Research And Application Of An Adaptive Manifold Learning Algorithm For Industrial Process Fault Detection

Fault detection technology is the key to ensure the safe and stable operation of modern industrial processes.The research of effective and feasible fault detection methods is of great significance for the smooth operation of industrial processes and the guarantee of production quality.With the arrival of the big data age,the mass production and operation data in the modern industrial process are recorded and preserved in real time.The fault detection method based on multivariate statistical process monitoring(MSPM)has been widely studied and applied.However,the actual industrial process is complex and nonlinear,and the traditional fault detection methods do not consider these factors.The MSPM method extracts global structure information and ignores the local features of the data.Moreover,the existing methods do not have the online learning ability,and the established fault detection model does not have the self-adaptive ability.It cannot effectively represent the running state of the system when monitoring objects change over time.Aiming at these problems,based on previous studies,an adaptive manifold learning algorithm is proposed for fault detection in industrial processes.And carries out case simulation to verify the effectiveness of the method.(1)The background and significance of this study are introduced,and describes the main research content and current situation in the field of industrial process fault detection.According to the data characteristics of the actual industrial process,the limitations of the traditional fault detection methods are analyzed in this thesis.The manifold learning method is introduced to analyze the advantages of manifold learning in fault detection,and illustrate the significance and value of adaptive manifold learning algorithm in fault detection.(2)The traditional fault detection methods do not consider data nonlinearity,ignore the local information of data and do not have online learning ability,so this thesis proposes an adaptive neighborhood preserving embedding(ANPE)algorithm with online learning capability.The overall manifold structure is extracted by extracting local structure features of data.Meanwhile,the algorithm has the online learning ability,and it can update the model adaptively to adapt to the new working conditions.(3)In view of the problem of the unicity of the update strategy of ANPE algorithm,the updating strategy of the algorithm is improved and the improved ANPE algorithm is proposed.The new algorithm combines the new sample's own monitoring results and updates the model using the data which independent from the training samples in the normal data.The performance of the ANPE algorithm and its improved algorithm are verified by the simulation of the cable joint of the ring network cabinet and the chemical process.The simulation experiment shows that the ANPE and its improved algorithm with online learning ability can effectively detect the fault in the industrial process.