基于时频图与双通道卷积神经网络的轴承故障识别模型

Bearing fault identification model based on time-frequency graph and dual channel CNN

采用传统的信号处理方法难以从轴承振动信号中提取能全面准确反映轴承运行状态的故障特征,并且实际工程中采集的数据量难以满足深度学习方法的要求(需要较大数据量),针对这些问题,提出了一种基于时频图与双通道卷积神经网络(CNN)的轴承故障识别模型(方法)。首先,基于样本熵和峭度,构造了新的目标函数,利用灰狼优化算法(GWO)对变分模态分解(VMD)方法进行了参数优化,当目标函数达到最小值时,得到了其最优参数组合;然后,使用经过参数优化后的变分模态分解(VMD)方法对轴承信号进行了处理,将处理后得到的模态分量进行了平滑伪Wigner Ville分布(SPWVD)计算,累加其计算结果后,最终得到了轴承的时频图;其次,利用连续小波变换(CWT)直接对原始信号处理得到了时频图;最后,将采用两种方式得到的时频图分别作为双通道CNN的输入,对网络进行了训练,由CNN提取了其时频图特征,并对轴承故障进行了识别分类和诊断。实验结果表明:采用该方法在轴承故障实验中得到的准确率为99.69%,在10次实验中的平均准确率达到了99.61%,相比于单通道CNN和支持向量机(SVM)等方法,该方法有着更高的准确率和更出色的稳定性。研究结果表明:将该方法应用在轴承故障诊断领域,具有准确率高、稳定性强的特点,能够有效地诊断轴承故障。

Aiming at the problem that it was difficult for traditional signal processing methods to extract fault features that could fully and accurately reflect the running state of bearings from bearing vibration signals,and the amount of data collected in practical engineering was difficult to meet the requirement of large data amount in deep learning,a bearing fault identification model based on time-frequency graph and dual-channel convolutional neural network(CNN)was proposed.Firstly,a new objective function was constructed based on sample entropy and kurtosis,the grey wolf optimization algorithm(GWO)was used to optimize the parameters of variational mode decomposition(VMD);when the objective function reached the minimum value,the optimal parameter combination was obtained.Secondly,the optimized VMD was used to process the bearing signal,and the obtained modal components were calculated by smoothing pseudo-Wigner Ville distribution(SPWVD),and the time-frequency graph of the bearing was obtained by summing the calculated results.Thirdly,the continuous wavelet transform(CWT)was used to process the original signal directly to obtain the time-frequency graph.Finally,the time-frequency graphs obtained by the two methods were respectively used as the input of dual-channel CNN to train the network,and the features of the time-frequency graphs were extracted by CNN and faults were classified.The experimental results show that the accuracy rate of the proposed method is 99.29%in the bearing fault experiments,and the average accuracy rate is 99.61%in ten experiments.Comparing with the single-channel CNN and support vector machine(SVM),the proposed method has higher accuracy and better stability.The results show that this method has the characteristics of high accuracy and strong stability in the field of bearing fault diagnosis,and can effectively diagnose bearing faults.