摘要
针对传统方法在滚动轴承故障诊断中无法自适应提取有效特征信息,且滚动轴承在强环境噪声干扰、复杂变工况等因素影响下诊断效果不佳,有抗噪性和泛化性下降的问题,提出了一种双路并行多尺度的改进残差神经网络(residual neural network, ResNet)的方法。该方法设计了多尺度的残差Inception模块,可以有效提取特征信息,同时加入注意力机制解决了数据的突变性和差异性,此外还使用多个空洞卷积的残差块扩大感受野,有助于提取更多特征信息,实现准确故障诊断。利用凯斯西储大学轴承数据集和东南大学变速箱数据集分别训练并测试了诊断效果,将该方法与其他卷积神经网络的方法在变噪声、变工况情况下作了对比,诊断准确率最高达到99.73%,平均准确率也在95%以上,均高于其他比较方法。结果表明,该方法在复杂多变的工况下具有较好的故障识别能力和泛化能力。
Here, aiming at problems of traditional methods being not able to adaptively extract effective feature information in rolling bearing fault diagnosis as well as poor diagnosis effect and dropping of the anti-noise ability and generalization ability under effects of strong environmental noise interference, complex variable working conditions and other factors, a dual-path parallel multi-scale improved residual neural network(ResNet) method was proposed. Using this method, a multi-scale residual inception module was designed to effectively extract feature information. Meanwhile, attention mechanism was added to solve the mutability and difference of data. In addition, residual blocks of multiple cavity convolutions were used to expand receptive field, it was helpful to extract more feature information, and realize accurate fault diagnosis. Diagnosis effects were trained and tested by using bearing data set of Case Western Reserve University, US and gearbox data set of Southeast University, respectively. The proposed method was compared with other convolutional neural network methods under variable noise condition and variable working conditions, it was shown that the proposed method’s diagnosis accuracy can reach 99.73% at the highest, and its average accuracy is also above 95%, both of them are higher than those of other methods;the proposed method can have better fault recognition ability and generalization ability under complex and variable working conditions.
作者
赵小强
张毓春
ZHAO Xiaoqiang;ZHANG Yuchun(School of Electrical Engineering and Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China;Key Laboratory of Advanced Control of Industrial Processes of Gansu Province,Lanzhou 730050,China;National Electrical and Control Engineering Laboratory Teaching Center,Lanzhou University of Technology,Lanzhou 730050,China)
出处
《振动与冲击》
EI
CSCD
北大核心
2023年第3期199-208,共10页
Journal of Vibration and Shock
基金
国家自然科学基金(61763029)
甘肃省科技计划资助(2021YF5GA072,21JR7RA206)
甘肃省教育厅产业支撑计划项目(2021CYZC-02)。
关键词
故障诊断
滚动轴承
变工况
注意力机制
多尺度ResNet
fault diagnosis
rolling bearing
variable working condition
attention mechanism
multi-scale ResNet
