Democritus University of Thrace (DUTH)
Neural network models for system identification, control and fault detection
Abstract
dc:descriptionThe demand for accurate and reliable system identification, control and fault detection is raising over the past years, due to the increasing complexity and cost of many industrial processes. The proposed research thesis tries to satisfy this requirement and so is devoted to devising system identification, control and fault detection approaches utilizing modern techniques based on neural networks and algorithms of deep learning. A common practice in many industrial applications is the build of mathematical models of dynamical unknown systems and fault diagnosis schemes, based on observed data from sensors attached to system parts. Modern neural networks and deep learning approaches fit with this practice as extractors of features or attributes from raw data, contrary to conventional machine learning (ML) techniques in which manual extraction of features, from prior expertise knowledge, is necessary. Therefore, the dissertation exploits the efficiency of deep learning and modern neural networks to represent complex nonlinear dynamics mappings and presents innovative techniques in system identification, control, and fault detection, enhancing the performance and dealing with confirmed drawbacks and defections. The first part of the dissertation, which is made up of four chapters, provides a literature study on neural networks, on deep learning techniques, on system identification and control and on methods of fault detection. The review equips the reader with the necessary background knowledge, presenting both well-established techniques and recent state of the art algorithms and approaches. So, the first review part of the dissertation assists as a basis, establishing the required knowledge for the reader to understand the motivation of the thesis. Subsequently, the second part of the thesis presents novel methods of system identification and control. So, an innovative fusion neural architecture and the use of a novel Lyapunov theory-based algorithm are proposed in chapter 5, for the online approximation of the dynamics of unknown nonlinear systems. The proposed neural system, in combination with the proposed update rule of the neural weights, achieves fast convergence of the identification process, ensuring at the same time stability of the error system in the sense of Lyapunov theory. The fusion neural system combines the features that are extracted from two independent neural streams, a feed-forward and a diagonal recurrent one, satisfying different design criteria of the identification task. Simulation results are included where we show the superiority of the proposed scheme over competitive state of the art methods. Also, additional experiments demonstratethe effectiveness in cases of parameter variations and additive noise. In the sequel of the second part of the thesis (chapter 6), an Adaptive One-Step Model Predictive Controller is introduced, overcoming the drawbacks of of-line learning in model predictive control systems. Indeed, neural network model predictive control has been widely used in the field of automatic control. However, the off-line training of the neural predictor and the appearance of disturbances and parameter variations, during the control phase, affect the system output by inducing steady state error. To ensure the convergence of the adaptive process, we use Lyapunov stability theory, which guides the updating of the weights of the deep neural network. Likewise, a Lyapunov based algorithm guides the updating of the control signal of the one-step model predictive controller. Simulation results for two cases demonstrate the effectiveness of the presented control scheme. Another common drawback of many closed-loop feedback controllers is the transfer of additive noise and disturbances in the output of the system. So, the last chapter of Part 2 (chapter 7) is devoted to the above problem with the proposition of an artificial neural network controller, for circumstances where noise robustness is a crucial matter. The presented artificial controller has many applications in diverse areas where it is important to keep accurate track of a specified reference output under the presence of noise. The proposed neural controller consists of two parts, an Auto-encoder which transforms the input data in a robust to noise representation and a Recurrent Neural Network, which is trained initially off-line with data taken from conventional controllers and subsequently online in an indirect feedback control scheme. The stability of the overall scheme is guaranteed by using Lyapunov stability arguments, which guide the convergence of the neural weights. The performance of the proposed artificial controller is tested on the control of a dc motor system. The simulation results show a satisfactory reduction of noise in the controlling process and accurate output track, as it is desired. Moving on Part 3, the current thesis contributes on the field of fault diagnosis with the introduction of two deep learning fault detection schemes. At first place in chapter 8, a one-class fault detection scheme for multi-dimensional problems is proposed, based on the unsupervised training of a Generative Adversarial Network. The generator tries to learn the manifold of normal behavior of the process, while the final decision of fault occurrence is taken from the discriminator. The network architectures of the discriminator and the generator and the hyper-parameters, that are used during training, are crucial for the stability of the model during training. To ensure system convergence and to enhance the accuracy of the unsupervised classifier, a model selection algorithm is used. The latter evaluates each candidate generative model on a validation set of the normal training dataset, based on a proposed performance metric which uses the reconstruction cost of an Auto-encoder. Finally, experimental results illustrate the superiority of the proposed system over One-Class SVM, Isolation Forest algorithm and an Autoencoder based algorithm, three state of the art anomaly detection methods, on two simulation cases. In the second chapter of Part 3, the Attentive Dense Convolutional Neural Network is introduced and employed for the problem of bearing fault detection. The latter considers the temporal coherence of the data samples, by the combination of Dense Convolutional blocks with an attention mechanism. In more details, the presented deep learning scheme exploits the ability of a Densely connected Convolutional Neural Network to extract features, belonging to different level of representation. Furthermore, the simple attentive mechanism, which consists of a feed-forward NN followed by a softmax layer, enhances the extracted information with the focus on specific spatial regions of the features and subsequently arranges the temporal coherence with the estimation of a simple weighted average output map. The proposed neural scheme has fewer unknown learning parameters and achieves accurate results with less training data as it appears from simulation cases on two famous rolling bearings fault detection benchmarks.Finally, the contribution, the overall evaluation and the feature research directions are presented in the final concluding section of the thesis.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2020
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Plakias, Spyridon
- Πλακιάς, Σπυρίδων
Subjects
dc:subject × 15- Νευρωνικά δίκτυα
- Αναγνώριση συστήματος
- Νευρωνικός έλεγχος
- Μη γραμμικά συστήματα
- Ανίχνευση βλαβών
- Neural networks
- Deep learning
- System identification
- Neural control
- Nonlinear systems
- Fault detection
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
dc:rights- Statement dc:rights
-
- BY
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/47209