A Novel Fault Diagnosis Method based on Bayesian Optimized Multi-Scale Convolutional Neural Network in Chemical Processes

Fault diagnosis can help operators make faster judgment and more accurate countermeasures in chemical processes. Though the traditional convolutional neural network method can extract the features of data, the in-depth and comprehensive analysis is still needed for accurately discriminating the fault types. The multi-scale convolutional neural network is proposed to solve this problem, however, with the increasing of scale, the hyper-parameters of the network are complex and numerous. The conventional manual debugging method cannot satisfy the requirements of the rapidity, reliability, and accuracy. To overcome these limitations, a novel Bayesian optimized multi-scale convolutional network fault classification method, named BOMSCNN for short, is proposed in this paper, which extracts the multi-scale features of the chemical data by multi-scale convolutional network and automatically optimizes the hyper-parameters of the complex network to reduce the workload of operators and achieve the desired results. Based on the proposed BOMSCNN applied in Tennessee Eastman process, it is verified that the accuracy of the fault classification reaches 94.64%.