570

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001

A Neuro-Fuzzy Supervisory Control System for Industrial Batch Processes

Christian W. Frey and Helge-Bjrn Kuntze
AbstractThe automation of complex industrial batch processes is a difficult problem due to the extremely nonlinear and variable system behavior or the conflicting goals within the different process phases. The introduction of a single multiple-input multiple-output controller is not useful because of the rather high design effort and the low transparency of its complex structure. A more suitable hierarchical fuzzy-logic (FL) based supervisory control concept is proposed in this paper. It permits the decomposition of the complex control problem into a series of smaller and simpler ones. In the upper level of the hierarchy the FL-based supervisory controller classifies the actual process phase in terms of the available process sensor signals and activates dynamically the appropriate situation specific low-level controllers. This paper presents the generic concept of the FL supervisory controller that comprises both a FL process diagnosis and a control mode selection as well as experiences with the industrial application.
Fig. 1. Scheme of the proposed FL-based supervisory control.

Index TermsMode-selector, neural network, neuro-fuzzy, process diagnosis, supervisory control, wavelet.

I. INTRODUCTION HE number of successful industrial fuzzy logic (FL) control applications has tremendously increased over the last decades. Theoretically founded by Zadeh [1] and explored by Mamdani [2] in the early 1970s the wide industrial application of FL control began almost a decade later in Japan [3]. With the availability of commercial FL control development tools and the better understanding of synthesis and analysis methods, more and more industrial companies in USA, Japan, and Europe have successfully introduced this new technology for their specific applications [4]. The vast majority of FL control concepts proposed in the literature prefer a monolithic multiple-inputmultiple-output (MIMO) control structure. For low order applications this is the most appropriate architecture, however, for heavily disturbed high order nonlinear industrial plants more flexible and transparent hierarchical control schemes are required. There are many complex industrial processes that are difficult to control because their dynamic and static system behavior depends strongly on the process phase. Moreover, the performed requirements may change in the different process phases or situations, respectively. In order to cope with these difficulties a two-stage FL-based supervisory control scheme has been developed and realized, which permits the reliable identification of the actual process phase and the decomposition
Manuscript received November 17 2000; revised May 15, 2001. The authors are with the Fraunhofer Institute for Information and Data Processing, Research Group Industrial Sensor and Control Systems, 76131 Karlsruhe, Germany (email: fry@iitb.fhg.de). Publisher Item Identifier S 1063-6706(01)06534-1.

of the complex control problem into a series of smaller and more transparent ones. The proposed concept is characterized by two hierarchical levels. As illustrated in Fig. 1 the lower stage of the control structure consists of different low level controllers that are optimized with respect to specific process phases. The various low level controllers, which may have any structure (FL control, model-based or hybrid) are activated by a supervisory controller in the upper stage of the hierarchy. Depending on the identified process phase a FL mode selector activates the most appropriate control mode, which can comprise both 1) the adaptation of control parameter sets without structural changes, or 2) the switching of different situation-specific low level controller structures. In cases of slowly changing operating set points the FL-based adaptation of control parameters is the most suitable measure provided no stability problems occur. Switching between the low-level controllers is more useful if the performance requirements have to be changed. For a fast and precise change between very different setpoints, e.g., it may be useful to apply a time-optimal bangbang control strategy for large state differences and a well damped PID controller in the vicinity of the target state. Of course, structural switching requires necessarily the compatibility of the alternative controller structures. Moreover, shock disturbances have to be avoided by introducing a FL-based soft switching strategy as explained in [5], [6]. Focal point of this paper is the functionality of the higher level FLsupervisory controller, particularly with regard to the underlying process diagnosis scheme. The concept and performance of the lower level FL control has been extensively discussed elsewhere [11], [21], [22].

10636706/01$10.00 2001 IEEE

FREY AND KUNTZE: A NEURO-FUZZY SUPERVISORY CONTROL SYSTEM

571

Fig. 2. Fault and state diagnosis scheme.

Fig. 3. Different categories of measured signals.

II. PROCESS DIAGNOSIS In order to diagnose automatically malfunctions and special phases or states of a technical process, the successive steps of feature extraction and evaluation have to be performed [7]. This general abstract model, shown in Fig. 2, allows a broad variety of different realizations and implementations. Within the feature extraction step the available sensor signals of the process are preprocessed for the purpose of extracting relevant signal characteristics. The generated features can be regarded as a condensed signal representation, ideally containing all important signal information. Basically, feature extraction in technical processes is performed by following methods [7]: 1) signal based methods, e.g., threshold comparison, frequency analysis, pattern recognition, or 2) analytical process-model based methods, e.g., parameter estimation techniques. In the next stage, the feature evaluation, a small number of meaningful features that optimally represent the given process phase without redundancy are identified. Finally, the classification of the actual situation is carried out, i.e., the selected features for a given situation are assigned to a specific class by suitable methods like, e.g., [8] 1) statistical evaluation, e.g., Bayes linear classifier, -nearest neighbor or polynomial classifier; 2) artificial intelligence methods, e.g., neural networks, fuzzy logic, and neuro-fuzzy classification. A. Feature Extraction Due to the complexity of industrial processes in the majority of cases, neither an analytical process model nor a sufficient knowledge base for describing the process behavior is available. Thus, feature generation in the application considered in Chapter 3 is focused on signal based methods. Choosing adequate methods for signal-based feature extraction strongly depends on the type of signals to be processed [9]. Generally, it is possible to break down the measured sensor signals into two broad categories. The first one is concerned with low bandwidth measurements such as trends, e.g., in temperature signals; while the second one exemplifies high-bandwidth measurements like, e.g., process values containing spikes and oscillations (Fig. 3).
Fig. 4. Conventional and FL based threshold evaluation.

The possible features of signals associated with the first category may develop slowly over time and can be extracted with standard methods like, e.g., threshold evaluation or trend detection. High-frequency phenomena usually require signal processing methods evaluating the signal over an adequate time span, e.g., the windowed Fourier transform or particular pattern classification techniques [10]. In the proposed diagnosis scheme, advanced signal-based feature extraction techniques for both types of signals have been employed. They will be outlined in the following sections. B. FL Feature Extraction The reliable identification of, e.g., faults and process phases in a technical process depends strongly on the extracted signal features and their entropy. To point out this relation, threshold evaluation as a common technique in feature extraction for low bandwidth signals [9] is considered as an example. In order to check whether a signal is below or above, given thresholds and , the following equation has to be employed: (1) Evaluating this relationship by standard logic gives three possible signal featuresthe signal can be between (B), smaller and . Fig. 4 (S) or larger (L) the given thresholds shows the result of evaluating the inequality with standard logic graphically. Obviously, standard logic provides a poor description of the transition between the possible signal featuresthe information about how close the signal is to the thresholds is lost. In particular, for diagnosis of processes (e.g., chemical plants), where the different process phases cannot be separated exactly, the evaluation of these features would not give a satisfying classification result. To overcome this restriction, the standard threshold evaluation can be enhanced by FL. By introducing membership functions and rules, FL gives the possibility to describe the transition between the signal features in a more detailed and continuous way. The following set of fuzzy rules

572

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001

has to be employed to performthreshold evaluation (TH) with FL: if if if is small, then is between, then is larger, then is is is (2)

As can be seen in Fig. 4, the FL evaluation of (1) leads to a continuous transition between the possible signal characteristics. In practical applications the combination of FL with standard signal processing methods gives a powerful tool for feature extraction. In this context, several other methods have been developed, e.g., adaptive FL threshold evaluation or FL based trend detection; a detailed illustration can be found in [11]. C. Wavelet Neural Networks for Feature Extraction Measured sensor signals of dynamical processes are frequently characterized by a nonstationary and dynamic time behavior. Particularly faults or special process phases can often be determined by typical patterns in the measured process signals. For such patterns timefrequency representations are highly desirable, with a view to deriving meaningful features. Generally the frequency characteristics in conjunction with temporal behavior can be described with respect to the Heisenbergs uncertainty principle [10]; from the variety of different approaches in the developed diagnosis scheme the linear one-dimensional (1-D) wavelet transform (WT) is preferred. In mathematical terms, the WT is expressed as an inner product with a wavelet analyzing function of a signal [12][14] (3) (4) The variable represents scale parameter; shift parameter of the wavelet function; denotes complex conjugate of . normalizes the wavelet for all The weighting factor values of the scale parameter to a constant energy. In contrast to the Fourier transform, which breaks up the signal into sine and cosine functions of different frequencies, the WT uses time limited analyzing functions localized in time and frequency by the shift and scale parameters, respectively. Wavelet analyzing functions have to satisfy several constraints, which are explained in detail in the literature [13], [14]. Frequently used wavelet functions are modulated windows, i.e., they are composed as a product of two functions, e.g., an oscillating term that is modulated by a Gaussian function. A typical example of a wavelet function is the real Morlet wavelet (5) Depending on the scale parameter , the wavelet function is a dilated low-frequency function, whereas for small values of , the wavelet is constricted, which yields to a high-frequency The output of the wavelet node can be interpreted as the correlation between the modified wavelet and the input signal . The MLP in the upper part of the WNN represents the classification part, which bases the classification decision on the wavelet

Fig. 5. Frequency domain analysis of a nonstationary using WT.

function. Due to the time limited analyzing function, the WT is capable of simultaneously providing time and frequency information of a given signal. As an example, Fig. 5 points out this advantage of the WT in conjunction with analyzing a nonstationary signal. The given signal is characterized by two different frequencies at different time intervals. With the FT, both frequencies can be determined but the information at that time instant, the different frequencies occur is lost. On the other hand the WT provides timefrequency map of the analyzed signal, enabling the extraction of frequency features, which vary in time. This attribute makes the WT an ideal tool for feature extraction of nonstationary signals. The concept of so-called wavelet neural networks (WNN) for classification tries to combine aspects of the WT for the purpose of efficient frequency extraction with the decision capabilities of neural-network approaches [16], [17]. WNN can be described as an expanded perceptron with so-called wavelet nodes as a frequency domain feature extraction preprocessing unit. The basic structure of a WNN is shown in Fig. 6. The input layer, composed of wavelet nodes, is responsible for preprocessing the input signal and extracting signal characteristics that are passed to a multi-layer perceptron (MLP) for class assignment. The wavelet nodes are shifted and modulated versions of the mother wavelet. The -th wavelet node is described by the shift parameter and the scale parameter , corresponding to the parameters of the WT. Formally the wavelet nodes output is the inner product of the wavelet and the input signal . Assuming only real valued wavelets in the wavelet layer the output of the -th wavelet node is calculated by following equation: (6)

FREY AND KUNTZE: A NEURO-FUZZY SUPERVISORY CONTROL SYSTEM

573

Fig. 6.

Wavelet neural network for pattern recognition. Fig. 7. NEFCLASS neuro-fuzzy system for feature evaluation.

nodes output. The neurons in the perceptron layers are interconnected by directed and weighted links. A neurons output is processed to all neurons in the succeeding layer, its inputs are the weighted outputs of the neurons in the preceding layer. A is defined by the weighted sum of the outneurons output of the neurons in the previous layer and its activation puts function. In the proposed diagnosis scheme sigmoid activation functions are used. The output of a sigmoid neuron in the layer is calculated as follows: (7) During the training phase, the training patterns are repeatedly propagated through the network, whereas the parameters are adjusted to minimize the least-square error of the distance between the desired output and the networks output (8) The minimization problem is solved by an iterative gradient technique; the so-called generalized delta-rule [15]. The parameters of the neural network are adjusted by calculating the partial derivatives of the error with respect to the parameters of the net, the shifts and the scales A work, i.e., the weights complete derivation and several approaches of the learning algorithm are presented in the literature [15]. D. Neuro-Fuzzy Feature Evaluation The purpose of the feature evaluation step is reliable assignment of the available signal features to the possible diagnosis statements (Fig. 7). With respect to the problem considered in Chapter III the process of feature evaluation can be understood as a classification problem, i.e., the generated features within a typical process phase are classified and assigned to the corresponding process phase. Several methods for feature evaluation have been proposed in the past [7], [18], which can roughly be divided into statistical and artificial intelligence based methods. All methods differ considerably in with respect to their practical realizations

and the developer interaction in the design phase. For instance, developing an evaluation system based on the Bayes linear classifier, as one of the most commonly used statistical method [7], the user has to select the relevant features manually and adjust the parameters of the probability distributions carefully by hand to obtain an optimal classification system. On the other hand artificial neural networks (ANNs) are capable of managing this user-controlled tuning automatically by applying their learning strategy in a self organizing optimal manner. Besides these aspects concerning the implementation and design of the evaluation module, in practical applications, the interpretability of the system and the possibility to include available expert knowledge is of outmost importance. Standard neural networks, e.g., the MLP, cannot provide these essential features. In a MLP, the inherent decision knowledge is implied in specific weight coefficients of the underlying network structure and cannot be accessed or supplemented easily. The so-called black-box-behavior represents a substantial restriction of standard neural networks. Desirable is a feature evaluation module that stores the decision knowledge in an interpretable and modifiable form. In this respect FL provides an ideal tool for realizing feature evaluation modules. FL gives the possibility to describe knowledge by linguistic rules like, e.g., if Feature 1 is Feature 2 is Feature is then Process Phase and and

(9)

The implementation of a fuzzy module for feature evaluation can be very difficult with an increasing number of features taken into account. The problem of finding appropriate membership functions and rules is often a tiring process of trail and error. Just like linear classifiers, FL systems require in contrast to ANNs manual tuning to obtain good classification results. In order to

574

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001

Fig. 9.

Normalized measured control variables over a whole charge.

Fig. 8.

Scheme of a glass tube manufacturing plant.

automate the design phase of the FL system in the proposed diagnosis scheme, neuro-fuzzy approaches are used for designing FL feature evaluation modules. Basically neuro-fuzzy approaches can be understood as the employment of learning strategies derived from the domain of neural network theory to support and accelerate the development of a FL system. Several neuro-fuzzy concepts have been described in the literature, in the proposed diagnosis scheme we follow the NEuro Fuzzy CLASSification (NEFCLASS) approach proposed by Nauck et al. [20]. The NEFCLASS model provides a neuro-fuzzy classification approach derived from the generic fuzzy perceptron [19]. The structure of the model is illustrated in Fig. 7. The neuro-fuzzy model is characterized by a three layer topology. The input nodes I in the input layer are connected by fuzzy sets with the rule nodes R in the hidden layer. For semantical reasons, each rule unit is assigned to a single output node C, in the output layer, in order to avoid weighted rules. These weights are fixed to one. To obtain an optimal classification result, the learning algorithm creates the rules and adjusts the fuzzy sets from training examples. For initialization, the user has to define the initial fuzzy sets for partitioning the domains of the several inputs and the maximum number of rules that may be created in the hidden layer. After training, the system corresponds to a simple fuzzy system, the classification knowledge can be easily accessed and extended by the user. A detailed description of the NEFCLASS system and the implemented learning strategies is given in [20]. III. APPLICATION TO AN INDUSTRIAL BATCH PROCESS The proposed supervisory control concept has been applied to an industrial glass tube manufacturing plant shown in Fig. 8.

Fig. 10.

Developed control scheme for an industrial batch process.

During operation of the process, a glass cylinder is fed with constant speed into a furnace where it is heated up to its softening temperature and pulled out with a controllable speed. Simultaneously, an inert gas is blown into the cylinder in order to create a difference pressure. The controlled geometrical variables are the tube diameter and the wall thickness. The process is characterized by the following difficulties: 1) high complexity of system dynamics due to a highly nonlinear rheology; 2) nonlinear coupled multi-variable system behavior with time varying, distributed parameters; 3) large time varying dead times caused by large sensor distances and changing pulling speed; 4) stochastic disturbances caused by inhomogeneous material; 5) large drift disturbances caused by thermal effects during start, transient, and end phases of the process. Fig. 9 gives an overview of the normalized measured control variables diameter, wall thickness, and temperature over a whole charge. The process can be divided into three major process phasesthe start, stationary and transient phase. Particularly the transient process phase, resulting from a weld in the raw glass cylinder, is characterized by heavy disturbances of the process variables.

FREY AND KUNTZE: A NEURO-FUZZY SUPERVISORY CONTROL SYSTEM

575

Fig. 13. phase.

Generated feature setpoint diameter within start and transient process

Fig. 11.

Detecting welds with a WNN based on the measured diameter.

Fig. 14. Generated feature setpoint temperature within start and transient process phase. Fig. 12. Generated feature weld within start and transient process phase.

A. Supervisory Control Concept In the last years several control strategies for the different process phases have been developed. A detailed discussion of the FL controllers is presented in [21], [22]. 1) A FL start-up control strategy that brings up the process state as fast as possible to the desired setpoints without overshooting. 2) A stationary FL control concept that is applied in the stationary process phases in order to minimize stochastic disturbances while considering the extraordinarily large dead times. 3) A heuristic FL control concept for transient process phases with heavy disturbances in order to minimize the waste within this period. So far these control schemes have been manually activated by operators. Often the operators have been overwhelmed by

the task of choosing the adequate time instant for activating the several control schemes. Therefore, it was of utmost importance to develop a supervisory control scheme that is capable of activating automatically the different process phase, specific low-level controllers. Fig. 10 shows the basic structure of the developed control scheme. The lower level contains the well-tried controllers of the different process phases. Based on the identified actual process phase these control strategies are activated by the mode selector [5]. To achieve fast recognition of the process phases only the measured sensor signal diameter and temperature are used as inputs to the diagnosis module. The measured wall thickness is ignored due to its large measuring dead time. B. Process Diagnosis The functionality of the developed control scheme depends strongly on the reliable identification of the actual process phase

576

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001

Fig. 15.

Process phase classification based on a neuro-fuzzy feature evaluation module.

by the diagnosis scheme. Due to the complexity of the considered industrial batch process, only the functionality of a simplified diagnosis module for detecting the three process phases will be outlined. As already mentioned above, transient process phases result from welds in the raw glass cylinder. A typical time response of the diameter within a transient process phase is shown in Fig. 11. In order to detect such typical patterns in the diameter a WNN signal classifier is trained, the generated feature weld (FW) testifies if the pattern occurs in the considered signal or not. The implemented WNN contains five wavelet nodes and four, respectively, one sigmoid neuron in the hidden and output layer. To obtain a reliable pattern classification result, several training patterns have been extracted from different recorded process cycles. With respect to the supervised learning algorithm of the WNN, all patterns are assigned to corresponding output values, i.e., if the pattern results from a weld, the desired output FW of the network is 1, otherwise 0. Fig. 12 shows the simulated feature FW over a whole charge. Two welds are detected; the firstweld occursduringthestart phase and the second one in the transient process phase. The generated feature FW is capable of detecting welds, but it is not sufficient to distinguish between the start and transient process phase. To determine the transient phase and the start phase correctly more features have to be incorporated into the feature evaluation module. To obtain more features the process variables diameter and temperature are analyzed by FL setpoint evaluation. As an example, the following set of fuzzy is implemented in the setpoint detection module of the diameter if if if Diameter is small, then Setpoint is N Diameter is between, then Setpoint is Y Diameter is larger, then Setpoint is N

Fig. 16. Normalized measured control variables and detected process phases over a whole charge.

(10)

Fig. 15 illustrates the implemented neuro-fuzzy module. The three generated features FW, FSD and FST, are used as inputs to the process phase evaluation module. Before training, the system was initialized with two fuzzy sets per input and three outputs corresponding to the process phases to be classified. The training data is composed out of several different recordings from measured process data. As an example, in Fig. 15 in the left column, the three features within the three different process phases are displayed. During the training phase, the learning algorithm generated five rules and adjusted the fuzzy sets until the classification performance goal was reached. After training, the evaluation module represents a simple FL system, it can be interpreted in the form of rules, e.g., the rule for classifying the transient phase is if FW is Y FSD is N FST is Y, then transient process phase

Figs. 13 and 14 shows the generated features setpoint diameter (FSD) and setpoint temperature (FST) within the start and transient process phase. As can be seen, the feature evaluation step leads to a significant compression of the signal information from which the different process phases can be identified by the following FL evaluation module.

(11)

FREY AND KUNTZE: A NEURO-FUZZY SUPERVISORY CONTROL SYSTEM

577

In the left column of Fig. 15, the simulated output of the FL module is shown. As can be seen, the classification result is very accurate compared with the desired response. Fig. 16 shows the complete result of the process phase classification over a whole charge. The presented diagnosis module for detecting the different process phase is exemplary. The developed and implemented diagnosis systems uses 16 different features generated out of 12 measured process signals and contains seven neuro fuzzy modules to determine several process phases and states of the manufacturing process. IV. CONCLUSION In this paper, a new neuro-fuzzy based approach for online process diagnosis is outlined that represents the key component of a two-stage FL based supervisory control scheme. It relies on a reliable identification of the actual process phase and the decomposition of the complex control problem into a series of smaller and more transparent ones. Thus, for each identified process phase by means of a fuzzy logic mode selector the adaptation of control parameter sets without structural changes and the switching of different situation-specific low level controller structures can be achieved. The concept has been successfully applied to an industrial batch process in the glass industry. REFERENCES
[1] L. Zadeh, Outline of a new approach to the analysis of complex synthesis with a fuzzy logic controller, Int. J. Man-Mach. Stud., vol. 7, no. 1, pp. 113, 1975. [2] E. H. Mamdani, Application of fuzzy logic to approximate reasoning using linguistic synthesis, IEEE Trans. Computers, vol. C-26, pp. 11821191, Dec. 1977. [3] M. Sugeno, Ed., Industrial Applications of Fuzzy Control. Amsterdam, The Netherlands: North Holland , 1985. [4] H.-J. Zimmermann and v. Altrock, Fuzzy LogicAnwendungen. Mnchen, Germany: Oldenburg-Verlag, 1994. [5] P. P. Bonissone, V. Bandami, K. H. Chiang, P. S. Khadkar, K. W. Marcelle, and M. J. Schutten, Industrial applications of fuzzy logic at General Electric, Proc. IEEE, vol. 83, pp. 450465, Mar. 1995. [6] M. Koch, T. Kuhn, and J. Wernstedt, Fuzzy Control. Mnchen, Germany: Oldenburg-Verlag, 1996. [7] P. M. Frank, Diagnoseverfahren in der Automatisierungstechnik, in atAutomatisierungstechnik, Feb. 1994, pp. 4764. [8] A. S. Willsky and H. L. Jones, A generalized likelihood ratio testing approach for the detection and estimation of jumps in linear systems, IEEE Trans. Automat. Contr. , vol. AC-21, pp. 108112, Jan. 1976. [9] R. Isermann, Ueberwachung und Fehlerdiagnose. Duesseldorf, Germany: VDI-Verlag GmbH, 1994.

[10] K. Kroschel, Statistische Nachrichtentheorie, New York, Berlin, Germany; Vienna, Austria: Springer-Verlag, 1996. [11] C. F. Frey and H.-B Kuntze, A neuro-fuzzy supervisory control system for industrial batch processes, in Proc. 9th Int. Conf. Fuzzy Systems FUZZ-IEEE 2000, San Antonio, TX, May 710, 2000, pp. 116121. [12] A. K. Louis, P. Maa, and A. Rieder, Wavelets, in Teubner Studienbcher, Germany: Stuttgart, 1998. [13] G. Strang and T. Nguyen, Wavelets and Filter Banks. Cambridge, MA: Wellesley-Cambridge, 1996. [14] S. Mallat, Wavelet Signal Processing, New York: Academic , 1996. [15] Y. Mallat, D. Coomans, J. Kautsky, and O. De Vel, Classification using adaptive wavelets for feature extraction, IEEE Tran. Pattern Anal. Machine Intell., vol. 19, Oct. 1997. [16] Q. Zhang and A. Benveniste, Wavelet networks, IEEE Trans. Neural Networks, vol. 3, Nov. 1992. [17] H. Dickhaus and H. Heinrich, Classifying biosignals with wavelet networks, IEEE Eng. Med. Biol. Mag., vol. 15, no. 5, Sept./Oct. 1996. [18] H. Kiendl, Fuzzy Control Methodenorientiert. Mnchen, Germany: Oldenburg-Verlag, 1997. [19] D. Nauck, Neuro-fuzzy systems: Review and Prospects, in Proc. 5th Euro. Congr. Intelligent Techniques and Soft Computing (EUFIT97), Aachen, Germany, Sept. 811, 1997, pp. 10441053. [20] D. Nauck, F. Klawonn, and R. Kruse, Neuronale Netze und Fuzzy Systeme, Germany: Vieweg-Verlag, 1994. [21] M. Sajidman, H.-B Kuntze, and A. Jacubasch, A fuzzy-logic concept for highly fast and accurate position control of industrial robots, in Proc. IEEE Int. Conf. Robotics and Automation ICRA95, Nagoya, Japan, May 21-27, 1995. [22] C. W. Frey, M. Sajidman, and H.-B Kuntze, Ein neuro-adaptives regelungskonzept mit on-line-fhigem kalman-filter lernverfahren fr stochastisch gestrte nicht-lineare prozesse, in VDI/VDE-GMA-Tagung Computational Intelligence, Berlin, Germany, Mar. 3/4, 1998.

Christian W. Frey was born in Freudenstadt, Germany, in 1968. He received the Master degree (Dipl.-Ing.) in electrical engineering from the University of Karlsruhe, Karlsruhe, Germany, in 1995. Since 1996, he has been with the Fraunhofer Institute for Information and Data Processing, Research Group Industrial Sensor and Control Systems, Karlsruhe. His research interests include the application of neural networks, fuzzy logic systems, and wavelets to a wide range of industrial diagnosis and control systems.

Helge-Bjrn Kuntze received the Master degree (Dipl.-Ing) in mechanical engineering and the Ph.D. degree (Dr.-Ing.) in control systems theory from the Technical University of Dresden, Dresden, Germany. From 1977 to 1980, he was with the Research Laboratory of Asea Brown Boveri Cie, Mannheim, Germany. Since 1981, he has been with the Fraunhofer Institute for Information and Data Processing where he is both Deputy Head of the Business Area Production Optimization Systems and Manager of the research group Industrial Sensor and Control Systems. His present research interests include the application of neural networks and fuzzy logic systems to a wide range of industrial surveillance, diagnosis, and control systems.