Journal of Materials Processing Technology 113 (2001) 322±330
Selected examples of cutting process monitoring and diagnostics Andrzej Sokoøowski*, Jan Kosmol Department of Machine Technology, Silesian University of Technology, Gliwice, Poland
Abstract This paper deals with one of the signi®cant factors in cutting related areas, which can be improved to give better overall machining performance. The authors propose how to shift the human machine tool operator position to a higher level of supervision by using an intelligent system to perform the cutting process and machine tool monitoring and diagnostics. The two speci®c approaches based on a socalled Intelligent Monitoring System (IMS) are considered. The IMS is implemented as an Intelligent Monitoring System Designer (IMSD) and it is applied as a so-called Intelligent Tool. In order to con®rm the applicability of the above-mentioned approaches, several examples are discussed in this paper. # 2001 Published by Elsevier Science B.V. Keywords: Cutting process; Machine tools; Diagnostics and monitoring; Arti®cial intelligence
1. Introduction This paper deals with one of the signi®cant factors in cutting related areas, which can be improved to give better overall machining performance. The authors propose how to shift the human machine tool operator position to a higher level of supervision by using an intelligent system to perform the cutting process and machine tool monitoring and diagnostics. From a general point of view, the research presented aimed at productivity and quality improvement in machining that seems to be obvious and does not require any justi®cation [1]. In this paper, the authors deal with one of the possible ways of such improvement, i.e. some aspects of modern monitoring systems applied for speci®c problems related to metal cutting are discussed. In order to obtain a machined part of an assumed quality, several points must be taken into consideration. In many cases, these points are opposite to the productivity requirements that obviously should be considered also. The machined part quality depends on the following cutting parameters: the workpiece, the applied cutting tool, the realisation of the cutting process and the machine tool condition. From the diagnostics point of view, this means that the workpiece, the cutting tool, the machine tool and the cutting process should be monitored in order to avoid any disturbances that can affect the quality of the machined part.
* Corresponding author. E-mail address:
[email protected] (A. Sokoøowski).
0924-0136/01/$ ± see front matter # 2001 Published by Elsevier Science B.V. PII: S 0 9 2 4 - 0 1 3 6 ( 0 1 ) 0 0 6 2 3 - 9
Two speci®c approaches to monitoring and diagnostics are considered in this paper. Both of them are based on the concept of an Intelligent Monitoring System (IMS) [1±3]. The ®rst approach considers practical implementation of IMS in the form of an Intelligent Monitoring System Designer (IMSD). In the second case, the so-called Intelligent Tool is analysed. First, both of the above-mentioned approaches are discussed in the next part of this paper. Then, the IMSD is characterised in order to describe the main algorithms and procedures tested in this research. Finally, the selected examples of the conducted tests are shown. In this case, the focus is on measured signal feature selection and integration, mainly. It should be underlined that the detailed description of the used data can be found in [1±5]. Therefore, in this paper the focus is on recently obtained results with an attempt being made to generalise the conclusions summarising research. 2. Considered approaches As mentioned above, the ®rst approach is based on the socalled IMS (Fig. 1). In order to introduce this approach, the functions performed by a skilled machine tool operator are analysed [2]. During machining, the operator observes the cutting process, analyses the machining sound and vibration or qualitatively assesses the manufactured products. Based on experience, he/she tries to estimate whether the process is ``normal'' and the desired quality of the products can be expected. When a ``not-normal'' situation occurs, he/she can react and try to avoid any disturbances. Trying to build an
A. Sokoøowski, J. Kosmol / Journal of Materials Processing Technology 113 (2001) 322±330
Fig. 1. Block diagram of an IMS.
algorithm of the human operator behaviour, it is assumed that the observation of the process is considered as a measurement of certain signals. Following this assumption, one can distinguish three main steps. In the ®rst step, the socalled feature extraction step, several features of the measured signals are estimated. In the next step, the so-called feature-selection step, the operator selects the most important features based on personal experience. Then, in the third step, the so-called feature integration step, the selected features are integrated within a decision-making procedure, i.e. the operator assesses the process as ``normal'' or ``notnormal''. Similar steps can be distinguished while analysing the functions performed by a conventional monitoring system. Despite these similarities, however, there are strong differences also. In the case of the conventional system, the set of measured signal features is strictly constrained to that assumed by the system designer. Also, such a system cannot update its strategy by analysing new incoming information, i.e. in contrast with the human operator, the system cannot learn. The differences stated above have been taken into consideration while de®ning a concept of IMS [2,4]. The de®nition considered in this paper is expressed in the form of recommendations and requirements (Fig. 1). First of all, the intelligent system should be able to automatically test and select the best con®guration of sensors and signal processing methods. This means that the system should automatically select measured signal features, which show the best correlation to the observed phenomenon. Moreover, the IMS should possess self-learning abilities in order to automatically model a relationship between the selected features and the observed phenomenon, i.e. the system should learn from provided data without human interaction.
323
Finally, the IMS should be capable of representing the acquired knowledge in a human comprehensive form. Practical implementation of the above de®ned concept can be, however, dif®cult or expensive since very sophisticated signal processing methods would have to be applied. Therefore, a different approach is proposed. Instead of building the IMS, it is proposed to apply a concept of such a system while designing conventional monitoring systems. On the level of designing, there is a possibility of interacting and assessing different solutions. Thus, if some of the signal processing methods do not ful®l the requirements, the designer can support the search for the best solution. The general objective of such an approach is to minimise the time (cost) that a designer would have to spend on analysing, testing and selecting the best con®guration of sensors and signal processing methods. In other words, the present research aimed at developing a tool Ð the IMSD Ð that would provide facilities, i.e. methods and algorithms, to automatically design monitoring systems (Fig. 2). The IMSD is not, however, the only way of practical implementation of the concept of an IMS. Currently, extensive research is being conducted on adaptation of this concept to the requirements of the Intelligent Tool (Fig. 3). The Intelligent Tool is a software±hardware system acting in a certain environment (machine tool, workpiece), which can automatically assess its current condition, predict
Fig. 2. Two possible ways of designing a monitoring system.
324
A. Sokoøowski, J. Kosmol / Journal of Materials Processing Technology 113 (2001) 322±330
Fig. 3. Block diagram of the Intelligent Tool.
its condition in the near future and assess the state of other phenomena via combination of the information on its condition with the other information available in the considered environment. Such a de®nition assumes the importance of the cutting tool as an element uniquely pointing at the type of machining and, thus, type of monitoring system to be applied. The implementation of the Intelligent Tool starts with a selection of a certain tool. This causes the sending of appropriate information from the Data Base to the Information Bank and the Execution Block. This means that, ®rst of all, speci®cally for this tool, sensory data and cutting parameters are available in the Information Bank. Also, the intelligent procedures dedicated to this tool are activated in the Execution Block and the procedures are equipped with adequate parameters. For example, if a drill is to perform ®nishing machining, the arti®cial neural network
for estimating drill wear and a neuro-fuzzy logic system for drilled hole surface roughness assessment are activated. 3. IMSD characteristics The characteristics of the IMSD [2] can be presented in two steps that re¯ect the two main parts of this system, i.e. hardware and software (Fig. 4). It is convenient to recall that the characteristics presented here aim at a description of the methods and algorithms applied in this research. The hardware part is basically designated to acquire the data. Therefore, one can distinguish in this part a set of sensors and equipment for measured signal ampli®cation, conditioning and data storing. The software part consists of several modules that are described in the next sections of this paper.
Fig. 4. Block diagram of the IMSD software.
A. Sokoøowski, J. Kosmol / Journal of Materials Processing Technology 113 (2001) 322±330
3.1. Feature extraction module During the feature extraction phase, a user determines a set of measured signal features that potentially show the highest correlation to the states of the observed phenomenon (Fig. 4). The feature extraction is performed in the two steps. In the ®rst step, the data is preliminary processed. The signal processing can be done in the time or the frequency domain. In the ®rst case, a user can apply smoothing with the moving average method or the true-RMS method. Also, a user can calculate the signal envelope, remove the signal mean value, calculate the signal trend and remove this trend from the signal. The preliminary processing in the frequency domain relies on the application of Fast Fourier Transformation, i.e. a user can compute the frequency spectrum, the power spectrum, the autocorrelation function and the cepstrum function. After the pre-processing, the primary functions of the feature extraction module are realised. In the case of the time domain, it is possible to calculate the mean value, the variation, the standard deviation, the RMS value, skew and kurtosis, the mean-square value and the zero crossing rate. The feature extraction in the frequency domain is based on the previously calculated spectral characteristics. First, a user speci®es the frequency bands in which the feature extraction is performed. Then, the program can calculate several values in each band, e.g. the maximum of the magnitudes or the RMS value. Also, frequencies corresponding to the maximal magnitudes can be evaluated. 3.2. Feature-selection module Feature selection is considered as one of the most important phases of data processing. In this phase, the number of sensor features (number of sensors) for proper modelling of the observed phenomenon is decided. The two groups of methods for feature selection are implemented in IMSD. The ®rst group is based on the Scatter Matrix (SM) method [4,6]. In order to provide the SM method with different feature combinations, the Sequential Forward Search (SFS) algorithm is commonly used. In the present research, the authors tried to overcome some drawbacks of the SFS algorithm by replacing it with the Genetic Algorithm (GA) [2]. The combination of the SM method and GA is considered here as a second method for feature selection. The two methods described above can be used to perform the feature selection with no assumption as to further feature integration techniques. In contrast, the methods discussed next in this section are based on a feed forward back propagation (FFBP) neural network that serves as a feature selection and feature integration method at the same time. Four procedures are implemented in IMSD. The ®rst one is referred to as the weight pruning method [4,7]. The weight pruning examines each weight of an already trained network and tries to eliminate some of them. Eventually, each input to the network is described with the number of weights that
325
did not ``survive'' the process of elimination. It is assumed that the higher is this number, the less important is the respective input. The second method can be called the weight sum method [4]. The method is also applied to the already trained network. The importance of each input is estimated based on the sum of the absolute values of the weights outgoing from this input. The third method is based on the sensitivity analysis that determines an in¯uence of each input on the neural net outputs. The last approach is based on a modi®ed form Karin's weight pruning procedure [2]. In this procedure, each weight is ranked based on the sum of its changes during the training. A small sum of changes mean that the weight does not contribute to the modelling of a phenomenon and, therefore, can be neglected, as is done in the weight pruning method. 3.3. Feature integration module At the last step of designing of a monitoring system, the selected features are integrated within a model of the relationship between the states of the observed phenomenon and the information provided by the sensors. In order to perform the feature integration, a user can apply an FFBP neural network or a neuro-fuzzy logic system. Since the FFBP neural network is relatively well known, the description of this network is constrained to a list of the main techniques supporting its application. In order to train the network, a user can select one of three different training procedures, such as the incremental update procedure, the cumulative update procedure and the incremental learning procedure with pattern extraction. In order to determine the number of nodes in a hidden layer, two options are available. The ®rst one allows a user to apply the node pruning technique [7]. The second method for deciding the network structure is called here an automatic algorithm [2,7]. This algorithm is based on the concept of a dynamically changing network structure during training. A fuzzy logic algorithm with learning abilities is implemented in IMSD as a second method for feature integration [4,7]. The application of training algorithms to fuzzy logic gives it the capability to accurately tune its parameters and improve its performance. Also, it is possible to easily extract knowledge from the trained system and improve the understanding of relationships between the sensory information and the observed phenomenon. 4. Selected examples 4.1. The academic task The ®rst test presented in this paper is based on a so-called academic task. In such a task the solutions are already known and, therefore, it is possible to properly assess the methods implemented in IMSD. It should be emphasised that the test is focused on the feature-selection methods that
326
A. Sokoøowski, J. Kosmol / Journal of Materials Processing Technology 113 (2001) 322±330
Fig. 6. Example of the estimated input importance.
Fig. 5. Graphical representation of the academic task.
seem to be the most crucial in the whole process of monitoring the system design. The task analysed is depicted graphically in Fig. 5. The FFBP neural network is applied to model a certain surface based on the two-dimensional input vector (inputs WE-1 and WE-2). Since interest is in the classi®cation task, the simulation performed with the neural network was constrained to a binary case in which the net was supposed to detect whether a point belongs (or not) to the two areas shown in Fig. 5. The test of feature-selection methods rely on augmenting the neural network input vectors (containing the basic inputs WE-1 and WE-2) with additional inputs representing redundant information. Obviously, the methods should reveal such information and point at respective inputs as candidates to be dropped. Four different redundant inputs were used, i.e. input representing constant value with noise (CONST), input representing corrupted with noise input WE-1 (REP), input representing computer generated noise (NOISE) and input being the sum of inputs WE-1 and WE-2 (SUM). An example of estimated importance is shown in Fig. 6. Generally, the results can be considered satisfying for a potential user. In most cases the importance of the redundant inputs was much lower in comparison to the WE-1 and WE2 inputs. In such cases, the redundant inputs are to be dropped. However, some undesirable differences were revealed, as well. The tested methods assessed the redundant inputs with various importances. This is to say, in some cases the analysed importance was on the level of that associated with the basic inputs WE-1 or WE-2. This can cause dif®culties in making a decision as to the candidate inputs to be neglected. Also, one can point at the simulation in which the REP input was introduced to the network. In this case, Karin's method did not assess properly the input, which suggests a need for careful review of this method.
Similar results were obtained while feeding the FFBP network with all redundant inputs at the same time (Fig. 7). Analysing these results, the high applicability of three feature-selection methods, i.e. the weight pruning method, the weight sum method and sensitivity analysis, can be underlined. These methods can properly detect the redundant information not only for a single input case but in the case in which several redundant inputs are fed to the network also. From another point of view, Karin's method failed in proper importance estimation, again. This last conclusion, however, should be considered carefully since some of implemented methods can be data-dependent. This means that it is not necessary to neglect them at all but apply them to a restricted type of tasks only. 4.2. Drill wear monitoring during multi-spindle machining The goal of the research presented in this section was to develop a strategy for monitoring tool wear while drilling with a multi-spindle drilling machine [6,7]. Drill wear-monitoring systems for such machining have to ful®l additional requirements compared to single-drill machining. Several drills machining at one time represent different, independent sources of signals, hence the signals interfere
Fig. 7. Input importance estimated while feeding the FFBP network with all redundant inputs.
A. Sokoøowski, J. Kosmol / Journal of Materials Processing Technology 113 (2001) 322±330
327
Fig. 8. (a) The schematics of the multi-spindle measuring set-up; (b) the importance estimated with the weight pruning method; (c) and (d) the performance of the FFBP fed with selected features.
and overlap with each other and create a cumulative signal. The schematics of the applied measuring set-up are shown in Fig. 8(a). For the experiments, 10 high-speed steel twist drills of diameter 10.5 mm were used. Two electromagnetic current transducers were applied to the power lines of the main spindle motor (sensor A) and the feed motor (sensor B). A pair of one vibration sensor (C) and one AE sensor (E) was installed in the back of the workpiece mounting plate and another pair of AE (F) and vibration sensor (D) were installed on the bushing mask of the machine. Just after acquiring data, four statistical features were extracted for each measured sensor signal: (1) mean value, (2) standard deviation, (3) slope of the linear trend of the measured signal, (4) standard deviation of the signal with subtracted linear trend. The four selected features multiplied by the number of sensors led to a matrix of 24 sources of information. This matrix augmented with feed (p) and workpiece material (M) was a base for further signal processing. In the next step, feature selection and integration was performed. It is convenient to recall that the goal of the
analysis presented here is to select the most important features and then integrate them within a model of the relationship between these features and the drill-wear levels. The main tests consisted in training the FFBP nets with subsets of input features from which the ``worst'' inputs were gradually removed. Obviously, the extraction of inputs (input features) was done following the importance estimated with the feature-selection methods. An example of estimated input importance is shown in Fig. 8(b). The process started with removing already 19 inputs and continuing to neglect features until only four features were left. In Fig. 8(c) and (d), each simulation is characterised with a graphical representation of input vector and respective neural net performance. Analysing the results obtained, it can be seen that removing 19 features did not cause any deterioration in the network performance (the initial performance for 26 features was at the level of 95%). Especially, very high performance was retained while removing inputs following the importance estimated with the weight pruning method. It must be stated clearly that ®nal subset of four features allows the user to substantially reduce the
328
A. Sokoøowski, J. Kosmol / Journal of Materials Processing Technology 113 (2001) 322±330
number of sensors to be applied in an assumed monitoring system. Following the discussed results, a user could apply only two sensors and monitor the drill wear with a performance of 94%. This regular trend has not been observed while training an FFBP net with combinations selected with the weight sum method. Here, the performance started decreasing and reached the level of only 80% (Fig. 8(d)). Such a case points at the differences between the input selection methods, as was concluded in the previous section. 4.3. Burr formation modelling The research presented in this part of this paper deals with burr formation modelling. The main goal of the research was to build a model of the relationship between the burr height created during the drilling operation and the signal representing the axial drilling force [5]. The measurements were conducted while drilling austenitic steel (DIN) X5CrNi1810 (C: 0.028%, Mn: 1.26%, Si: 0.29%, P: 0.039%, S: 0.025%, Cr: 17.75%, Ni: 9.10%, Cu: 0.30%) with twist drills of diameter 5 mm. The analysis of the drilling force focused on two main tasks. In the ®rst case, the drill path length ``dpl'' was to be decided. This length describes the position of the drill tip from which the cutting force potentially contains important information on burr height. The second task was related to the applied measured-signal-processing methods, i.e. the signal was represented with its derivative and smoothed with the moving average method. Before applying the feature-selection algorithms, a conventional analysis has been carried out. In this analysis, a step-by-step procedure based on statistics was applied to build several burr formation models. For example, in order to determine the optimal drill path length the statistical models were compared for correlation factors R and the sum of residual values SRVs, as shown in Fig. 9(a). Such a way, however, cannot be considered satisfactory. As can be noted, one would have to spend a lot of time in order to test a high number of models to point at the best parameters of the signal processing methods and the optimal tool path length. In contrast, the application of the FFBP network involved feeding the 9-3-1 net with all values representing different drill paths lengths, at the same time. The results obtained (Fig. 9(b)) fully corresponded to those shown in Fig. 9(a). During the conventional analysis, the optimal drill path length was selected ®rst and then different signal features were tested. Ideally, such an analysis should be performed in parallel so that the selection of the optimal drill path length would not affect the selection of the signal feature. From the neural network point of view, this means that the FFBP network should be trained with input vectors containing values representing different drill path lengths and the different analysed signal features. The results obtained in this case are shown in Fig. 10. Analysing the results shown in Fig. 10, the relative importance of the computed signal features should be taken
Fig. 9. Results of: (a) the conventional approach; (b) the FFBP neural network application (SRVs Ð sum of residual values; R Ð correlation factor).
into consideration, ®rst of all. In this case, it is possible to conclude that the feature-selection methods uniquely ranked the mean and RMS values with higher importance than the standard deviation. Then, all methods estimated similar drill path lengths (MN-5, MN-6, SD-6 and RMS-6) with high importance. This result again corresponds entirely to the result of the conventional analysis. 4.4. Tool wear monitoring in turning The main goal of the research presented in this section was to work out a cutting tool wear-monitoring strategy which makes possible far reaching independence of the wear symptoms from the cutting conditions [2,4]. It should be mentioned that the example analysed here serves as the basis for a global comparison of the input selection methods applied in the present research. Therefore, the data used is characterised and then the performances of the FFBP network applied to assess various subsets of the selected inputs are discussed. The data have been taken during turning operations. The measurements consisted in measuring the vibration velocities in the thrust direction and the respective cutting tool ¯ank wear VB. The feature extraction was performed based on the spectral characteristics of the measured signal. First, the spectral characteristics were split into eight equal bands and the RMS values of magnitudes within each band were computed. Next, each set of the RMS values was augmented
A. Sokoøowski, J. Kosmol / Journal of Materials Processing Technology 113 (2001) 322±330
329
Fig. 10. Input importance estimated with the sensitivity method (MN Ð mean value; S.D. Ð standard deviation; RMS Ð root-mean-square value; 1±9 drill path lengths).
Fig. 11. FFBP neural network performances achieved for different number of input features.
with information on cutting parameters, insert material and tool overhang. In this way the input vectors containing 13 inputs were obtained. The input vectors were then processed with the input selection methods. In order to compare the input selection methods, the FFBP performances obtained for a gradually decreasing number of inputs are depicted in Fig. 11. Analysing the results obtained it can be noted that the best results were achieved with the input subsets selected with the weight pruning method. Also, the selection performed with the SM-based procedures must be considered satisfying. The other methods, however, are dif®cult to assess in a unique way. Karin's method allowed the selection of inputs for a restricted number of inputs, i.e. the high performance was retained for nine inputs, only. This relatively high number of inputs cannot ful®l the expectations of a potential user. Next, the weight sum method and the method based on sensitivity analysis gave relatively high similar performances. However, the performances varied, reaching the level of 91% for seven input features and only 88% for nine input features. Generally, this means that the selected feature combinations vary in terms of the performances obtained while decreasing the number of inputs. At this point, it is convenient to introduce another criterion that may be taken into consideration. A feature-selection method
to put the selected input subsets in a speci®c order may be expected, i.e. the performance of the feature integration method should deteriorate with decreasing number of inputs. In this case, a user could stop reviewing the combinations at the moment when the performance drops below an assumed value. In the light of this criterion, the weight sum method and sensitivity analysis seem again to be less ef®cient compared to the weight pruning or SM-based methods. 5. Summary In the form of a summary of this paper, it is convenient to recall the main bene®ts related to the proposed approaches. First of all, a potential user can be provided with the most important information on the generally considered correlation between measured signal features and the states of the observed phenomenon. This can be done through the application of a certain set of procedures and algorithms that follow the de®nition of the Intelligent System. In each example presented, the procedures selected information allowed substantially reducing the number of features suf®cient to model the phenomenon under consideration. Obviously, the number of sensors for an assumed diagnostic
330
A. Sokoøowski, J. Kosmol / Journal of Materials Processing Technology 113 (2001) 322±330
system was reduced, also. Such a case is of interest here since it allows decreasing in the general cost of machine diagnostics. However, the application of the feature-selection methods revealed some differences in their performance. Such a case suggests the continuation of careful tests. The tests would potentially admit the existence of relationships between the considered feature-selection method, the analysed data and the feature integration method. The above-stated conclusion re¯ects the directions of the authors current and assumed future work. References [1] A. Sokoøowski, J. Kosmol, On some aspects of productivity and quality improvement in machining, in: Proceedings of the Eighth International Conference on Productivity and Quality Research, Vaasa, Finland, 1999.
[2] A. Sokoøowski, A. Kolka, J. Kosmol, A new approach to aid designing of monitoring systems, in: Proceedings of the Matador Conference, Manchester, 1997. [3] A. Sokoøowski, J. Kosmol, Arti®cial intelligence in cutting process monitoring, Advances in Technology of the Machines and Equipment, Vol. 23, No. 3, Poland, 1999. [4] A. Sokoøowski, J. Kosmol, Intelligent Monitoring System Designer, in: Proceedings of the Japan±USA Symposium on Flexible Automation, Boston, USA, 1996. [5] A. Sokoøowski, J. Kosmol, Feature selection for burr height estimation, in: Proceedings of the Fifth International Conference on Monitoring and Automatic Supervision in Manufacturing, Warszawa, 1998. [6] W. KoÈnig, D.A. Dornfeld, M. Rehse, A. Sokoøowski, On designing of a tool wear monitoring system for multispindle drilling machine, in: Proceedings of the Fourth International Conference on Monitoring and Automatic Supervision in Manufacturing, Warszawa, 1995. [7] A. Sokoøowski, D.A. Dornfeld, Intelligent system for cutting parameter optimization and design of cutting process monitoring systems, in: Proceedings of the First S.M. Wu Symposium on Manufacturing Science, SME, Chicago, USA, 1994.