摘要
针对传感器故障检测方法对早期微弱故障信息不敏感以及抗异常值干扰能力差的问题,提出了一种基于因果卷积改进的自注意力长短期记忆网络(CCALSTM)模型和Shapiro-Wilk检验与阈值比较法相结合的故障预检测方法。首先在长短期记忆网络(LSTM)模型中引入基于因果卷积的自注意力机制,以提取局部信息特征,减少异常值对预测精度的影响;然后将预测结果与测量值进行残差计算,并利用滑动窗口选取合适长度的残差序列;最后将残差序列通过Shapiro-Wilk检验和阈值比较法相结合的故障检测方法进行故障预检测。通过传感器原始数据进行仿真实验,并与支持向量机(SVM)、误差反向传播网络(BP)、双向长短期记忆网络(Bi-LSTM)等常见预测模型进行对比,结果表明CCALSTM模型取得了更高的预测精度结果,且具有更高的鲁棒性;同时,所提出的故障预检测方法表现出对传感器早期微弱故障敏感,能够在故障潜伏期及时检测出故障。
Aiming at the problem that the sensor fault detection method is insensitive to early weak fault and is susceptible to outliers,a model based on improved self-attention with Long Short-Term Memory(LSTM)and a sensor fault pre-detection method combing Shapiro-Wilk test and threshold comparison method were proposed.First,a Causal Convolution-based self-Attention mechanism(CCALSTM)was introduced into the LSTM model to extract local information features and reduced the impact of outliers on prediction accuracy.Then the residuals of the predicted results and measured values were calculated,and the residuals with appropriate length were selected by using the sliding window.Last,the residual sequence was input into the fault detection method combining Shapiro-Wilk test and threshold comparison method for fault pre-detection.The experiments were carried out on the original sensor data and CCALSTM was compared with SVM(Support Vector Machine),BP(Back Propagation)neural network and Bi-LSTM(Bidirectional Long Short-Term Memory).Experimental results show that the accuracy of the proposed model is higher than those of other models,the proposed model is more robust,and the proposed method of fault pre-detection is sensitive to the early weak fault of the sensor and can detect the fault in time during the fault incubation period.
作者
林涛
付崇阁
吉萌萌
LIN Tao;FU Chongge;JI Mengmeng(College of Artificial Intelligence and Data Science,Hebei University of Technology,Tianjin 300401,China)
出处
《计算机应用》
CSCD
北大核心
2021年第S01期31-35,共5页
journal of Computer Applications
基金
国家自然科学基金资助项目(61976242)。
