Research On Data-driven Fault Diagnosis Method Based On Deep Belief Network

With the advent of technology, modern equipment of aeronautics and spaceflight becomes increasingly complicated, intelligent and integrated. Traditional diagnostic methods based on artificial reliability engineering have been unable to adapt the changes of equipment. Aiming to improve the safety and reliability of complex systems, under the background of the development in data acquisition and storage technology, it has theoretical and practical significance in researching on intelligent and data-driven fault diagnosis technology for complex systems. Deep Belief Network, as a new method in the field of machine learning, has already obtained brilliant achievements in some domains such as image recognition due to the strong ability of automatic feature extraction. But it has yet to be developed of the application in fault diagnosis. This paper starts from the background of complex system fault diagnosis. Because of the deficiency that traditional data-driven methods do not have enough capacity to learn the complex non-linear relationships in fault diagnosis issues, the author carries out research in data-driven fault diagnosis based on deep belief network.Firstly, the capacity of DBN's feature extraction and classification is discussed on the basis of the principle analysis. The ability of feature extraction is studied through researching a single Restricted Boltzmann Machine, which primarily includes different number of iterations and different hidden layer nodes. Also, the classification performance is described under different depth of DBN. Then, the author studies the single sensor fault diagnosis method relayed on DBN. This method combining feature extraction with state recognition can automatically extract fault features from the original time-domain signal. The ability of fault feature extraction and status recognition is verified and analyzed with simulated data sets and bearing data sets. Finally, the author studies the multi-sensor information fault diagnosis based on DBN from the perspective of multidimensional time-series classification. For the current multidimensional time classification does not consider the importance of the connection between multidimensional time-series variables and the problem of the dimension of different sample matrices is not corresponding complete after reduction. Common Principal Component Analysis(CPCA) and Dynamic Time Warping(DTW) methods are introduced into the DBN model, a novel multidimensional time sequence fault diagnosis method data-driven fault diagnosis method is proposed named based CPCA_DTW_DBN, which has achieved better classification performance in the publicly available data sets and Tennessee Eastman process fault case.The effectiveness of the proposed method is validated using datasets from publicly data sets, simulations and bearings. The diagnosis results show that the proposed method is able to not only accomplish fault feature extraction and diagnosis from raw signals effectively under various operating conditions, fault locations and fault degree, but also obtain superior diagnosis accuracy compared with the existing methods. The proposed approach of CPCA_DTW_DBN can be effectively applied to the multisensor equipment fault diagnosis.