一维卷积神经网络在往复式压缩机气阀故障诊断中的应用

Application of One-Dimensional Convolutional Neural Network in Fault Diagnosis of Reciprocating Compressor Air Valve

针对往复式压缩机气阀故障诊断问题,对气阀盖上的振动信号进行分析,提出了一种基于一维卷积神经网络(1D-CNN)的故障诊断模型。首先,将原始一维振动信号经傅里叶变换从时域转换为频域;然后,将频域信号作为1D-CNN的输入,利用卷积层实现自适应提取特征;最后,网络输出层利用Softmax函数实现多种故障的模式识别。在往复式压缩机故障模拟实验台上进行了气阀正常、阀片裂纹、阀片断裂、弹簧失效4种工作状况下气阀盖振动信号的测量并对提出的模型进行验证。结果表明,气阀盖上的振动信号能够明显反映气阀的工作状态,而且信号易提取、十分适合用于气阀的故障诊断;将振动信号从时域转换成频域作为1D-CNN的输入明显地提高了故障分类的准确率;与采用原始一维振动信号作为1D-CNN输入的模型相比,采用频域信号作为输入的故障诊断模型具有更优越的表现,准确率更高,可达100%,而且模型结构简单,能够实现端到端的快速故障诊断。

Aiming at the fault diagnosis of reciprocating compressor air valves,a fault diagnosis model based on one-dimensional convolutional neural network(1D-CNN)is proposed for analyzing vibration signals on the air valve cover.First,the original one-dimensional vibration signal is converted from the time domain to the frequency domain via Fourier transformation;then,the frequency domain signal is used as the input of 1D-CNN to realize adaptive feature extraction by the convolutional layer;finally,the network output layer could recognize modes of multiple faults by Softmax function.At the reciprocating compressor failure simulation test bench,we have measured the valve cover vibration signal under four working conditions,namely normal air valve,valve plate crack,valve plate fracture and spring failure,and verified the proposed models.The result shows that the vibration signal on the air valve cover could obviously reflect the working state of the air valve,and the signal is easy to extract and very suitable for fault diagnosis of the air valve;vibration signals are converted from time domain to frequency domain as input of 1D-CNN,which significantly improves the accuracy of fault classification;compared with the model with original one-dimensional vibration signal as input,the fault diagnosis model with frequency domain signal as input has better performance and higher accuracy(up to 100%).Moreover,it is simple in structure,and could realize end-to-end fast fault diagnosis.