Abstract: Aiming at the fact that there are few fault signals in the operation and maintenance process of gear equipment,a fault identification method combining Discrete Wavelet Transform (DWT) and two-dimensional Convolutional Neural Networks (2D-CNN) was proposed,which weighted the classification label obtained by a small number of signals through 2D-CNN and the wavelet energy of the signal for realizing the fault identification of gears.To fully obtain the information in small samples to train the neural network,DWT,image transformation and Markov transition field methods were used to increase and transform the sample signals.Through the verification of the gearbox dataset,96% training accuracy and 87.5% classification accuracy were obtained,and ablation experiments and comparison experiments were conducted,which proved that the proposed method could effectively overcome the noise interference in the small sample data,enhance the data,and have good practical significance in gear fault identification.

Key words: fault diagnosis, small samples, two-dimensional convolutional neural networks, wavelet transform

摘要： 针对齿轮设备运维过程中故障信号较少的情况,提出一种将离散小波变换(DWT)与二维卷积神经网络(2D-CNN)相结合的故障识别方法。该方法通过将少量信号经卷积神经网络得到的分类标签与信号的小波能量进行权值分配,实现对齿轮的故障识别。为了充分获取小样本中的信息来训练神经网络,利用离散小波分解、图像变换和Markov变迁场方法对样本信号进行增量和转换。通过验证齿轮箱数据集得到96%的训练准确率和87.5%的分类准确率,同时通过消融实验和对比实验证明,该方法可以有效克服小样本数据中的噪声干扰,使数据得到增强,在齿轮故障识别中具有很好的现实意义。

关键词: 故障诊断, 小样本, 二维卷积神经网络, 小波变换