6608-7528; e-mail: [email protected]). *SMIEEE, Engineering & Industrial Services, Tata Consultancy Services Limited, Limited, Pune, Pune, 411006 411006 *SMIEEE, & Services, Tata Consultancy *SMIEEE, Engineering Engineering & Industrial Industrial Services, Tata [email protected]). Consultancy Services Services Limited, Limited, Pune, Pune, 411006 411006 India (Tel: 20-6608-7528; e-mail: India (Tel: (Tel: 20-6608-7528; 20-6608-7528; e-mail: e-mail: [email protected]). [email protected]). India India (Tel:Engineering, 20-6608-7528; e-mail: [email protected]). ** Department of Chemical Indian Institute of Technology Bombay, Mumbai, 400076 ** ** Department Department of of Chemical Chemical Engineering, Engineering, Indian Indian Institute Institute of of Technology Technology Bombay, Bombay, Mumbai, Mumbai, 400076 400076 ** Engineering, Indian Institute ** Department Department of of Chemical Chemical India Engineering, Institute of of Technology Technology Bombay, Bombay, Mumbai, Mumbai, 400076 400076 (e-mail: Indian [email protected]). India (e-mail: [email protected]). India (e-mail: [email protected]). India India (e-mail: (e-mail: [email protected]). [email protected]). Abstract: This paper briefly reports experience of our research group in developing and deploying some Abstract: This paper briefly briefly reports experience of of ourcrucial research group in in for developing and deploying deploying some promising This approaches for virtual andexperience soft sensing parameters use in control systems some of a Abstract: paper reports of our research group developing and Abstract: This paper briefly reports experience of our research group in developing and some Abstract: This paper briefly reports experience of ourcrucial research group in for developing and deploying deploying some promising approaches for virtual and soft sensing of parameters use in control systems of byaa process or plant. It briefs on the constraints and limitations of the measurement of crucial parameters promising approaches approaches for for virtual virtual and and soft soft sensing sensing of of crucial crucial parameters parameters for for use use in in control control systems systems of of promising a promising approaches for on virtual andand softviability sensing ofgocrucial parameters for use The in crucial control systems ofare process or plant. It briefs the constraints and limitations of the measurement of parameters bya physical means and hence the need to for virtual/soft sensing.. approaches used process or or plant. plant. It It briefs briefs on on the the constraints constraints and and limitations limitations of of the the measurement measurement of of crucial crucial parameters parameters by by process process plant. It briefs on constraints and limitations ofsupervised the measurement of crucial parameters by physicalor means and hence thethe need and topologies viability to and go for for virtual/soft sensing.. The approaches used are variants of Artificial Neural Network their and partly supervised training physical means and hence the need and viability to go virtual/soft sensing.. The approaches used are physical means and hence the need and viability to go for virtual/soft sensing.. The approaches used are physical means and hence the need and viability to go for virtual/soft sensing.. The approaches used are variants of Artificial Neural Network topologies and their supervised and partly supervised training algorithms. The paperNeural provides a brieftopologies overview of virtual sensor development based ontraining these variants of Artificial Artificial Neural Network topologies andthe their supervised and partly partly supervised supervised training variants of Network and their supervised and variants of for Artificial Network andthe their supervised andjustify partlythe supervised algorithms. Theselected paperNeural provides a brief brieftopologies overview of virtual sensor to development based on ontraining these approaches process situations and provides validation results approaches used. algorithms. The paper provides a overview of the virtual sensor development based these algorithms. The paper provides aa brief overview of the virtual sensor development based on these algorithms. The paper provides brief overview of the virtual sensor development based on these approaches for selected process situations and provides validation results to justify the approaches used. The industryfor sectors forprocess which these solutions were developed areresults automotive, cement grinding process approaches selected situations and provides validation to justify the approaches used. approaches for selected process situations and provides validation results to the approaches used. approaches forsectors selected process situations and provides validation to justify justify thebeen approaches used. The industry for which process these solutions were developed developed areresults automotive, cement grinding process and kiln process, and chemical for pH control. Some of these approaches have implemented The industry sectors for which these solutions were are automotive, cement grinding process The industry sectors for which these solutions were developed are automotive, cement grinding process The industry sectors for which these solutions were developed are automotive, cement grinding process and kiln process, and chemical process for pH control. Some of these approaches have been implemented in the industry underlining the significant the virtual/soft mechanisms could play in process and kiln process, and chemical process forrole pH control. Some ofsensing these approaches have been implemented and kiln process, and process pH Some these have been implemented and kiln process, and chemical chemical process for forrole pH control. control. Some of ofsensing these approaches approaches have been implemented in the industry underlining the significant significant the virtual/soft virtual/soft mechanisms could play in process process operations. in the industry underlining the role the the sensing mechanisms mechanisms could play in in the industry underlining the significant role virtual/soft sensing could play in in the industry underlining the significant role the virtual/soft sensing mechanisms could play in process process operations. operations. operations. © 2016, IFAC (International Federation Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Keywords: virtual sensing, soft sensing,ofartificial neural networks, wavelets, wavenets operations. Keywords: virtual sensing, soft sensing, artificial neural networks, wavelets, wavenets Keywords: networks, Keywords: virtual virtual sensing, sensing, soft soft sensing, sensing, artificial artificial neural neural networks, wavelets, wavelets, wavenets wavenets Keywords: virtual sensing, soft sensing, artificial neural networks, wavelets, wavenets are RNN (Recurrent Neural Network) and WNN (Wavenet 1. INTRODUCTION are RNN (Recurrent Neural Network) WNN (Wavenet Neural Network) which address the and slower and faster are RNN (Recurrent Neural Network) and WNN (Wavenet 1. are RNN (Recurrent Neural Network) and WNN (Wavenet 1. INTRODUCTION INTRODUCTION are RNN (Recurrent Neural Network) and WNN (Wavenet Neural Network) which address the slower and faster 1. INTRODUCTION respectively in the underlying process along with Increasing global competition poses a challenge of dynamics Neural Network) which address the slower and faster 1. INTRODUCTION Neural Network) which address the and faster Neural Network) which address the slower slower and faster dynamics respectively in the underlying process along with Increasing global competition poses a challenge of the nonlinear behaviour. optimizing the various process/plant operations in the dynamics respectively in the underlying process along with Increasing global competition poses a challenge of dynamics respectively in the underlying process along Increasing global competition poses aa challenge of respectively in the underlying process along with with Increasing global competition poses challenge of dynamics the nonlinear behaviour. optimizing the various process/plant operations in the industry. The process optimizers usually help in optimizing the nonlinear nonlinear behaviour. behaviour. optimizing the the various various process/plant process/plant operations operations in in the the the optimizing the nonlinear behaviour. optimizing the various process/plant operations in the industry. process optimizers usually optimizing economic The objectives, as consumption raw 2. VIRTUAL SENSING IN VEHICLE POWERTRAIN industry. The process such optimizers usually help helpofin inenergy, optimizing industry. The process optimizers usually help optimizing industry. The process optimizers usually helpofin inenergy, optimizing economic objectives, such as consumption materials, utility etc. For process optimization, it is usually economic objectives, objectives, such such as as consumption consumption of of energy, energy, raw raw 2. 2. VIRTUAL VIRTUAL SENSING SENSING IN IN VEHICLE VEHICLE POWERTRAIN POWERTRAIN economic raw VIRTUAL SENSING IN POWERTRAIN economic objectives, such as consumption of energy, raw 2. 2. VIRTUAL SENSING IN VEHICLE VEHICLE POWERTRAIN materials, utility etc. For process optimization, it is usually required to operate the process at operational and quality Wide band air-fuel ratio sensor (WAFR) is one of the critical materials, utility etc. For process optimization, it is usually materials, utility etc. For process optimization, it is usually materials, utility etc. For process optimization, it is usually required to operate the process at operational and quality Wide band air-fuel ratio sensor (WAFR) is one of critical constraints, which requires tighter quality control. sensors mounted on SI engines in an automotive. Itthe serves as required to operate the process at operational and quality Wide band air-fuel ratio sensor (WAFR) is one of the critical required to operate the at operational and band air-fuel ratio sensor (WAFR) is one of the critical required to which operate the process process at operational and quality quality Wide Wide band air-fuel ratio sensor (WAFR) is one of the critical constraints, requires tighter quality control. sensors mounted on SI engines in an automotive. It serves as the sensor part in the feedback path of the closed loop lambda constraints, which requires tighter quality control. To achieve tighter quality control, it is necessary to have sensors mounted on SI engines in an automotive. It serves as constraints, which requires tighter quality control. sensors mounted on SI engines in an automotive. It serves as constraints, which requires tighter quality control. sensors mounted on SI engines in an automotive. It serves as the sensor part in the feedback path of the closed loop lambda To achieve tighter quality control, it is necessary to have (normalized AFR) control that ensures the quality of online feedback of the crucial quality parameters. Though the sensor part in the feedback path of the closed loop lambda To achieve achieve tighter tighter quality quality control, control, it it is is necessary necessary to to have have the sensor part in the feedback path of the closed loop lambda To the sensor part in the feedback path of the closed loop lambda (normalized AFR) control that ensures the quality of To achieve tighter quality control, it is necessary to have online feedback of the crucial quality parameters. Though combustion and hence the emission levels. However, the online sensors are available for many of the physical (normalized AFR) control that ensures the quality of feedback of the crucial quality parameters. Though (normalized AFR) control that ensures the quality of online feedback of the crucial quality parameters. Though (normalized AFR) control that ensures the quality of combustion and hence the emission levels. However, the online feedback of the crucial quality parameters. Though sensors are available for many of the physical sensor has its limitations such as its operating temperature for variables involved in the process e.g. temperature, pressure, combustion and hence the emission levels. However, the online sensors are available for many of the physical combustion and hence the emission levels. However, the online sensors are available for many of the physical combustion and hence the emission levels. However, the sensor has its limitations such as its operating temperature for online sensors are available for many of the physical variables involved in the process e.g. temperature, pressure, reliable feedback value, operational life and reliability of its flow, velocity, displacement, etc., reliable and pressure, accurate sensor sensor has has its its limitations limitations such such as as its its operating temperature temperature for for variables involved in the process e.g. temperature, variables involved in process e.g. pressure, sensor has its limitations such as its operating operating temperature for reliable feedback value, life reliability its variables involved in the the process e.g. temperature, temperature, pressure, flow, velocity, displacement, etc., reliable and accurate characteristics over its operational operational lifeand (Prokhov et.of al., online sensors for the quality variables are scarce, too costly, reliable feedback value, operational life and reliability of its flow, velocity, displacement, etc., reliable and accurate reliable feedback value, operational life and reliability of its flow, velocity, displacement, etc., reliable and accurate reliable feedback value, operational life and reliability of its characteristics over its operational life (Prokhov et. al., flow, velocity, displacement, etc.,quality reliable andtoois accurate online sensors for the quality variables are scarce, costly, 2005). In the case of its substantial degradation or complete or sometimes less reliable. The control often characteristics over its operational life (Prokhov et. al., online sensors for the quality variables are scarce, too costly, characteristics over its operational life (Prokhov et. al., online sensors for the quality variables are scarce, too costly, characteristics over its operational life (Prokhov et. al., 2005). In the case of its substantial degradation or complete online sensors for thereliable. quality variables are scarce, too special costly, failure or less The control thethe control build around or sensor dependent on the offline laboratory measurements 2005). In In theclosed case of ofloop its substantial substantial degradation orthecomplete complete or sometimes sometimes less reliable. The quality quality controloris is often often 2005). case its degradation or sometimes less reliable. The quality control is often 2005). In thethe case ofloop its substantial degradation orthe complete failure the closed control build around sensor or sometimes less reliable. The that quality controlorand isspecial often dependent on the offline laboratory measurements deprives of feedback and switches over open loop mode test measurements using methods are tedious time failure the closed loop control build around the dependent on on the the offline offline laboratory laboratory measurements measurements or or special special failure the closed loop control build around the sensor sensor dependent failure the closed loop control build around the sensor deprives of the feedback and switches over open loop mode dependent on the offline laboratory measurements or special test measurements using methods that are tedious and time thus causing somewhat inferior lambda control and hence the consuming. When these measurements are made less deprives of the feedback and switches over open loop mode test measurements measurements using using methods methods that that are are tedious tedious and and time time deprives of the feedback and switches over open loop mode test deprives of the feedback and switches over open loop mode thus causing somewhat inferior lambda control and hence the test measurements using methods that are tedious and time consuming. When these measurements are made less emissions. The virtual sensor steps in this situation to make frequently (say at hourly or bihourly rate), tighter quality thus causing somewhat inferior lambda control and hence the consuming. When these measurements are made less thus causing somewhat inferior lambda control and hence the consuming. When these measurements are made less thus causing somewhat inferior lambda control and hence the emissions. The virtual sensor steps in this situation to make consuming. When these measurements are made less frequently (say at hourly or bihourly rate), tighter quality the lambda value available for the control loop and the control becomes difficult to achieve. Soft-sensors or the emissions. The virtual sensor steps in this situation to make frequently (say at hourly or bihourly rate), tighter quality emissions. The virtual sensor steps in this situation to make frequently (say at hourly or bihourly rate), tighter quality emissions. The virtual sensor steps in this situation to make the lambda value available for the control loop and the frequently (say at hourly or bihourly rate), tighter quality control becomes difficult to achieve. Soft-sensors or the vehicle can drive, though compromised on quality a little, to virtual sensors find their application in such scenarios to the lambda value available for the control loop and the control becomes difficult to achieve. Soft-sensors or the the lambda value available for the control loop and the control becomes difficult to achieve. Soft-sensors or the the lambda value available for the control loop and the vehicle can drive, though compromised on quality a little, control becomes difficult to achieve. Soft-sensors or the virtual sensors find their application in such scenarios to the service station to get the sensor replaced. The virtual meet this challenge. We define soft sensors as scenarios models that vehicle can can drive, drive, though though compromised compromised on on quality quality aa little, little, to to virtual sensors find their application in such to vehicle to virtual sensors find their application in such scenarios to vehicle can drive, though compromised on quality a little, to the service station to get the sensor replaced. The virtual virtual sensors find their application in such scenarios to meet this challenge. We define soft sensors as models that sensor running in parallel with the hardware sensor provides make use of less-frequently available measurement values of the service station to get the sensor replaced. The virtual meet this challenge. We define soft sensors as models that the service station to get the sensor replaced. The virtual meet this challenge. We define soft sensors as models that the service station to get the sensor replaced. The virtual sensor running in parallel with the hardware sensor provides meet this challenge. Wesoft-sense define soft as models that make of available measurement values of analytical to monitor health sensor of the hardware quality parameter they andsensors the virtual sensors sensor runningredundancy in parallel parallel with with the hardware hardware sensor provides make use use of less-frequently less-frequently available measurement values as of the sensor running in the provides make use of available measurement values of running in parallel with the hardware sensor provides the analytical redundancy to monitor health of the hardware make use of less-frequently less-frequently available measurement values of sensor quality parameter they soft-sense and the virtual sensors as sensor and also complement it to refine the lambda control. models which estimate it making use of other relevant the analytical redundancy to monitor health of the hardware quality parameter parameter they they soft-sense soft-sense and and the the virtual virtual sensors sensors as as the analytical redundancy to monitor health of the hardware quality the analytical redundancy to monitor health of the hardware sensor and also complement it to refine the lambda control. quality parameter they soft-sense and the virtual sensors as models estimate of plant/process availableuse and sensor and also complement it to refine the lambda control. models which which measurables estimate it it making making use economically of other other relevant relevant sensor models which estimate it use of other sensor and and also also complement complement it it to to refine refine the the lambda lambda control. control. models which measurables estimate it making making use economically of describe other relevant relevant plant/process available and conveniently. The following sections such plant/process measurables available economically and 2.1 RNN for sensing lambda plant/process measurables available economically and plant/process measurables available economically and conveniently. The following sections describe such implementations of the soft/virtual sensorsdescribe which were 2.1 RNN for sensing lambda conveniently. The following sections such conveniently. The following sections describe such RNN for sensing lambda 2.1 RNN for sensing lambda conveniently. The following sections describe such 2.1 implementations of the soft/virtual sensors which were 2.1 RNN for sensing lambda carried out by group of researchers (Kamat et. al., 2005, implementations of the soft/virtual sensors which were For this application Recurrent Neural Network was the implementations of the soft/virtual sensors which were implementations of the soft/virtual sensors which were carried out by group of researchers (Kamat et. al., 2005, 2006) in the stated processes/plants. The structures attempted For this application Recurrent Neural was the carried out by group of researchers (Kamat et. al., 2005, preferred choice over TDNN(Time DelayedNetwork Neural Network) carried out by group of researchers (Kamat et. al., 2005, For this application Recurrent Neural Network was the this application Recurrent Neural Network was the carriedinout by group of researchersThe (Kamat et. al., 2005, For 2006) the stated processes/plants. structures attempted For this application Recurrent Neural Network was the preferred choice over TDNN(Time Delayed Neural Network) 2006) in the stated processes/plants. The structures attempted 2006) in the stated processes/plants. The structures attempted preferred choice over TDNN(Time Delayed Neural Network) preferred choice over TDNN(Time Delayed Neural Network) 2006) in the stated processes/plants. The structures attempted preferred choice over TDNN(Time Delayed Neural Network)
Copyright © 2016 IFAC 100 2405-8963 © 2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Copyright © 2016 IFAC 100 Copyright © 2016 IFAC 100 Peer review under responsibility of International Federation of Automatic Copyright © 2016 IFAC 100Control. Copyright © 2016 IFAC 100 10.1016/j.ifacol.2016.03.036
to capture the nonlinear dynamic character of the lambda phenomenon. A virtual sensor (Kamat et. al., 2006) is designed making use of the other sensed parameters/ actuation parameters at every engine event (control move deployment instant), namely, throttle position (TPS), engine rpm (RPM), manifold absolute pressure (MAP), manifold air flow (MAF), engine coolant temperature(ECT), fuel pulse width (FPW) calculated within control unit as shown in the figure 1. These variables are chosen based on our understanding of the key variables affecting the dynamics of the system.
101
batch learning. The learning algorithm is an approximate gradient descent type that minimizes objective function of NMSE (Normalized Mean Squared Error), E Total (batch type learning)
E Total
1 e( k )T e( k )) 2
…(2)
e(k ) d(k ) y (k )) d (k ) Cx(k )
…(3)
where, d is target vector and C is [1 0 0…] for a MISO system. Using steepest gradient method, the increments in the weight matrix w at any instant k is given by
w( k )
Fig. 1. RNN virtual sensor topology The recurrent neural network topology devised is equivalent to a non-linear state-space model that maps inputs and outputs. The hidden neurons are equivalent to the states. However, the number of neurons could be more than the number of states. The output of the hidden layer is fed back to itself via a bank of unit delays. The dynamic behaviour of the model is given by a pair of coupled equations similar to linear state-space model (Haykin, 1999) as:
Unit delays
1.35
bias Input vector u(k-1)
RNN Virtual sensor Low BW Sensor data
1.3
i 1...N
…(5)
RNN Virtual sensor High BW Sensor data
1.3
1.25
1.25
1.2
1.2
Lambda
Output y(k)
by
1.4
1.4
1.35
State vector x(k-1)
…(4)
The lower bound of the number of nodes in the hidden layer is dependent on the system dynamics. In the actual implementation the nodes chosen were in the range 5-35. After extensive trials a 14 node hidden layer was found adequate. The sensor model is trained using the test bed data from a chosen widely followed American drive cycle as well as specially designed drive pattern which included open loop and closed loop response data with detuned and tightly tuned controller and special test data at two select gears (Kamat et. al., 2006 (2)). Event based samples were used in preference to time based samples to correspond closer to the engine operation at all operating conditions. A two step approach has been followed for training with an initial training of forward path of RNN followed by a training of the full RNN with feedback links restored. The figure 3 shows results for such a pattern where the lambda signal has comparatively low and high band widths respectively.
z
z-1
w( k )
T T T xi (n) ( wix x(n 1) wiu u (n 1) wib xb ) y (n) Cx(n)
x(k ) f ( x(k 1), u(k 1), Wxx, Wxu, Wxb); …(1) y(k) g(x(k)) Cx(k); where, is a nonlinear tansigmoidal function vector characterizing the hidden layer and C is the vector of synaptic weights characterizing the output . The output is tapped at the output node, which itself is the one of the nodes of the hidden layer (Fig. 2). The training algorithm adopted is based on the algorithm developed by Williams and Zipser (Williams, et. al.) and popularly known as RTRL (Real Time Recurrent Learning). -1 z-1
ETotal ( k )
where, η is learning rate of sufficiently small value. The output is updated
Lambda
IFAC ACODS 2016 February 1-5, 2016. NIT Tiruchirappalli, India Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105
1.15
1.1
1.15
1.1
1.05
1.05 1
1 1
0.95
0.9 1000
1500
2000
2500
3000
Engine events
3500
4000
0.95
4500 3.68
0.9 1000
3.7
3.72 1500
3.74 3.76 Engine Events 2000
2500
3.78
3.8
3.82 4
x 10 3000
3500
4000
Engine events
Fig. 2. RNN virtual sensor topology
Fig. 3. Lambda : Sensor and RNN virtual sensor outputs (low and high bandwidth signals)
The weights of a fully connected recurrent network are adjusted in real time using either a supervised continuous or a
In similar manner the RNN virtual sensing was performed for the parameters, Manifold Absolute Pressure (inputs: TPS, RPM, MAF, FPW), Manifold Air Flow (inputs: TPS, RPM, 101
4500
IFAC ACODS 2016 102 Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105 February 1-5, 2016. NIT Tiruchirappalli, India
MAP, FPW), Engine Torque (inputs: TPS, RPM, MAP, FPW, ECT, Ignition angle) and the results are depicted in the figures 4,5. The torque is sensed on the test bed set and not available as a sensed value on-board the vehicle. All values are normalized in the range [-1 1] and displayed. These sensors are tuned independently. It is observed that RNN was good enough to capture these signals as they are limited bandwidth signals. It is obvious that all of these virtual sensors do not work at the same time, meaning only one could perform at a time. It can be seen that RNN with tansigmoial functions for all the nodes does not seem to have the discriminatory feature to capture with equal precision both the low frequency and high frequency components of the lambda response.
x n x
x2 e 2
y
Sensor data RNN Virtual sensor
k k
100
500
900
1300
Engine Events
RNN Virtual sensor Sensor data
i
it
b
The resulting wavenet topology is shown in Figure 6. With the initial network developed as stated above, optimal weights were obtained by minimization of MSE based on cumulative error, ETotal. With the weights available, the initial network is pruned using Akaike information criteria. For the pruned network, optimal values of the dilation and translational parameters, network weights and the bias were obtained by solving the Least Squares problem using Levenberg- Marquadt method.
0
5500
it
i 0 t1
…(8) where, X is the input vector; Dj, Di = Dilation vectors associated with scalon at level j and wavelon at level i respectively; NTj = number of translation vectors associated with scalon at level j; NTi = number of translation vectors associated with wavelon at level i; Tjs = Translation vector associated with scalon at level j, Tit = translation vector associated with wavelon at level i; b = bias; NM is chosen as 1 for the wavenet structure.
1700
1.0
5000
js
k1
Fig. 4. MAP and MAF: Sensor and RNN virtual sensor outputs
4500
j
n
Engine Events
-0.4
js
c x -1
-1
N L 1 NTi
w [D (X T )} w [(D (X T )] j 0 s1
MAF Normalized
MAP Normalized
Sensor data RNN Virtual sensor
… (7)
( x) Wavenet network has a feedforward topology with one hidden layer consisting of scalons and wavelons . The data points were examined at various scales and at each scale the distribution of data was examined to arrive at the initial set of translation vectors. In the present case a scale level NL=5 was found adequate. The resultant equation describing the wavenet is described by the equation N M 1 NT j
1
for n dimensions
x2 2e 2
1
Engine Torque Normalized
2
6000
6500
7000
7500
Engine events
Fig. 5. Torque: Sensor and RNN virtual sensor outputs 2.2 WNN for lambda sensing To overcome the deficiency of RNN, an alternate WNN topology (Kamat et. al., 2006) was proposed for this purpose. This structure is essentially a feed-forward topology, with the addition of time delays to bring in the dynamics into the model. The wavelet transforms (Bakshi et. al., 1993; Miller, 1998) could be adopted as activation functions to impart the ability to capture the local behaviour in addition to the mean (global) behaviour. A family of wavelets is derived from the translations and dilations of a single function x , called mother wavelet and is represented by
x u for s s
1
s, u R 2
Fig. 6. WNN topology (t: translations, d: dilations, w: weights of wavelons, c: feed through weights, b: bias, first node φ in layer is scalon)
…(6) The wavenet (WNN) trained has shown promising results as depicted in figure 7. In the lambda figures the line at magnitude 1.0 denotes stoichiometric combustion for which lambda =1.0, meaning the combustion is closer to stoichiometry.
The wavelet family selected for this exercise is Mexican hat which is a derivative of Gaussian function. The expressions for scalar function and vector mother wavelets are
102
IFAC ACODS 2016 February 1-5, 2016. NIT Tiruchirappalli, India Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105
4. SOFT / VIRTUAL PROCESSING
1.2
W NN Virtual sensor Sensor data
Lambda
1.15
1.1
103
SENSING
IN
MATERIAL
2.1 Soft sensing of quality parameter: Cement blaine
stoich 1.05
The cement manufacturing plant comprises of cement mill at its tail end that grinds cement clinkers (output from cement kiln) into finest powder which is the final product. The fineness of the product is determined by means of a parameter called blaine and is measured on sample basis at regular intervals at laboratory. While the lab measured values are available at intervals such as 1 hour, the control inputs to the cement mill are deployed at every 1 minute, or so. This justifies the need for a soft sensor that could estimate the blaine value at the sample time interval required by the mill control system. In this work the plant considered is rotary ball mill that grinds the clinkers into final product. The inputs to the plant are feed rate, separator fan rpm and power, motor drive power, accumulation of the material in the mill shell which are available at faster rate than the lab measurement. A time-unfolded feed-forward ANN is built using the past data of the mill available from the SCADA that governs the mill operation. All the inputs are posed as individual time-delayed inputs and the network is trained in supervisory mode using lab data as target available at rate Ts. This calls for substantial amount of test data. A generalization of this network is carried out at the rate ts by progressively replacing the inputs in such a manner that the u(k-1) is replaced by u(k) and so forth (Fig 9). This produces output value at the faster sample rate (ts << Ts ). This procedure is repeated until substantial data points are generated by this trained network.
1
0.95 0.96
0.98
1
1.02
1.04
1.06
1.08
1.1 4
Engine events
x 10
Fig. 7. Lambda: Sensor and WNN virtual sensor outputs (High BW) In these implementations the measurement noise is not explicitly addressed. 3. SOFT SENSING IN ELECTRIFIED VEHICLES Hybrid electric and all electric vehicles are going to be the future vehicles for known environmental reasons. These vehicles have Li-ion battery as electrical energy storage element of their powertrain. Essentially, the battery pack is composed of a string of series-parallel arrangement of basic cell in order to achieve net bus voltage and Ah capacity of the pack. The control and management of the battery parameters from its performance, life, diagnostics and safety points of view, the battery management systems have come into existence and are undergoing continuous improvement. Among many other vital parameters the state of charge (SoC) of the battery is very crucial to know about the battery’s status of holding the energy and hence the electric range it can travel safely. In the absence of an on-board sensor for SOC, this property needs to be inferred using various approaches for soft/ virtually sensing it. The input parameters are battery terminal voltage, battery current (both charge and discharge), and temperature at various cells in the pack. n this work the RNN and WNN topology are trained in a supervised mode using the the SoC reading obtained from the existing on-board SoC estimator. The topology is further assisted by making use of SoC reading made available from specially conducted test carried out at decided time intervals which are very few in number. Thus it is a soft sensor rather than a true virtual sensor. The results are shown in figure 8. The RNN shows more noise immunity than WNN.
A fresh ANN is now built with single time dimensioned inputs (Fig 9.c) and the generated target variable values, both of which are now available at faster rate. The target values at the lab measurement instant are the values obtained from the lab which are true values. The input dimensionality is now minimal and the network is re-trained using RNN to address the dynamics. The RTRL algorithm in teacher forcing mode is adopted to accommodate further lab measurement values. The method has element of redundancy and it is observed that, the lesser the interval between lab readings, better is the accuracy of soft sensor. The figure 10 shows results of the blaine soft sensor used as feedback element for the multi variable mill control where the controller is supposed to achieve the blaine within specified limit. . The soft-sensed blaine values under the closed loop control show acceptable quality output.
Existing SoC estimator WNN Virtual sensor
2.2 Virtual sensing of burning zone temperature in cement kiln 0
5000
10000
15000
20000
The ground raw mix ore is pyro-processed in a huge rotary inclined shell cement kiln. This mix is added from the top and a burner fires the pulverized coal powder mixed with fresh air flow to raise the temperature of the order 900 to 1400 deg C for the clinckerization reactions to take place in the zone called burning zone. This temperature is usually sensed by IR camera and faces limitations due to interruption
30000
Sample number
Fig. 8. SoC of Li-ion battery : Estimator, RNN and WNN virtual sensor outputs
103
IFAC ACODS 2016 104 Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105 February 1-5, 2016. NIT Tiruchirappalli, India
5. VIRTUAL SENSING IN CHEMICAL PROCESS (pH CONTROL)
.
. u1(k-2) u1(k-1)
FF ANN @Ts
. u2(k-2) u2(k-1)
u1(k-1) y(k) u1(k)
FF ANN @ Ts
. u2(k-1) u2(k)
(a) FF- ANN in training
A pH processes is highly nonlinear, may exhibit varying characteristics depending on the character of the input and conditions of operation. A typical characteristics may be drastic increase in process gain over a small region around operating point causing sharp transients as the operating point is approached from low gain regions. Another possible manifestation could be variations of steady state profile due to changes in character of input. In the case of pH measurement, there could be loss of sensor signal due to frequent fouling. These issues pose considerable challenge in developing a framework and methodology for such processes. The process model considered here is taken from Henson and Seborg (Henson et. al., 1994). It consists of an acid stream (q1), buffer stream (q2) and base stream (q3) that are mixed in a tank (Fig 12). The system involves 3 state variables including the reaction invariants of the effluent stream and height in the tank. The model also includes valve (unity gain, time constant p ) and transmitter dynamics (unity gain, time constant h) as well as hydraulic relationships for the tank outlet flow. The effluent pH is measured by probe (with time delay Δt) located downstream from the tank.
(b) FF- ANN in mode generalization mode y(k)
u1(k-1)
RNN @ts
u2(k-1)
(c) RNN trained using data generated by (b), ts<< Ts Fig. 9 Slower rate FF-ANN to faster rate RNN
H
Fig. 10 Cement blaine soft sensor output Fig. 12 pH (with buffer) neutralization scheme 1350
1300
In the case of RNN based virtual sensor (Kamat et. al., 2005) the size of training matrix required is high. In case where both buffer and base are perturbed, larger matrix is required. Dead time is accounted by delaying the inputs appropriately. Two lagged inputs at (k-1) and (k-2) are found adequate to capture dynamics for prediction of output at (k). Network is trained for a number of iterations by reducing the learning rate () for successive reductions in NMSE. The figure 13 shows results for randomly perturbed q3 flow whereas a steady state pH-q3 profile is shown in figure 14.
BZT deg C
1250
1200
1150
1100
1050
Sensed temperature W NN virtual sensor
1000
950
0
100
200
300 Sample number
400
500
600
Fig. 11 Cement kiln BZT virtual sensor In case of WNN based virtual sensor the training data and validation data used are the same as that in case of RNN. The wavenet being a static mapping of inputs and outputs, the dynamics is accounted by unfolding the inputs. The size of training data required is about one fifth for case when only the base flow is perturbed. In the case where both buffer and base flows are perturbed, a comparable size of training data is required. The figure 15 shows results for randomly perturbed q3 flow whereas a steady state pH-q3 profile is shown in figure 16.
of the view of the flame due to material turbulence. This led to complement the temperature measurement by a virtual sensor. Some of the crucial inputs used for this are coal feed, kiln rpm and power, material feed and feed material temperature, secondary (hot) air temperature, etc. The results of WNN sensor built using the aforementioned approach are shown in the figure 11.
104
IFAC ACODS 2016 February 1-5, 2016. NIT Tiruchirappalli, India Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105
Both the virtual sensors are be able to capture the trend in steady state profile reasonably well, while the ability to capture the dynamics is comparable.
to processes whose characteristics may range from mildly nonlinear to high nonlinearity. The networks also have the inherent flexibility to map operational conditions from near steady to transients of varied spectral character. The design approach of such structures has also been evolved using a careful analysis of system characteristics and operating data. The WNN has greater degrees of freedom in shaping the responses by the virtue of dilation and translation and the choice of the wavelet function. In the case of RNN for the same process, the dimension of input vector is relatively smaller since dynamic features are embedded in the RNN structure itself unlike in the WNN case. The virtual sensors augment the sensing part of a control system of a process and help in reducing the cost and time involved in the precise real life measurement of its crucial parameters.
11
10
pH value
9
8
7
6
5
Sensed pH data RNN virtual pH sensor
4
3 1000
1050
1100
1150
1200
1250
1300
Sample number
ACKNOWLEDGEMENT
Fig. 13 pH: Sensor and RNN virtual sensor outputs
The authors acknowledge the support from the management of TCS limited and participation of the former TCS associates, Jessy Smith, Vivek Diwanji, and Hossein Javaherian of General Motors while carrying out some of the reported work.
14
pH steady state values
12
10
Sensed pH data RNN virtual pH sensor
8
6
105
4
2
REFERENCES
0 4.4
4.6
4.8
5
5.2
5.4
5.6
5.8
base fluid flow perturbation inml/s
B E Miller, “Wavelet Networks for Characterization and Implementation of Environments for Haptic Display.” Masters Thesis, Northwestern University, 1998. Bhavik R. Bakshi, George Stephanopoulos,Wave-Net: a Multiresolution, Hierarchical Neural Network with Localized Learning. AIChE Journal, Jan1993,Vol.39, No.1. M. A. Henson and D. E. Seborg, “Adaptive Nonlinear Control of a pH Neutralization Process”. IEEE Trans. on Control Systems Technology, 2, 169, 1994. Simon Haykin, Neural Networks – A Comprehensive Foundation, (Pearson Education, 1999). Shivaram Kamat, Vivek Diwanji, Jessy George Smith, K. P. Madhavan, ‘’Modeling of pH Process Using Recurrent Neural Network and Wavenet”, CIMSA 2005 – IEEE International Conference on Computational Intelligence for Measurement Systems and Applications Giardini Naxos, Italy, 20-22 July 2005. Shivaram Kamat, V Diwanji, Jessy Smith, Hossein Javaherian, KP Madhavan, “Virtual Sensing of SI Engines Using Recurrent Neural Networks”, SAE 200601-1348. Shivaram S. Kamat, Hossein Javaherian, Vivek V. Diwanji, Jessy G. Smith, and K. P. Madhavan, “Virtual Air-Fuel Ratio Sensors for Engine Control and Diagnostics”, American Control Conference 2006 Williams, R. J, D. Zipser, A learning Algorithm for Continually Running Fully R ecurrent Neural Networks, Neural Computation, 1, 270-280. D. Prokhov, Virtual Sensors and Their Automotive Applications, Intelligent Sensors, Sensor Networks and Information Processing Conference,2005, 411-416
Fig. 14 pH steady state values comparison 10 9.5 9
pH value
8.5 8 7.5 7 6.5
Sensed pH data WNN virtual pHsensor
6 5.5
1000
1200
1400
1600
1800
2000
Sample number
Fig. 15 pH : Sensor and WNN virtual sensor outputs
14
pH steady state values
12
10
Sensed pH data W NN virtual pH sensor
8
6
4
2
0 4.4
4.6
4.8
5
5.2
5.4
5.6
5.8
base fluid flow perturbation in ml/s
Fig. 16 pH steady state values comparison 6. CONCLUSIONS The paper has reported the feasibility of developing soft/virtual sensors using RNN and WNN. The two structures were shown to complement each other and could be applied
105
ScienceDirect
IFAC-PapersOnLine (2016)Sensors 100–105 Developing ANN based Virtual 49-1 / Soft for Industrial Problems Developing ANN based Virtual / Soft Sensors for Industrial Problems Developing ANN based Virtual / Soft Sensors for Industrial Problems Developing ANN basedShivaram VirtualKamat* / Soft Sensors for Industrial Problems KP Madhavan**
Shivaram Shivaram Kamat* Kamat* KP KP Madhavan** Madhavan** Shivaram Kamat* Madhavan** KP Shivaram Kamat* Madhavan** *SMIEEE, Engineering & Industrial Services, Tata Consultancy Services Limited, Pune, 411006 KP *SMIEEE, Engineering & Industrial Services, Tata Consultancy Services India (Tel: 20-6608-7528; e-mail: [email protected]). *SMIEEE, Engineering & Industrial Services, Tata Consultancy Services Limited, Limited, Pune, Pune, 411006 411006 *SMIEEE, & Services, Tata Consultancy *SMIEEE, Engineering Engineering & Industrial Industrial Services, Tata [email protected]). Consultancy Services Services Limited, Limited, Pune, Pune, 411006 411006 India (Tel: 20-6608-7528; e-mail: India (Tel: (Tel: 20-6608-7528; 20-6608-7528; e-mail: e-mail: [email protected]). [email protected]). India India (Tel:Engineering, 20-6608-7528; e-mail: [email protected]). ** Department of Chemical Indian Institute of Technology Bombay, Mumbai, 400076 ** ** Department Department of of Chemical Chemical Engineering, Engineering, Indian Indian Institute Institute of of Technology Technology Bombay, Bombay, Mumbai, Mumbai, 400076 400076 ** Engineering, Indian Institute ** Department Department of of Chemical Chemical India Engineering, Institute of of Technology Technology Bombay, Bombay, Mumbai, Mumbai, 400076 400076 (e-mail: Indian [email protected]). India (e-mail: [email protected]). India (e-mail: [email protected]). India India (e-mail: (e-mail: [email protected]). [email protected]). Abstract: This paper briefly reports experience of our research group in developing and deploying some Abstract: This paper briefly briefly reports experience of of ourcrucial research group in in for developing and deploying deploying some promising This approaches for virtual andexperience soft sensing parameters use in control systems some of a Abstract: paper reports of our research group developing and Abstract: This paper briefly reports experience of our research group in developing and some Abstract: This paper briefly reports experience of ourcrucial research group in for developing and deploying deploying some promising approaches for virtual and soft sensing of parameters use in control systems of byaa process or plant. It briefs on the constraints and limitations of the measurement of crucial parameters promising approaches approaches for for virtual virtual and and soft soft sensing sensing of of crucial crucial parameters parameters for for use use in in control control systems systems of of promising a promising approaches for on virtual andand softviability sensing ofgocrucial parameters for use The in crucial control systems ofare process or plant. It briefs the constraints and limitations of the measurement of parameters bya physical means and hence the need to for virtual/soft sensing.. approaches used process or or plant. plant. It It briefs briefs on on the the constraints constraints and and limitations limitations of of the the measurement measurement of of crucial crucial parameters parameters by by process process plant. It briefs on constraints and limitations ofsupervised the measurement of crucial parameters by physicalor means and hence thethe need and topologies viability to and go for for virtual/soft sensing.. The approaches used are variants of Artificial Neural Network their and partly supervised training physical means and hence the need and viability to go virtual/soft sensing.. The approaches used are physical means and hence the need and viability to go for virtual/soft sensing.. The approaches used are physical means and hence the need and viability to go for virtual/soft sensing.. The approaches used are variants of Artificial Neural Network topologies and their supervised and partly supervised training algorithms. The paperNeural provides a brieftopologies overview of virtual sensor development based ontraining these variants of Artificial Artificial Neural Network topologies andthe their supervised and partly partly supervised supervised training variants of Network and their supervised and variants of for Artificial Network andthe their supervised andjustify partlythe supervised algorithms. Theselected paperNeural provides a brief brieftopologies overview of virtual sensor to development based on ontraining these approaches process situations and provides validation results approaches used. algorithms. The paper provides a overview of the virtual sensor development based these algorithms. The paper provides aa brief overview of the virtual sensor development based on these algorithms. The paper provides brief overview of the virtual sensor development based on these approaches for selected process situations and provides validation results to justify the approaches used. The industryfor sectors forprocess which these solutions were developed areresults automotive, cement grinding process approaches selected situations and provides validation to justify the approaches used. approaches for selected process situations and provides validation results to the approaches used. approaches forsectors selected process situations and provides validation to justify justify thebeen approaches used. The industry for which process these solutions were developed developed areresults automotive, cement grinding process and kiln process, and chemical for pH control. Some of these approaches have implemented The industry sectors for which these solutions were are automotive, cement grinding process The industry sectors for which these solutions were developed are automotive, cement grinding process The industry sectors for which these solutions were developed are automotive, cement grinding process and kiln process, and chemical process for pH control. Some of these approaches have been implemented in the industry underlining the significant the virtual/soft mechanisms could play in process and kiln process, and chemical process forrole pH control. Some ofsensing these approaches have been implemented and kiln process, and process pH Some these have been implemented and kiln process, and chemical chemical process for forrole pH control. control. Some of ofsensing these approaches approaches have been implemented in the industry underlining the significant significant the virtual/soft virtual/soft mechanisms could play in process process operations. in the industry underlining the role the the sensing mechanisms mechanisms could play in in the industry underlining the significant role virtual/soft sensing could play in in the industry underlining the significant role the virtual/soft sensing mechanisms could play in process process operations. operations. operations. © 2016, IFAC (International Federation Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Keywords: virtual sensing, soft sensing,ofartificial neural networks, wavelets, wavenets operations. Keywords: virtual sensing, soft sensing, artificial neural networks, wavelets, wavenets Keywords: networks, Keywords: virtual virtual sensing, sensing, soft soft sensing, sensing, artificial artificial neural neural networks, wavelets, wavelets, wavenets wavenets Keywords: virtual sensing, soft sensing, artificial neural networks, wavelets, wavenets are RNN (Recurrent Neural Network) and WNN (Wavenet 1. INTRODUCTION are RNN (Recurrent Neural Network) WNN (Wavenet Neural Network) which address the and slower and faster are RNN (Recurrent Neural Network) and WNN (Wavenet 1. are RNN (Recurrent Neural Network) and WNN (Wavenet 1. INTRODUCTION INTRODUCTION are RNN (Recurrent Neural Network) and WNN (Wavenet Neural Network) which address the slower and faster 1. INTRODUCTION respectively in the underlying process along with Increasing global competition poses a challenge of dynamics Neural Network) which address the slower and faster 1. INTRODUCTION Neural Network) which address the and faster Neural Network) which address the slower slower and faster dynamics respectively in the underlying process along with Increasing global competition poses a challenge of the nonlinear behaviour. optimizing the various process/plant operations in the dynamics respectively in the underlying process along with Increasing global competition poses a challenge of dynamics respectively in the underlying process along Increasing global competition poses aa challenge of respectively in the underlying process along with with Increasing global competition poses challenge of dynamics the nonlinear behaviour. optimizing the various process/plant operations in the industry. The process optimizers usually help in optimizing the nonlinear nonlinear behaviour. behaviour. optimizing the the various various process/plant process/plant operations operations in in the the the optimizing the nonlinear behaviour. optimizing the various process/plant operations in the industry. process optimizers usually optimizing economic The objectives, as consumption raw 2. VIRTUAL SENSING IN VEHICLE POWERTRAIN industry. The process such optimizers usually help helpofin inenergy, optimizing industry. The process optimizers usually help optimizing industry. The process optimizers usually helpofin inenergy, optimizing economic objectives, such as consumption materials, utility etc. For process optimization, it is usually economic objectives, objectives, such such as as consumption consumption of of energy, energy, raw raw 2. 2. VIRTUAL VIRTUAL SENSING SENSING IN IN VEHICLE VEHICLE POWERTRAIN POWERTRAIN economic raw VIRTUAL SENSING IN POWERTRAIN economic objectives, such as consumption of energy, raw 2. 2. VIRTUAL SENSING IN VEHICLE VEHICLE POWERTRAIN materials, utility etc. For process optimization, it is usually required to operate the process at operational and quality Wide band air-fuel ratio sensor (WAFR) is one of the critical materials, utility etc. For process optimization, it is usually materials, utility etc. For process optimization, it is usually materials, utility etc. For process optimization, it is usually required to operate the process at operational and quality Wide band air-fuel ratio sensor (WAFR) is one of critical constraints, which requires tighter quality control. sensors mounted on SI engines in an automotive. Itthe serves as required to operate the process at operational and quality Wide band air-fuel ratio sensor (WAFR) is one of the critical required to operate the at operational and band air-fuel ratio sensor (WAFR) is one of the critical required to which operate the process process at operational and quality quality Wide Wide band air-fuel ratio sensor (WAFR) is one of the critical constraints, requires tighter quality control. sensors mounted on SI engines in an automotive. It serves as the sensor part in the feedback path of the closed loop lambda constraints, which requires tighter quality control. To achieve tighter quality control, it is necessary to have sensors mounted on SI engines in an automotive. It serves as constraints, which requires tighter quality control. sensors mounted on SI engines in an automotive. It serves as constraints, which requires tighter quality control. sensors mounted on SI engines in an automotive. It serves as the sensor part in the feedback path of the closed loop lambda To achieve tighter quality control, it is necessary to have (normalized AFR) control that ensures the quality of online feedback of the crucial quality parameters. Though the sensor part in the feedback path of the closed loop lambda To achieve achieve tighter tighter quality quality control, control, it it is is necessary necessary to to have have the sensor part in the feedback path of the closed loop lambda To the sensor part in the feedback path of the closed loop lambda (normalized AFR) control that ensures the quality of To achieve tighter quality control, it is necessary to have online feedback of the crucial quality parameters. Though combustion and hence the emission levels. However, the online sensors are available for many of the physical (normalized AFR) control that ensures the quality of feedback of the crucial quality parameters. Though (normalized AFR) control that ensures the quality of online feedback of the crucial quality parameters. Though (normalized AFR) control that ensures the quality of combustion and hence the emission levels. However, the online feedback of the crucial quality parameters. Though sensors are available for many of the physical sensor has its limitations such as its operating temperature for variables involved in the process e.g. temperature, pressure, combustion and hence the emission levels. However, the online sensors are available for many of the physical combustion and hence the emission levels. However, the online sensors are available for many of the physical combustion and hence the emission levels. However, the sensor has its limitations such as its operating temperature for online sensors are available for many of the physical variables involved in the process e.g. temperature, pressure, reliable feedback value, operational life and reliability of its flow, velocity, displacement, etc., reliable and pressure, accurate sensor sensor has has its its limitations limitations such such as as its its operating temperature temperature for for variables involved in the process e.g. temperature, variables involved in process e.g. pressure, sensor has its limitations such as its operating operating temperature for reliable feedback value, life reliability its variables involved in the the process e.g. temperature, temperature, pressure, flow, velocity, displacement, etc., reliable and accurate characteristics over its operational operational lifeand (Prokhov et.of al., online sensors for the quality variables are scarce, too costly, reliable feedback value, operational life and reliability of its flow, velocity, displacement, etc., reliable and accurate reliable feedback value, operational life and reliability of its flow, velocity, displacement, etc., reliable and accurate reliable feedback value, operational life and reliability of its characteristics over its operational life (Prokhov et. al., flow, velocity, displacement, etc.,quality reliable andtoois accurate online sensors for the quality variables are scarce, costly, 2005). In the case of its substantial degradation or complete or sometimes less reliable. The control often characteristics over its operational life (Prokhov et. al., online sensors for the quality variables are scarce, too costly, characteristics over its operational life (Prokhov et. al., online sensors for the quality variables are scarce, too costly, characteristics over its operational life (Prokhov et. al., 2005). In the case of its substantial degradation or complete online sensors for thereliable. quality variables are scarce, too special costly, failure or less The control thethe control build around or sensor dependent on the offline laboratory measurements 2005). In In theclosed case of ofloop its substantial substantial degradation orthecomplete complete or sometimes sometimes less reliable. The quality quality controloris is often often 2005). case its degradation or sometimes less reliable. The quality control is often 2005). In thethe case ofloop its substantial degradation orthe complete failure the closed control build around sensor or sometimes less reliable. The that quality controlorand isspecial often dependent on the offline laboratory measurements deprives of feedback and switches over open loop mode test measurements using methods are tedious time failure the closed loop control build around the dependent on on the the offline offline laboratory laboratory measurements measurements or or special special failure the closed loop control build around the sensor sensor dependent failure the closed loop control build around the sensor deprives of the feedback and switches over open loop mode dependent on the offline laboratory measurements or special test measurements using methods that are tedious and time thus causing somewhat inferior lambda control and hence the consuming. When these measurements are made less deprives of the feedback and switches over open loop mode test measurements measurements using using methods methods that that are are tedious tedious and and time time deprives of the feedback and switches over open loop mode test deprives of the feedback and switches over open loop mode thus causing somewhat inferior lambda control and hence the test measurements using methods that are tedious and time consuming. When these measurements are made less emissions. The virtual sensor steps in this situation to make frequently (say at hourly or bihourly rate), tighter quality thus causing somewhat inferior lambda control and hence the consuming. When these measurements are made less thus causing somewhat inferior lambda control and hence the consuming. When these measurements are made less thus causing somewhat inferior lambda control and hence the emissions. The virtual sensor steps in this situation to make consuming. When these measurements are made less frequently (say at hourly or bihourly rate), tighter quality the lambda value available for the control loop and the control becomes difficult to achieve. Soft-sensors or the emissions. The virtual sensor steps in this situation to make frequently (say at hourly or bihourly rate), tighter quality emissions. The virtual sensor steps in this situation to make frequently (say at hourly or bihourly rate), tighter quality emissions. The virtual sensor steps in this situation to make the lambda value available for the control loop and the frequently (say at hourly or bihourly rate), tighter quality control becomes difficult to achieve. Soft-sensors or the vehicle can drive, though compromised on quality a little, to virtual sensors find their application in such scenarios to the lambda value available for the control loop and the control becomes difficult to achieve. Soft-sensors or the the lambda value available for the control loop and the control becomes difficult to achieve. Soft-sensors or the the lambda value available for the control loop and the vehicle can drive, though compromised on quality a little, control becomes difficult to achieve. Soft-sensors or the virtual sensors find their application in such scenarios to the service station to get the sensor replaced. The virtual meet this challenge. We define soft sensors as scenarios models that vehicle can can drive, drive, though though compromised compromised on on quality quality aa little, little, to to virtual sensors find their application in such to vehicle to virtual sensors find their application in such scenarios to vehicle can drive, though compromised on quality a little, to the service station to get the sensor replaced. The virtual virtual sensors find their application in such scenarios to meet this challenge. We define soft sensors as models that sensor running in parallel with the hardware sensor provides make use of less-frequently available measurement values of the service station to get the sensor replaced. The virtual meet this challenge. We define soft sensors as models that the service station to get the sensor replaced. The virtual meet this challenge. We define soft sensors as models that the service station to get the sensor replaced. The virtual sensor running in parallel with the hardware sensor provides meet this challenge. Wesoft-sense define soft as models that make of available measurement values of analytical to monitor health sensor of the hardware quality parameter they andsensors the virtual sensors sensor runningredundancy in parallel parallel with with the hardware hardware sensor provides make use use of less-frequently less-frequently available measurement values as of the sensor running in the provides make use of available measurement values of running in parallel with the hardware sensor provides the analytical redundancy to monitor health of the hardware make use of less-frequently less-frequently available measurement values of sensor quality parameter they soft-sense and the virtual sensors as sensor and also complement it to refine the lambda control. models which estimate it making use of other relevant the analytical redundancy to monitor health of the hardware quality parameter parameter they they soft-sense soft-sense and and the the virtual virtual sensors sensors as as the analytical redundancy to monitor health of the hardware quality the analytical redundancy to monitor health of the hardware sensor and also complement it to refine the lambda control. quality parameter they soft-sense and the virtual sensors as models estimate of plant/process availableuse and sensor and also complement it to refine the lambda control. models which which measurables estimate it it making making use economically of other other relevant relevant sensor models which estimate it use of other sensor and and also also complement complement it it to to refine refine the the lambda lambda control. control. models which measurables estimate it making making use economically of describe other relevant relevant plant/process available and conveniently. The following sections such plant/process measurables available economically and 2.1 RNN for sensing lambda plant/process measurables available economically and plant/process measurables available economically and conveniently. The following sections describe such implementations of the soft/virtual sensorsdescribe which were 2.1 RNN for sensing lambda conveniently. The following sections such conveniently. The following sections describe such RNN for sensing lambda 2.1 RNN for sensing lambda conveniently. The following sections describe such 2.1 implementations of the soft/virtual sensors which were 2.1 RNN for sensing lambda carried out by group of researchers (Kamat et. al., 2005, implementations of the soft/virtual sensors which were For this application Recurrent Neural Network was the implementations of the soft/virtual sensors which were implementations of the soft/virtual sensors which were carried out by group of researchers (Kamat et. al., 2005, 2006) in the stated processes/plants. The structures attempted For this application Recurrent Neural was the carried out by group of researchers (Kamat et. al., 2005, preferred choice over TDNN(Time DelayedNetwork Neural Network) carried out by group of researchers (Kamat et. al., 2005, For this application Recurrent Neural Network was the this application Recurrent Neural Network was the carriedinout by group of researchersThe (Kamat et. al., 2005, For 2006) the stated processes/plants. structures attempted For this application Recurrent Neural Network was the preferred choice over TDNN(Time Delayed Neural Network) 2006) in the stated processes/plants. The structures attempted 2006) in the stated processes/plants. The structures attempted preferred choice over TDNN(Time Delayed Neural Network) preferred choice over TDNN(Time Delayed Neural Network) 2006) in the stated processes/plants. The structures attempted preferred choice over TDNN(Time Delayed Neural Network)
Copyright © 2016 IFAC 100 2405-8963 © 2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Copyright © 2016 IFAC 100 Copyright © 2016 IFAC 100 Peer review under responsibility of International Federation of Automatic Copyright © 2016 IFAC 100Control. Copyright © 2016 IFAC 100 10.1016/j.ifacol.2016.03.036
to capture the nonlinear dynamic character of the lambda phenomenon. A virtual sensor (Kamat et. al., 2006) is designed making use of the other sensed parameters/ actuation parameters at every engine event (control move deployment instant), namely, throttle position (TPS), engine rpm (RPM), manifold absolute pressure (MAP), manifold air flow (MAF), engine coolant temperature(ECT), fuel pulse width (FPW) calculated within control unit as shown in the figure 1. These variables are chosen based on our understanding of the key variables affecting the dynamics of the system.
101
batch learning. The learning algorithm is an approximate gradient descent type that minimizes objective function of NMSE (Normalized Mean Squared Error), E Total (batch type learning)
E Total
1 e( k )T e( k )) 2
…(2)
e(k ) d(k ) y (k )) d (k ) Cx(k )
…(3)
where, d is target vector and C is [1 0 0…] for a MISO system. Using steepest gradient method, the increments in the weight matrix w at any instant k is given by
w( k )
Fig. 1. RNN virtual sensor topology The recurrent neural network topology devised is equivalent to a non-linear state-space model that maps inputs and outputs. The hidden neurons are equivalent to the states. However, the number of neurons could be more than the number of states. The output of the hidden layer is fed back to itself via a bank of unit delays. The dynamic behaviour of the model is given by a pair of coupled equations similar to linear state-space model (Haykin, 1999) as:
Unit delays
1.35
bias Input vector u(k-1)
RNN Virtual sensor Low BW Sensor data
1.3
i 1...N
…(5)
RNN Virtual sensor High BW Sensor data
1.3
1.25
1.25
1.2
1.2
Lambda
Output y(k)
by
1.4
1.4
1.35
State vector x(k-1)
…(4)
The lower bound of the number of nodes in the hidden layer is dependent on the system dynamics. In the actual implementation the nodes chosen were in the range 5-35. After extensive trials a 14 node hidden layer was found adequate. The sensor model is trained using the test bed data from a chosen widely followed American drive cycle as well as specially designed drive pattern which included open loop and closed loop response data with detuned and tightly tuned controller and special test data at two select gears (Kamat et. al., 2006 (2)). Event based samples were used in preference to time based samples to correspond closer to the engine operation at all operating conditions. A two step approach has been followed for training with an initial training of forward path of RNN followed by a training of the full RNN with feedback links restored. The figure 3 shows results for such a pattern where the lambda signal has comparatively low and high band widths respectively.
z
z-1
w( k )
T T T xi (n) ( wix x(n 1) wiu u (n 1) wib xb ) y (n) Cx(n)
x(k ) f ( x(k 1), u(k 1), Wxx, Wxu, Wxb); …(1) y(k) g(x(k)) Cx(k); where, is a nonlinear tansigmoidal function vector characterizing the hidden layer and C is the vector of synaptic weights characterizing the output . The output is tapped at the output node, which itself is the one of the nodes of the hidden layer (Fig. 2). The training algorithm adopted is based on the algorithm developed by Williams and Zipser (Williams, et. al.) and popularly known as RTRL (Real Time Recurrent Learning). -1 z-1
ETotal ( k )
where, η is learning rate of sufficiently small value. The output is updated
Lambda
IFAC ACODS 2016 February 1-5, 2016. NIT Tiruchirappalli, India Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105
1.15
1.1
1.15
1.1
1.05
1.05 1
1 1
0.95
0.9 1000
1500
2000
2500
3000
Engine events
3500
4000
0.95
4500 3.68
0.9 1000
3.7
3.72 1500
3.74 3.76 Engine Events 2000
2500
3.78
3.8
3.82 4
x 10 3000
3500
4000
Engine events
Fig. 2. RNN virtual sensor topology
Fig. 3. Lambda : Sensor and RNN virtual sensor outputs (low and high bandwidth signals)
The weights of a fully connected recurrent network are adjusted in real time using either a supervised continuous or a
In similar manner the RNN virtual sensing was performed for the parameters, Manifold Absolute Pressure (inputs: TPS, RPM, MAF, FPW), Manifold Air Flow (inputs: TPS, RPM, 101
4500
IFAC ACODS 2016 102 Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105 February 1-5, 2016. NIT Tiruchirappalli, India
MAP, FPW), Engine Torque (inputs: TPS, RPM, MAP, FPW, ECT, Ignition angle) and the results are depicted in the figures 4,5. The torque is sensed on the test bed set and not available as a sensed value on-board the vehicle. All values are normalized in the range [-1 1] and displayed. These sensors are tuned independently. It is observed that RNN was good enough to capture these signals as they are limited bandwidth signals. It is obvious that all of these virtual sensors do not work at the same time, meaning only one could perform at a time. It can be seen that RNN with tansigmoial functions for all the nodes does not seem to have the discriminatory feature to capture with equal precision both the low frequency and high frequency components of the lambda response.
x n x
x2 e 2
y
Sensor data RNN Virtual sensor
k k
100
500
900
1300
Engine Events
RNN Virtual sensor Sensor data
i
it
b
The resulting wavenet topology is shown in Figure 6. With the initial network developed as stated above, optimal weights were obtained by minimization of MSE based on cumulative error, ETotal. With the weights available, the initial network is pruned using Akaike information criteria. For the pruned network, optimal values of the dilation and translational parameters, network weights and the bias were obtained by solving the Least Squares problem using Levenberg- Marquadt method.
0
5500
it
i 0 t1
…(8) where, X is the input vector; Dj, Di = Dilation vectors associated with scalon at level j and wavelon at level i respectively; NTj = number of translation vectors associated with scalon at level j; NTi = number of translation vectors associated with wavelon at level i; Tjs = Translation vector associated with scalon at level j, Tit = translation vector associated with wavelon at level i; b = bias; NM is chosen as 1 for the wavenet structure.
1700
1.0
5000
js
k1
Fig. 4. MAP and MAF: Sensor and RNN virtual sensor outputs
4500
j
n
Engine Events
-0.4
js
c x -1
-1
N L 1 NTi
w [D (X T )} w [(D (X T )] j 0 s1
MAF Normalized
MAP Normalized
Sensor data RNN Virtual sensor
… (7)
( x) Wavenet network has a feedforward topology with one hidden layer consisting of scalons and wavelons . The data points were examined at various scales and at each scale the distribution of data was examined to arrive at the initial set of translation vectors. In the present case a scale level NL=5 was found adequate. The resultant equation describing the wavenet is described by the equation N M 1 NT j
1
for n dimensions
x2 2e 2
1
Engine Torque Normalized
2
6000
6500
7000
7500
Engine events
Fig. 5. Torque: Sensor and RNN virtual sensor outputs 2.2 WNN for lambda sensing To overcome the deficiency of RNN, an alternate WNN topology (Kamat et. al., 2006) was proposed for this purpose. This structure is essentially a feed-forward topology, with the addition of time delays to bring in the dynamics into the model. The wavelet transforms (Bakshi et. al., 1993; Miller, 1998) could be adopted as activation functions to impart the ability to capture the local behaviour in addition to the mean (global) behaviour. A family of wavelets is derived from the translations and dilations of a single function x , called mother wavelet and is represented by
x u for s s
1
s, u R 2
Fig. 6. WNN topology (t: translations, d: dilations, w: weights of wavelons, c: feed through weights, b: bias, first node φ in layer is scalon)
…(6) The wavenet (WNN) trained has shown promising results as depicted in figure 7. In the lambda figures the line at magnitude 1.0 denotes stoichiometric combustion for which lambda =1.0, meaning the combustion is closer to stoichiometry.
The wavelet family selected for this exercise is Mexican hat which is a derivative of Gaussian function. The expressions for scalar function and vector mother wavelets are
102
IFAC ACODS 2016 February 1-5, 2016. NIT Tiruchirappalli, India Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105
4. SOFT / VIRTUAL PROCESSING
1.2
W NN Virtual sensor Sensor data
Lambda
1.15
1.1
103
SENSING
IN
MATERIAL
2.1 Soft sensing of quality parameter: Cement blaine
stoich 1.05
The cement manufacturing plant comprises of cement mill at its tail end that grinds cement clinkers (output from cement kiln) into finest powder which is the final product. The fineness of the product is determined by means of a parameter called blaine and is measured on sample basis at regular intervals at laboratory. While the lab measured values are available at intervals such as 1 hour, the control inputs to the cement mill are deployed at every 1 minute, or so. This justifies the need for a soft sensor that could estimate the blaine value at the sample time interval required by the mill control system. In this work the plant considered is rotary ball mill that grinds the clinkers into final product. The inputs to the plant are feed rate, separator fan rpm and power, motor drive power, accumulation of the material in the mill shell which are available at faster rate than the lab measurement. A time-unfolded feed-forward ANN is built using the past data of the mill available from the SCADA that governs the mill operation. All the inputs are posed as individual time-delayed inputs and the network is trained in supervisory mode using lab data as target available at rate Ts. This calls for substantial amount of test data. A generalization of this network is carried out at the rate ts by progressively replacing the inputs in such a manner that the u(k-1) is replaced by u(k) and so forth (Fig 9). This produces output value at the faster sample rate (ts << Ts ). This procedure is repeated until substantial data points are generated by this trained network.
1
0.95 0.96
0.98
1
1.02
1.04
1.06
1.08
1.1 4
Engine events
x 10
Fig. 7. Lambda: Sensor and WNN virtual sensor outputs (High BW) In these implementations the measurement noise is not explicitly addressed. 3. SOFT SENSING IN ELECTRIFIED VEHICLES Hybrid electric and all electric vehicles are going to be the future vehicles for known environmental reasons. These vehicles have Li-ion battery as electrical energy storage element of their powertrain. Essentially, the battery pack is composed of a string of series-parallel arrangement of basic cell in order to achieve net bus voltage and Ah capacity of the pack. The control and management of the battery parameters from its performance, life, diagnostics and safety points of view, the battery management systems have come into existence and are undergoing continuous improvement. Among many other vital parameters the state of charge (SoC) of the battery is very crucial to know about the battery’s status of holding the energy and hence the electric range it can travel safely. In the absence of an on-board sensor for SOC, this property needs to be inferred using various approaches for soft/ virtually sensing it. The input parameters are battery terminal voltage, battery current (both charge and discharge), and temperature at various cells in the pack. n this work the RNN and WNN topology are trained in a supervised mode using the the SoC reading obtained from the existing on-board SoC estimator. The topology is further assisted by making use of SoC reading made available from specially conducted test carried out at decided time intervals which are very few in number. Thus it is a soft sensor rather than a true virtual sensor. The results are shown in figure 8. The RNN shows more noise immunity than WNN.
A fresh ANN is now built with single time dimensioned inputs (Fig 9.c) and the generated target variable values, both of which are now available at faster rate. The target values at the lab measurement instant are the values obtained from the lab which are true values. The input dimensionality is now minimal and the network is re-trained using RNN to address the dynamics. The RTRL algorithm in teacher forcing mode is adopted to accommodate further lab measurement values. The method has element of redundancy and it is observed that, the lesser the interval between lab readings, better is the accuracy of soft sensor. The figure 10 shows results of the blaine soft sensor used as feedback element for the multi variable mill control where the controller is supposed to achieve the blaine within specified limit. . The soft-sensed blaine values under the closed loop control show acceptable quality output.
Existing SoC estimator WNN Virtual sensor
2.2 Virtual sensing of burning zone temperature in cement kiln 0
5000
10000
15000
20000
The ground raw mix ore is pyro-processed in a huge rotary inclined shell cement kiln. This mix is added from the top and a burner fires the pulverized coal powder mixed with fresh air flow to raise the temperature of the order 900 to 1400 deg C for the clinckerization reactions to take place in the zone called burning zone. This temperature is usually sensed by IR camera and faces limitations due to interruption
30000
Sample number
Fig. 8. SoC of Li-ion battery : Estimator, RNN and WNN virtual sensor outputs
103
IFAC ACODS 2016 104 Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105 February 1-5, 2016. NIT Tiruchirappalli, India
5. VIRTUAL SENSING IN CHEMICAL PROCESS (pH CONTROL)
.
. u1(k-2) u1(k-1)
FF ANN @Ts
. u2(k-2) u2(k-1)
u1(k-1) y(k) u1(k)
FF ANN @ Ts
. u2(k-1) u2(k)
(a) FF- ANN in training
A pH processes is highly nonlinear, may exhibit varying characteristics depending on the character of the input and conditions of operation. A typical characteristics may be drastic increase in process gain over a small region around operating point causing sharp transients as the operating point is approached from low gain regions. Another possible manifestation could be variations of steady state profile due to changes in character of input. In the case of pH measurement, there could be loss of sensor signal due to frequent fouling. These issues pose considerable challenge in developing a framework and methodology for such processes. The process model considered here is taken from Henson and Seborg (Henson et. al., 1994). It consists of an acid stream (q1), buffer stream (q2) and base stream (q3) that are mixed in a tank (Fig 12). The system involves 3 state variables including the reaction invariants of the effluent stream and height in the tank. The model also includes valve (unity gain, time constant p ) and transmitter dynamics (unity gain, time constant h) as well as hydraulic relationships for the tank outlet flow. The effluent pH is measured by probe (with time delay Δt) located downstream from the tank.
(b) FF- ANN in mode generalization mode y(k)
u1(k-1)
RNN @ts
u2(k-1)
(c) RNN trained using data generated by (b), ts<< Ts Fig. 9 Slower rate FF-ANN to faster rate RNN
H
Fig. 10 Cement blaine soft sensor output Fig. 12 pH (with buffer) neutralization scheme 1350
1300
In the case of RNN based virtual sensor (Kamat et. al., 2005) the size of training matrix required is high. In case where both buffer and base are perturbed, larger matrix is required. Dead time is accounted by delaying the inputs appropriately. Two lagged inputs at (k-1) and (k-2) are found adequate to capture dynamics for prediction of output at (k). Network is trained for a number of iterations by reducing the learning rate () for successive reductions in NMSE. The figure 13 shows results for randomly perturbed q3 flow whereas a steady state pH-q3 profile is shown in figure 14.
BZT deg C
1250
1200
1150
1100
1050
Sensed temperature W NN virtual sensor
1000
950
0
100
200
300 Sample number
400
500
600
Fig. 11 Cement kiln BZT virtual sensor In case of WNN based virtual sensor the training data and validation data used are the same as that in case of RNN. The wavenet being a static mapping of inputs and outputs, the dynamics is accounted by unfolding the inputs. The size of training data required is about one fifth for case when only the base flow is perturbed. In the case where both buffer and base flows are perturbed, a comparable size of training data is required. The figure 15 shows results for randomly perturbed q3 flow whereas a steady state pH-q3 profile is shown in figure 16.
of the view of the flame due to material turbulence. This led to complement the temperature measurement by a virtual sensor. Some of the crucial inputs used for this are coal feed, kiln rpm and power, material feed and feed material temperature, secondary (hot) air temperature, etc. The results of WNN sensor built using the aforementioned approach are shown in the figure 11.
104
IFAC ACODS 2016 February 1-5, 2016. NIT Tiruchirappalli, India Shivaram Kamat et al. / IFAC-PapersOnLine 49-1 (2016) 100–105
Both the virtual sensors are be able to capture the trend in steady state profile reasonably well, while the ability to capture the dynamics is comparable.
to processes whose characteristics may range from mildly nonlinear to high nonlinearity. The networks also have the inherent flexibility to map operational conditions from near steady to transients of varied spectral character. The design approach of such structures has also been evolved using a careful analysis of system characteristics and operating data. The WNN has greater degrees of freedom in shaping the responses by the virtue of dilation and translation and the choice of the wavelet function. In the case of RNN for the same process, the dimension of input vector is relatively smaller since dynamic features are embedded in the RNN structure itself unlike in the WNN case. The virtual sensors augment the sensing part of a control system of a process and help in reducing the cost and time involved in the precise real life measurement of its crucial parameters.
11
10
pH value
9
8
7
6
5
Sensed pH data RNN virtual pH sensor
4
3 1000
1050
1100
1150
1200
1250
1300
Sample number
ACKNOWLEDGEMENT
Fig. 13 pH: Sensor and RNN virtual sensor outputs
The authors acknowledge the support from the management of TCS limited and participation of the former TCS associates, Jessy Smith, Vivek Diwanji, and Hossein Javaherian of General Motors while carrying out some of the reported work.
14
pH steady state values
12
10
Sensed pH data RNN virtual pH sensor
8
6
105
4
2
REFERENCES
0 4.4
4.6
4.8
5
5.2
5.4
5.6
5.8
base fluid flow perturbation inml/s
B E Miller, “Wavelet Networks for Characterization and Implementation of Environments for Haptic Display.” Masters Thesis, Northwestern University, 1998. Bhavik R. Bakshi, George Stephanopoulos,Wave-Net: a Multiresolution, Hierarchical Neural Network with Localized Learning. AIChE Journal, Jan1993,Vol.39, No.1. M. A. Henson and D. E. Seborg, “Adaptive Nonlinear Control of a pH Neutralization Process”. IEEE Trans. on Control Systems Technology, 2, 169, 1994. Simon Haykin, Neural Networks – A Comprehensive Foundation, (Pearson Education, 1999). Shivaram Kamat, Vivek Diwanji, Jessy George Smith, K. P. Madhavan, ‘’Modeling of pH Process Using Recurrent Neural Network and Wavenet”, CIMSA 2005 – IEEE International Conference on Computational Intelligence for Measurement Systems and Applications Giardini Naxos, Italy, 20-22 July 2005. Shivaram Kamat, V Diwanji, Jessy Smith, Hossein Javaherian, KP Madhavan, “Virtual Sensing of SI Engines Using Recurrent Neural Networks”, SAE 200601-1348. Shivaram S. Kamat, Hossein Javaherian, Vivek V. Diwanji, Jessy G. Smith, and K. P. Madhavan, “Virtual Air-Fuel Ratio Sensors for Engine Control and Diagnostics”, American Control Conference 2006 Williams, R. J, D. Zipser, A learning Algorithm for Continually Running Fully R ecurrent Neural Networks, Neural Computation, 1, 270-280. D. Prokhov, Virtual Sensors and Their Automotive Applications, Intelligent Sensors, Sensor Networks and Information Processing Conference,2005, 411-416
Fig. 14 pH steady state values comparison 10 9.5 9
pH value
8.5 8 7.5 7 6.5
Sensed pH data WNN virtual pHsensor
6 5.5
1000
1200
1400
1600
1800
2000
Sample number
Fig. 15 pH : Sensor and WNN virtual sensor outputs
14
pH steady state values
12
10
Sensed pH data W NN virtual pH sensor
8
6
4
2
0 4.4
4.6
4.8
5
5.2
5.4
5.6
5.8
base fluid flow perturbation in ml/s
Fig. 16 pH steady state values comparison 6. CONCLUSIONS The paper has reported the feasibility of developing soft/virtual sensors using RNN and WNN. The two structures were shown to complement each other and could be applied
105