A Model Transfer Learning Based Fault Diagnosis Method for Chemical Processes With Small Samples

Traditional fault diagnosis methods relies on sufficient fault samples, but it is unrealistic since the fault is a low possibility event in real industrial scenes. To address the above issue, this paper proposed a fault diagnosis method for chemical processes with small samples. First, a data self-generating-based transfer learning (DSGTL) method is presented to expand the fault samples. The characteristic of fault data is learned by adversarial relation and transferred to the generated data. Moreover, a model-based transfer learning strategy is adopted to improve the robustness of the proposed method to the quality of generated data. Second, the sample reconstruction-based convolutional neural network (SR-CNN) is proposed which adaptively extracts features from both spatial domain and time domain and identifies the fault type of industrial process with small samples. Finally, the experimental result of the Tennessee Eastman (TE) process proves the validity and the feasibility of the proposed method.