基于小波变换和改进卷积神经网络的刚性罐道故障诊断

Fault diagnosis of rigid guide based on wavelet transform and improved convolutional neural network

  • 摘要: 现有刚性罐道故障诊断方法有的仅适用于小样本数据集,有的虽适用于大样本数据集,但忽略了实际工作环境中的多工况背景。基于卷积神经网络的刚性罐道故障诊断方法存在数据和运算量庞大,易产生过拟合等问题。针对上述问题,提出了一种基于小波变换和改进卷积神经网络的刚性罐道故障诊断方法。首先,在刚性罐道设置错位与间隙2种缺陷,采集多工况下提升容器振动加速度信号。其次,利用小波变换将采集的振动加速度信号转换为二维时频图像,采用试凑法最终确定经Complex Morlet 小波基函数处理后的二维时频图像的时间和频率分辨率最佳。然后,通过改进卷积神经网络模型结构,即保留第1层和第5层池化层,将第2,3,4层池化层替换为小尺度卷积层,以防止过拟合现象。最后,将二维时频图像输入改进后的卷积神经网络模型。实验结果表明: ① 改进模型经过训练后,在训练集上的平均准确率为99%左右,在测试集上的平均准确率为99.5%。② 当数据训练至200步后,改进模型的准确率达99%以上,改进模型的损失函数值趋近于0,说明改进模型收敛性能较好,模型的泛化能力得到了增强,在学习过程中对于过拟合的抑制效果明显。③ 在验证集混淆矩阵上,间隙缺陷和错位缺陷识别准确率为100%,无缺陷识别准确率为92%。④ 与EMD−SVD−SVM、小波包−SVM、EMD−SVD−BP神经网络、小波包−BP神经网络相比,基于小波变换和改进卷积神经网络的刚性罐道故障诊断方法准确率达99%。

     

    Abstract: Some of the existing fault diagnosis methods for rigid guide are only suitable for small sample data sets. Although some methods are suitable for large sample data sets, they ignore the multi-condition background in the actual working environment. The method of rigid guide fault diagnosis based on the convolutional neural network has the problems of huge data and computation, and easy to produce over-fitting. In order to solve these problems, a fault diagnosis method of rigid guide based on wavelet transform and improved convolutional neural network is proposed. Firstly, two kinds of defects, dislocation and gap, are set in the rigid cage guide. The vibration acceleration signals of the hoisting container under multiple working conditions are collected. Secondly, the collected vibration acceleration signals are converted into two-dimensional time-frequency images by wavelet transform. The time and frequency resolution of the two-dimensional time-frequency images processed by the Complex Morlet wavelet basis function is determined to be the best by trial and error method. Thirdly, the structure of the convolutional neural network model is adjusted. The first pooling layer and the fifth pooling layer are reserved. The second pool layer, the third pooling layer and the fourth pooling layer are replaced by small-scale convolutional layers to prevent the over-fitting phenomenon. Finally, the two-dimensional time-frequency image is input into the improved convolutional neural network model. The experimental results show the following points. ① After training, the average accuracy of the improved model is about 99% on the training set and 99.5% on the test set. ② When the training data reaches 200 steps, the accuracy of the improved model is more than 99%, and the loss function of the improved model approaches 0. These results show that the improved model has good convergence performance, and the generalization of the model is enhanced. The inhibition effect on over-fitting in the learning process is obvious. ③ On the confusion matrix of the validation set, the identification rate of gap defect and dislocation defects is 100%. The identification rate of no defect is 92%, and 8% of the defect are mistakenly identified as gap defects. ④ Compared with EMD-SVD-SVM, wavelet packet-SVM, EMD-SVD-BP neural network and wavelet packet-BP neural network, the accuracy of rigid guide fault diagnosis method based on wavelet transform and the improved convolutional neural network reaches 99%.

     

/

返回文章
返回