Research on Fault Diagnosis of Reciprocating Compressor Based on GAF and Inception-ResNet-v2

The fault signal of the reciprocating compressor has many excitation sources and nonlinear. Deep learning can solve such problems. However, with the deepening of the network, the amount of model parameters is increased, and the phenomenon of gradient disappearance or explosion may occur. In addition, one-dimensional signals are used for diagnosis, which is easy to lose time dependence, resulting in unstable accuracy. In view of the fact that the two-dimensional images generated by the Gramian Angular Field contain the temporal correlation of one-dimensional sequences, and the Inception-ResNet-v2 can reduce parameters and avoid the phenomenon of gradient disappearance or explosion, this paper proposes a fault diagnosis method for reciprocating compressor based on GAF and Inception-ResNet-v2 to solve the above problems. First, the GAF method is used to convert the one-dimensional vibration signal into two-dimensional images to preserve the time correlation of the one-dimensional vibration data. Then, the generated two-dimensional images are input into Inception-ResNet-v2 for feature extraction, and the Adam optimizer is used for optimization to improve the convergence speed of the model. Finally, the extracted features are input into the softmax classifier to obtain the classification results. The experimental results show that the method can effectively identify the fault types of the reciprocating compressor, with an accuracy rate of 99.7%. In order to prove the superiority of this method, the methods of PIL+Inception-ResNet-v2, GAF+Inception-ResNet-v2, PIL+ResNet50 and GAF+ResNet50 are selected for comparison. The GAF+Inception-ResNet-v2 method has the highest accuracy rate, and the PIL+ResNet50 method has the lowest accuracy rate of 91.25%. Inception-ResNet-v2 converges faster than ResNet50. The results show that the images generated by GAF can retain more feature information, and Inception-ResNet-v2 improves the accuracy of network classification while reducing the amount of parameters and computation. GAF and Inception-ResNet-v2 can improve the overall model's effect in reciprocating compressor fault diagnosis.