基于自编码器的调压器在线故障诊断方法

On-line Fault Diagnosis Method of Regulator Based on Auto-encoder

通过对多台在役调压器的出口压力数据进行分析,提出一种基于自编码器的燃气调压器在线故障诊断系统。该系统工作流程主要包括数据预处理、特征自动提取、出口压力稳定值判断、调压器故障诊断及早期预警。对每个工作流程进行介绍。每分钟采集一次调压器的出口压力,将预处理后的所有调压器出口压力数据分割成时间长度为24 h的样本(每组样本为1440维,即1440个数据)。将每组样本输入稀疏自编码器,得到200个稀疏特征。对出口压力稳定点预测模型,输入数据为200个稀疏特征,输出数据为1个出口压力稳定值。对调压器故障检测模型,输入数据为200个稀疏特征和1个出口压力稳定值,输出数据为1个调压器故障类型编号。设计了一种基于自编码器重构误差的用户习惯刻画方式,将自编码器的重构误差用来衡量调压器运行数据偏离历史数据的程度,重构误差越大偏离程度越大。当新的样本重构误差过高时则需要进行预警;若新样本持续偏离原有数据特征,且经排查后调压器运行正常,则说明用户习惯发生了改变,需要对历史数据库进行更新。

Based on the analysis of the outlet pressure data of multiple in-service regulators,an online fault diagnosis system for gas regulator based on auto-encoder is proposed.The workflow of the system mainly includes data preprocessing,automatic feature extraction,judgement of outlet pressure stability value,regulator fault diagnosis and early warning.Each workflow is introduced.The outlet pressure of the regulator is collected per minute,and all the outlet pressure data of the regulator after preprocessing are divided into samples with a time length of 24 h(Each set of samples is 1440 dimensions,or 1440 data).Each set of samples is inputted into the sparse auto-encoder,and200 sparse features are obtained.For the outlet pressure stability point prediction model,the input data are200 sparse features,and the output datum is 1 outlet pressure stability value.For the regulator fault detection model,the input data are 200 sparse features and 1 outlet pressure stability value,and the output datum is1 regulator fault type number.A user habit description method based on the reconstruction error of auto-encoder is designed.The reconstruction error of auto-encoder is used to measure the degree of the operating data of the regulator deviating from the historical data.The larger the reconstruction error,the greater the deviation degree.When the new sample reconstruction error is too high,an early warning is required.If the new sample continues to deviate from the original data characteristics,and the regulator operates normally after troubleshooting,which indicates that the user habits have changed,and the historical database needs to be updated.