An Adaptive Interval Forecast CNN Model for Fault Detection Method

Abstract

The machine fault detection (MFD) is critical for the safety operation of the petrochemical production. Aiming to automatically optimizing the pre-warning bounds of the control chart, an interval forecasting convolutional neural network (IFCNN) model has been proposed to forecast the warning interval of the signal with the raw dynamic data. Essentially, the IFCNN model is an improved convolutional neural network with dual output value to construct the warning interval directly and adaptively. To guide the model to learn the interval automatically during the model training, the loss function is customized to improve the fault detection accuracy. The proposed method is compared with the fixed threshold and the adaptive interval method with exponentially weighted moving average on a petrochemical equipment data set. The results indicated that the proposed method is of stronger robustness with lower failure rate in the fault detection of the petrochemical pump.