基于多尺度小波包启发卷积网络的旋转机械故障诊断  

Multiscale wavelet packet-inspired convolutional network for fault diagnosis of rotating machinery

作　　者：卢一相 钱冬生[1,2] 竺德 孙冬 赵大卫 高清维[1,2] LU Yixiang;QIAN Dongsheng;ZHU De;SUN Dong;ZHAO Dawei;GAO Qingwei(MOE Key Lab of Intelligent Computing and Signal Processing,Anhui University,Hefei 230601,China;College of Electrical Engineering and Automation,Anhui University,Hefei 230601,China)

机构地区：[1]安徽大学计算智能与信号处理教育部重点实验室,合肥230601 [2]安徽大学电气工程与自动化学院,合肥230601

出　　处：《振动与冲击》2024年第17期203-213,共11页Journal of Vibration and Shock

基　　金：安徽省青年基金(2308085QF224,2208085QF206);安徽省教育厅高校自然科学重点项目(KJ2021A0013);中国博士后科学基金面上项目(2023M730009);国网安徽省电力有限公司科技项目(521203240005)。

摘　　要：在工程实践中,旋转机械故障诊断常面临噪声干扰、故障样本稀缺以及工况变化等各种复杂情况,这给先验知识缺乏的数据驱动深度学习方法应用带来了新的挑战。传统基于小波分析的故障诊断方法可提取到故障丰富的先验知识,但固定(结构化)或单一的小波基难以直接适应复杂故障场景。针对上述问题,在传统多尺度小波包分析思想启发下,提出一种基于多尺度小波包启发卷积网络(multiscale wavelet packet-inspired convolutional network, MWPICNet)的端到端旋转机械故障诊断方法。MWPICNet在神经网络内部实现了时频域转换与滤波降噪、特征提取与分类过程的有机耦合。首先,通过交替使用多尺度小波包启发卷积层和软阈值激活层进行信号分解和非线性变换,逐层挖掘多尺度时频故障特征和过滤噪声冗余信息,该过程的多次迭代可近似视为小波包阈值去噪算法在多个可学习滤波器和可学习阈值下的多层深度展开;然后,设计频带加权层动态调整各频带通道的权重;最后,引入全局功率池化层提取有助于故障状态识别的判别性频带能量特征。在三种不同应用场景下分别采用对应的机械故障数据集进行案例研究,验证了所提模型在复杂故障场景下的可行性和有效性。In engineering practice,fault diagnosis of rotating machinery often faces various complex situations such as noise interference,limited fault samples and variable working conditions,these pose new challenges to application of data-driven deep learning methods with lack of prior knowledge.Traditional fault diagnosis methods based on wavelet analysis can extract rich prior knowledge of faults,but fixed structuring or single set of wavelet bases are difficult to directly adapt to complex fault scenarios.Here,aiming at above problems,inspired by inspired by the idea of traditional multiscale wavelet packet analysis,an end-to-end multiscale wavelet packet-inspired convolutional network(MWPICNet)was proposed for fault diagnosis of rotating machinery.The proposed MWPICNet could realize the organic interactions among time-frequency domain conversion and filtering denoising,feature extraction and classification processes within a neural network architecture.First,the multiscale wavelet packet-inspired convolutional layer and soft-thresholding activation layer were alternately used for signal decomposition and nonlinear transformation,layer by layer extracting multiscale time-frequency fault features and filtering out the noise.The multiple iterations of this process could be approximated as a multi-layer deep unfolding of the wavelet packet threshold denoising algorithm under multiple learnable filters and learnable thresholds.Then,a frequency band weighting layer was designed to dynamically adjust weights of various frequency band channels.Finally,a global power pooling layer was introduced to extract discriminative frequency band energy features helpful for fault state recognition.Case studies were performed using corresponding mechanical fault datasets in 3 different application scenarios to verify the feasibility and effectiveness of the proposed MWPICNet in complex fault scenarios.

关 键 词：小波包变换 卷积神经网络 多小波基融合 故障诊断 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]