Nonlinear Dynamic Fault Dignosis Method Based on DAutoencoder

In order to detect faults in chemical industry process effectively, a nonlinear dynamic fault detection method using DAutoencoder is proposed. Correlation analysis is applied firstly to establish autoregressive model. Then weights of Autoencoder can be obtained by improved differential evolution (DE) algorithm. Meanwhile, the least square method is used to prune nodes every layer to simplifying network structure. Features of training sample and reconstruction residuals can be extracted by DAutoencoder. Monitoring statistic is developed and confidence limit is computed by kernel density estimation at last. According to correlation between measured variables and nonlinear features, the contribution of each variable is calculated to give contribution plots. Simulation results of Tennessee Eastman (TE) process show that DAutoencoder-based method is more effective than KPCA (Kernel Principal Component Analysis) for process monitoring, and it can also realize fault identification.