Research On Artificial Immune System Based Fault Diagnosis Method

With the increasing complexity of modern industrial systems, the research on process monitoring technique has attracted more attention in recent years. An Artificial Immune System based process monitoring technique is investigated in this dissertation, which aims to explore the potential of negative selection and clonal selection.A new fault detection method is proposed in the case of the training data set contains normal data only and once the faulty data are included in the training data set, the corresponding diagnosis method is developed.The fault detection method can be described as three stages: detector generation, fault detection and detector update, in which the last two stages are carried out by turns. In detector generation, negative selection is performed with the normal data as self to generate initial detector set. In fault detection, the detectors are examined with the new observations. If a detector gets activated with current observation, an abnormal or a fault is indicated. Then clonal selection is introduced to update the detector set dynamically and increase the adaptability of the method.In the case of samples in different fault classes are also contained in training set, fault diagnosis can be implemented. Two strategies, negative selection based and fault sample based strategies are applied to generate the initial detectors. The shortest distance rule and the largest quantity rule are employed to train the initial detectors derived based on negative selection in order to form different fault classifiers. Since the detectors may overlap each other and a sample could be detected by more than one detector simultaneously, the affiliation of the detected sample is discussed in this thesis.The performances of the proposed methods are demonstrated by the applications on the TE simulation data. Experimental results indicate that the fault detection method performs well, especially after the detector set update mechanism has been introduced. While the performance of the proposed fault diagnosis method varies with different fault classes, because of the different representation and distribution characteristics of training samples. However, as compared with the performances of PCA based methods, the proposed methods are superior both in detection and diagnosis.