Research on fault prediction models based on deep learning

Abstract Failure prediction is of great significance to modern industry. Effective fault prediction can reduce losses by identifying potential problems with equipment in a timely manner. In actual industrial processes, some failures often occur when not only a single variable changes. These variables are often accompanied by certain trends. To address such problems, this paper combines the Chemical Reaction Optimization (CRO) algorithm with the Attention Mechanism (ATT) and the Long Short-Term Memory (LSTM) network. The advantages of LSTM are utilized for future time series prediction, after which the predicted future values are used as inputs to a trained eXtreme Gradient Boosting (XGBOOST) model to predict the future state of the device. The method is validated on the Tennessee Eastman (TE) data set, and the results show that the CRO-ATT-LSTM proposed in this paper reduces the error rate by 25% compared to the traditional LSTM. The diagnostic model XGBOOST improves the accuracy by 17.5% and 10% compared to BP and SVM, respectively; the fault can be detected two time windows earlier.