Research And Application On Hybrid Intelligent Diagnosis Method Based On Adaptive Resonance Theory

Diagnosis and assessment of the condition and performance of the complicated mechanical system is an important component of advanced manufacturing system, is also an important way that ensure the stability of production capacity and quality control. When the intelligent fault diagnosis technique based on the traditional neural network is used to diagnose and assess the conditions of mechanical equipment, the neural network often undergo the stability-plasticity dilemma. The neural network based adaptive resonance theory can solve the problem. In order to make full use of the merit of the neural network, and make the neural network apply to the diagnosis and assessment of the condition and performance of the mechanical equipment, the adaptive resonance theory is used as foundation in this dissertation, and the hybrid intelligent fault diagnosis technique based the adaptive resonance theory is studied, and the technique is applied to the diagnosis and assessment of condition and performance of the self-made sub-high speed feed system.Firstly, the basic theory and character of fuzzy ART are analyzed. According to the drawback that performance of fuzzy ART is affected by input order of training samples and merits of ART and similarity classifier based on Yu's norm, a new unsupervised ART-similarity classification method based on Yu's norm is proposed. Lots of feature parameters are extracted from the original signal and used to describe the condition and performance of the mechanical equipment, and the sensitive features are selected by the distance discrimination technique. Then they are input the classifier to diagnose the fault of equipment. By the application to the fault diagnosis of gear, the results indicate that the diagnosis accuracy is very high, and the performance of the ART-similarity based on Yu's norm is superior to that of the fuzzy ART neural network.According to the problem that the "hard-competition" classification mechanism of the fuzzy ARTMAP neural network brings, an improved fuzzy ARTMAP neural network based on hybrid classification mechanism is suggested. After the fuzzy ARTMAP is trained, the center of each pattern node can be obtained through the data samples that the pattern node encloses. Given a data sample, it can be classified by two classification mechanism, namely, classification method based on Yu's norm and original classification mechanism of fuzzy ARTMAP. Using the fault diagnosis of gear as an example, the diagnosis results verify the effectiveness of the proposed method, and the generalization is verified by the bootstrap method.Because of the drawback of the fuzzy ARTMAP neural network, viz. its performance is affected by the input order of the training samples, and lack of information that the feature parameters describe the condition and performance of the equipment make the diagnosis accuracy is not high, a selective ensemble fuzzy ARTMAP based multi-symptom-domain is proposed. Many feature parameter sets are extracted to represent the condition of equipment from different point of view, and the sensitive feature parameter sets are selected from each original feature parameter set, then they are input fuzzy ARTMAP respectively. The number of the neural network is obtained by the correlation method. And the final diagnosis result is obtained by the Bayesian belief method. By the application to the fault diagnosis of bear different category and severity, the diagnosis results verify the effectiveness and generalization and robustness of the proposed diagnosis method.Different feature parameters have different importance degree to the different condition and performance of mechanical system.In order to strengthen the compactness of knowledge that each node of fuzzy ARTMAP contains, an intelligent diagnosis method based weighted fuzzy ARTMAP is proposed. Time-domain feature parameters and frequency-domain feature parameters are extracted to represent the condition of mechanical equipment. And the sensitive feature parameters and the corresponding weight coefficient can be obtained by the modified distance evaluation technique. Then they are combined with the weighted fuzzy ARTMAP to identify the fault and performance of the mechanical equipment. Using the fault diagnosis of bearing as an example, the diagnosis results verify the superiority of the suggested diagnosis method.When fuzzy ARTMAP neural network is utilized to diagnose the fault, it often suffers from the "black box" problem. Especially the data samples and their corresponding class names are not known, a fault diagnosis method and rule extraction method based fuzzy ART&ARTMAP is proposed. When the training samples and their corresponding class names are known, they can be used to train the fuzzy ARTMAP. The trained neural network can classify testing samples, and diagnosis rules can be obtained. At the same time the testing samples whose class names are not known can be classified into one class. For these data samples, they can be classified and their corresponding class names can be obtained by the fuzzy ART. Then, these data samples and their class names can be utilized to train fuzzy ARTMAP newly, and the corresponding diagnosis rule can be extracted through the prediction samples. By the application to diagnosis of bearing different damage level fault, the effectiveness of the suggested diagnosis method is verified.Every prediction method has its own advantages and disadvantages, the performance of RBF neural network is affected because of that the number of nodes in the hidden layer can not be ascertained dynamically. For these, a combined prediction method based ART-RBF neural network is proposed. Firstly time varying autoregressive model and LS-SVM are utilized to predict the same time series respectively. Then these prediction results are used as the input of the ART-RBF neural network to obtain the final prediction results. Using the Mackey-Glass time series as an example, the superiority of the proposed combined prediction method is verified.Finally, using the HUST-FS-001 feed system as an object, three diagnosis methods, namely ART-similarity based Yu's norm and improved fuzzy ARTMAP and selective ensemble fuzzy ARTMAP based multi-symptom-domain, are applied to the diagnosis of the different preload conditions of bearing. The weighted fuzzy ARTMAP and fuzzy ART&ARTMAP are used to identify the axis position error level and extract the diagnosis rules respectively, the combined prediction method is utilized to predict its trend. It is shown that these diagnosis methods are effective to diagnose and assess the condition and performance of feed system.