基于改进卷积神经网络的齿轮箱故障诊断方法

Fault diagnosis method of gearbox based on improved convolutional neural network

针对传统卷积神经网络(CNN)对振动信号中的时序信息特征提取不敏感,导致对齿轮箱故障诊断准确率不高的问题,提出基于CNN与双向长短时记忆网络(BiLSTM)结合的改进模型用于齿轮箱故障诊断,利用BiLSTM神经网络能够根据时序信息建模的自适应特征提取能力,使改进网络同时具有空间和时间上的特征提取能力,对振动信号中故障特征进一步深度挖掘,同时采用支持向量机(SVM)代替Softmax分类器,克服了Softmax计算速度较慢,受噪声干扰较大的缺点。通过公开数据集验证,改进CNN网络模型故障诊断准确率为98.7%,损失为0.045,比单独的CNN模型准确率高18.8%,比Softmax分类的CNN-BiLSTM模型准确率高3.75%,实验结果证明了该方法比传统CNN网络诊断准确率更高、更有效。

To address the problem that the traditional convolutional neural network(CNN)is insensitive to feature extraction of temporal information in vibration signals,resulting in low accuracy in gearbox fault diagnosis,an improved model based on the combination of CNN and bi-directional long and short term memory network(BiLSTM)is proposed for gearbox fault diagnosis,using the adaptive feature extraction capability of BiLSTM neural network capable of modelling according to temporal information,so that The improved network has both spatial and temporal feature extraction capabilities,further deep mining of fault features in vibration signals,while a support vector machine(SVM)is used instead of the Softmax classifier,overcoming the disadvantages of the slower computation speed of Softmax and greater interference by noise.The improved CNN network model fault diagnosis accuracy is 98.7%with a loss of 0.045,which is 18.8%more accurate than the CNN model alone and 3.75%more accurate than the CNN-BiLSTM model with Softmax classification,as verified by the public dataset.The experimental results demonstrate that the method is more accurate and effective than the traditional CNN network diagnosis.