串行集成CNN模型的两种机械故障诊断算法

Two Mechanical Fault Diagnosis Algorithms Based on Serial Integrated Convolutional Neural Network Model

表征机械故障的数据种类和数量多样,数据质量参差不齐,故障信息的价值密度也较低,对机械故障诊断提出了挑战。由于浅层神经网络模型自学能力较弱,无法达到精确诊断故障的要求,因此对故障特征提取和故障诊断模型进行了深入研究。利用时间序列排序转换和连续小波变换,分别构建了基于时间序列图像和时频图的两种2D-CNN故障特征数据集,结合卷积神经网络(convolutional neural network,CNN)和梯度提升决策树(gradient boosting decision tree,GBDT)算法的优点,提出了基于GBDT的CNN-Z和CNN-F两种故障诊断算法。通过实验,两种算法的故障诊断精度分别为94.33%和98.62%。与传统CNN算法相比较,实验结果展示了所提的两种故障诊断算法的精度更高,另外基于GBDT的CNN-F故障诊断算法的误差收敛时间为前者的三分之一,验证了所提算法的有效性和精确性。

The various types and huge amount of data that characterize mechanical failures,as well as the uneven quality of the data and the low value density of failure information,pose a big challenge to mechanical failure diagnosis.However,the shallow neural network model can not meet the requirements of accurate fault diagnosis due to its week self-learning ability.Therefore,the research group conducted indepth research on fault feature extraction and fault diagnosis models.With time series permutation transform and continuous wavelet transform,two 2D-CNN fault feature dada sets were constructed based on images of time series permutation and time-frequency images respectively.Combined with the advantages of convolutional neural network(CNN)and gradient boosting decision tree(GBDT)algorithm,two fault diagnosis algorithms of CNN-Z and CNN-F based on GBDT were proposed.Through experiments,the two algorithms have achieved higher fault diagnosis accuracy,as high as 94.33%and 98.62%respectively.Compared with the traditional CNN algorithm,the experimental results show that the two fault diagnosis algorithms proposed in the paper have higher accuracy.In addition,the error convergence time of the fault diagnosis algorithms of CNN-F based on GBDT is only one third of the former,which verifies the effectiveness and accuracy of the algorithm proposed in this paper.