软阈值时序卷积网络在冷水机组传感器故障诊断中的应用  

Application of Temporal Convolutional Network with Soft-Threshold Adaptive Module in Fault Diagnosis of Chiller Sensor

作　　者：洪琳 李冬辉 高龙 赵墨刊 HONG Lin;LI Donghui;GAO Long;ZHAO Mokan(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072

出　　处：《西安交通大学学报》2023年第2期67-77,共11页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(61873180)。

摘　　要：为了提高冷水机组传感器的故障诊断性能,提出了一种基于软阈值时序卷积网络的编码-解码器重构模型(ST-TCN),并建立基于该模型的传感器故障诊断方法。采用时序卷积网络(TCN)充分挖掘冷水机组传感器的时间相关性、热力学物理量间的数据相关性以及动态响应差异性特征。在TCN的残差块中引入软阈值自适应模块剔除冗余信息,降低噪声干扰。依托ST-TCN模型“端到端”的网络结构优势,将绝对重构残差向量与故障阈值向量进行比较,直接定位故障传感器。在实际压缩式冷水机组平台上采集传感器数据进行实验,结果表明,软阈值自适应模块能有效地增强网络模型的重构能力,从而提高故障传感器的诊断性能。以压缩机吸气温度传感器T1为例,ST-TCN的平均偏差故障识别率比改进前提升了45.9%;与其他故障诊断方法相比,所提的最新框架获得了较高的偏差故障识别率。In order to improve the capacity to identify the faults of air-cooled chiller sensor,an encoder-decoder reconstruction model based on temporal convolutional network(TCN)with a soft-threshold adaptive module(ST-TCN)is proposed,and an ST-TCN-based method for sensor fault diagnosis is established.In this method,TCN is used to fully explore the temporal correlation of the chiller sensors,the data correlation among the physical quantities,and the dynamic response characteristics of the physical quantities.In addition,a soft-threshold adaptive module is introduced into the residual module of TCN to eliminate redundant information and reduce noise interference.Relying on the advantage of the“end-to-end”network structure of the ST-TCN model,the fault sensor is directly located by comparing the absolute reconstruction error vector with the fault threshold vector.According to datasets collected from an actual air-cooled chiller platform,it is verified that the soft-threshold adaptive module will effectively improve the reconstruction performance of the network,thus facilitating diagnosis performance of faulty sensors.Taking the compressor suction temperature sensor T1 as an example,compared with other fault diagnosis methods,ST-TCN improves the average fault identification accuracy by 45.9%.It shows that the latest framework proposed in this paper yields higher fault identification accuracy.

关 键 词：时序卷积网络 编码-解码器 软阈值化 冷水机组 传感器故障诊断 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置] TP306.3[自动化与计算机技术—控制科学与工程]