基于特征融合与时间卷积自编码器的工业过程故障检测

打开文本图片集

中图分类号：TP277 文献标志码：A DOI：10.3969/j.issn.1003-9015.2025.04.011

Abstract： In order to solve the problem of multi-scale timeseries feature extraction of industrial process data，a fault detection method based on feature fusionand temporal convolutional autoencoder was proposed.Firstly，the multi-layer temporal convolutional network structurewasused toextractfeatures fromtheinputtimeseriesatdiferent scales，andamulti-scale temporal convolutional autoencoder wasconstructed.Secondly，afeature fusion module basedon eficient channelatention wasdesigned，which was added tothe temporalconvolutional autoencoder through jump joining and connected the temporal series features of diferent scales across channels.It generated corresponding weights to weight and fusethe features，soas to capture richer temporal informationand enhance the model'sdiscrimination between normal sequenceand abnormal sequence reconstruction eror.Finally，the statistics were established byreconstructing the error，andthe kerneldensity estimation wasused todeterminethe control limit to realize fault detection.The proposed detection method is applied to numerical cases and Tennesse-Eastman process，and the experimental results show that the proposed method has good fault detection performance，which can provide a certain reference for fault detection in complex industrial processes.

Key words： fault detection; temporal convolutional networks; autoencoder； attention mechanisms ; feature fusion

1前言

随着工业过程自动化、信息化水平不断提高，生产过程中积累了丰富的数据信息，为基于数据驱动的故障检测方法提供有力支持[1]。（剩余17040字）