Abstract: Deep learning methods are widely concerned in the field of mechanical equipment intelligent fault diagnosis.To extract and identify fault features from multi-sensor raw fault monitoring data more effectively and solve the limitation of a single algorithm in fault diagnosis,a new deep recurrent convolution neural network algorithm based on improved Long Short Term Memory neural network—Global average pooling Convolutional Neural Network (LSTM-GCNN) was proposed for fault intelligent diagnosis of mechanical equipment.A Long Short Term Memory (LSTM) neural network was designed to extract time-related memory features from multi-sensor raw data.Then,these feature data was input into a one Dimension Convolutional Neural Network (1D-CNN) for micro difference feature identification.Furthermore,to reduce the number of model parameters and improve the detection speed of the algorithm,a one Dimension global average pooling layer was designed to improve the fully connected layer structure of traditional CNN.Finally,the proposed method was applied to the three-channel vibration signal data of rolling bearing under various load conditions including in 1 horsepower,2 horsepower and 3 horsepower,and the diagnostic accuracy was 100%,99.85% and 99.78% respectively.Compared with the traditional DNN,LSTM and CNN algorithm,the experimental results showed that the proposed method had better diagnostic performance.The diagnosis accuracy rate of 8-channel raw fault data of the gearbox under no-load and load-bearing conditions was up to 99.93% and 99.8% respectively.The proposed method had good universal migration performance.

Key words: intelligent fault diagnosis, deep learning, recurrent neural network, convolutional neural network, multi-sensor

摘要： 深度学习作为一种智能高效的模式识别技术,已得到基于大数据驱动的机械装备故障诊断领域学者的广泛关注。为了更加有效地从多传感器原始故障数据中提取出故障特征,解决单一诊断算法提取时序数据特征时的信息丢失问题,提出一种基于改进的长短期记忆循环神经网络全局均值池化卷积神经网络(LSTM-GCNN)的深度循环卷积神经网络新算法,用于机械装备大数据的故障智能诊断。该算法首先运用长短时记忆循环神经网络(LSTM)从多通道原始数据中提取时间关联性记忆特征,然后再将特征数据输入到一维卷积神经网络(1D-CNN)中进行微小差异特征辨识,并且为了减少模型参数量和提高算法检测速度,设计了一个一维全局均值池化层用于代替传统1D-CNN算法中的全连接层结构。通过将提出的算法用于滚动轴承在1马力、2马力和3马力多种负载工况下采集的3通道振动信号数据进行诊断验证,分别得到100%、99.85%和99.78%的诊断准确率,实验结果相比传统的DNN、LSTM和CNN算法具有更加优越的诊断性能;对齿轮箱在空载和承载两种运行工况下的8通道原始数据进行故障诊断的准确率分别高达99.93%和99.8%,具有良好的迁移通用性能。

关键词: 智能故障诊断, 深度学习, 循环神经网络, 卷积神经网络, 多传感器