Research On Metric Learning And Similarity Measure Of Time Series Based Fault Diagnosis Method

Along with the fast development of industrial production technology, modern industrial systems have the common characteristics of sophisticated and large-sized. Due to the increasing amount of the running data, the nonlinearity of the system and various disturbances, the industrial systems are easier to be affected by improper conditions nowadays which become principal factors prevent the safety and reliability of modern industrial systems. Meanwhile the difficulty of mechanism modeling for industrial systems makes the traditional methods based on analysis model no longer applicative. To ensure the safety of industrial systems, many state variables are observed in industrial production for a long period. Motivated by this observation, it is of great interest for us to judge system operating status only based on running data which can enhance productivity and ensure quality of products. Therefore, academic study of fault diagnosis systems based on data-driven methods is of great significance in practical fields.Based on theoretical analysis of traditional data-driven fault diagnosis methods, in this thesis we firstly propose a metric learning based fault detection framework in fault detection. The proposed algorithm just needs the observational data of industry system status. By adding constraints of sample pairs, metric learning method aims at finding a metric matrix to describe the distance between samples. In order to improving the fault detection accuracy of metric learning method, we integrate metric learning method and time series method to remedy its defect in retrieving and simplifying information. To demonstrate the advantages of the fault detection method based on metric learning and time series, we carried out experiments on the dataset of TE process and compare it to a classical method PCA. The results indicate that time series method can retrieving fault information effectively and the combination of metric learning algorithm and time series method is efficient in fault classification and diagnosis for large-scaled complex industrial systems.