A Multi-task Learning Approach For Chemical Process Abnormity Locations And Fault Classifications With Applications

Currently,it is seen that deep learning-based fault classification methods have received increasing attentions.However,the data characteristics of high dimensionality,strong noise and insufficient labels discourage the use of deep learning based fault classifications in chemical processes.In response,this thesis investigates a multi-task learning approach to chemical process fault classifications.Aiming at data denoising,abnormity localizations and fault classifications,a multi-task learning network is established which is able to achieve satisfied classification results in the case of insufficient fault labeled samples.The main work accomplished and the results obtained in the thesis are as follows.1.A contribution-based dynamic task weight(DTW)algorithm is proposed for multi-task learning networks,which can make two learning tasks with large differences converge simultaneously;it is applied to chemical process sample denoising and fault classification,thus giving a multi-task learning-based chemical process data denoising method,which adds an auxiliary task of data denoising to the main target task of network learning and can The model has robustness by adding regularization effect.2.A multi-task learning cascade method for locating anomalies and classifying faults in chemical processes is proposed.Firstly,using the comparative Granger causality analysis,process variable nodes in the fault tracing path are located with anomaly labels.Subsequently,a multi-task learning cascade network is created,which involves the secondary learning task of abnormity localizations and the primary learning task of fault classifications is created.The feature sharing mechanism between the two tasks can help resolve the problem of insufficient labeled fault samples.Additionally,the DTW algorithm is improved so as to balance the learning weights of data denoising,abnormity localizations and fault classifications.3.The proposed method is applied to an actual chemical coal gasification process.Historical data of the coal gasification plant are collected,and fault classification experiments are carried out under different working conditions,achieving satisfactory results.Moreover,the proposed method is compared with conventional fault classification methods,demonstrating tangible benefits of the contributions.