Fault Detection Based on Hierarchical Sampling Feature Extraction Network for Multirate Process

Due to limitations of measurement technology and hardware costs, monitoring all industrial variables with a uniform sampling rate is unfeasible. Therefore, while the multirate process is common, single-rate based methods often perform ineffectively in such scenarios. To realize effective multirate process fault detection, a novel hierarchical sampling feature extraction network (HSFEN) is proposed. The HSFEN not only alleviates the influence of higher-rate variables on lower-rate ones during data imputation, but also rectifies the neglect of sampling features encountered in other methods. And the HSFEN consists of two modules: multiscale imputation with interwoven aggregation and sampling-perception-driven transformer. Then, an adaptive feature enhancement learning mechanism is developed to balance the weight bias between variables owing to the differences in sample size. Finally, an inverter system is adopted to demonstrate the effectiveness of the proposed method, showing an average reduction of 3.69% in false alarm rate (FAR) and an average improvement of 7.95% in fault detection rate (FDR) compared to other classical approaches.