Research On Intelligent Diagnosis Method Based On Manifold Learning

The essence of fault diagnosis is pattern recognition and its main contents include signal acquisition, feature extraction and pattern classification. Feature extraction is the most critical and difficult section of fault diagnosis, which directly affects the accuracy of fault diagnosis and the reliability of fault early warning. Thus, it’s a big challenge to find a feature extraction technique extracting the most salient features beneficial to improving classification performance and simultaneously decreasing feature dimension under complicated operating conditions. The foundation of this dissertation is manifold learning algorithm, and it deeply explores feature extraction and diagnosis technique based on manifold learning algorithm.The fault features of complicated equipments are high-dimensional. Numerous redundant and irrelevant features may increase computation burden and degrade the classification accuracy. Hence, the diagnosis model based on Marginal Fisher analysis (MFA) is proposed. On behalf of deriving more detailed fault information, the model employs multiple signal processing methods to extract several feature parameters from various domains to reflect the operating conditions of equipments. By utilizing MFA algorithm, it extracts the most representative low-dimensional manifold features embedded in the raw high-dimensional feature set. The low-dimensional features are finally fed into K-nearest neighbor classifier to recognize different faults. By the application to the incipient fault diagnosis of rolling element bearings, it is validated that the model is feasible and effective.Aiming at the small sample size problem of pattern recognition for mechanical equipments fault diagnosis, regularized kernel MFA algorithm (RKMFA) for feature extraction and the diagnosis model based on the algorithm are presented. The model directly extracts the low-dimensional manifold features from the raw high-dimensional vibration signals, by utilizing RKMFA algorithm. The few manifold features including discriminate information are finally fed into K-nearest neighbor classifier to recognize different operating conditions. The model is separately applied to the recognition of bearing fault categories and inner fault severity. The experiment results demonstrate that RKMFA algorithm is an effective feature extraction method, and simultaneously validate its superiority. It’s costly and time consuming to collect a large amount of labeled faulty samples, so semi-supervised kernel MFA (SSKMFA) algorithm for feature extraction and the diagnosis model based on the algorithm are proposed. The model implements SSKMFA algorithm and directly learns the raw high-dimensional vibration signals. It extracts the discriminate low-dimensional manifold features and acquires the label information of the whole manifold under the guide of labeled faulty samples, by utilizing a number of cheap unlabeled faulty samples and few expensive labeled faulty samples to estimate the intrinsic manifold structure of faulty data. Thus, the cheap unlabeled faulty samples have good classification performance. The model is separately applied to the fault diagnosis of bearing fault categories, fault severity and three kinds of gear fault categories. The diagnosis results indicate that the model is able to greatly improve fault recognition accuracy, and simultaneously degrade the computational burden.The low-dimensional manifold features extracted by manifold learning algorithm are no physical meaning, so it’s not easy to understand in fault diagnosis. Aiming at the shortcoming, MFA score algorithm for feature selection and the diagnosis model based on the algorithm and support vector machine classifier is presented. The model employs multiple signal processing methods to analyze the faulty signals and acquires a high-dimensional feature set constructed by several feature parameters. By utilizing Marginal Fisher analysis score, it exploits its inherent law hidden in the original high-dimensional feature and picks out the sensitive feature subset fully reflecting the nature of the failure. The sensitive features are finally fed into support vector machine classifier to recognize different fault conditions. By the application to the fault diagnosis of bearing categories and inner fault severity, the results demonstrate its superiority.