Intelligent Fault Diagnosis of Reciprocating Compressor Based on Attention Mechanism Assisted Convolutional Neural Network Via Vibration Signal Rearrangement

Abstract

Reciprocating compressor is extensively used in petrochemical industry and other fields. However, due to the nonlinearity of the system, it is usually difficult for traditional methods to extract reliable fault features from its vibration signal and achieve satisfactory diagnostic accuracy under the condition of high intensity noise. In this paper, a novel fault recognition method for reciprocating compressor is proposed on the basis of signal rearrangement and attention mechanism assisted convolutional neural network. Firstly, to enhance the features of the raw signal without information loss and avoid artificial feature extraction, a novel signal rearrangement method, that can convert the raw data into 2-D format, is developed. The proposed signal rearrangement method can bring the data points into a straight line (45 degrees counterclockwise from the horizontal), which can reinforce the characteristics of the raw data and make it more intuitive. Besides, to enable the network to make adequate use of the characteristics of different channels and take global feature into consideration, the attention mechanism is introduced into the convolutional neural network classifier through the SE module of the SENet. The effectiveness of the proposed method is verified by experiments, and the experimental results show that, the diagnostic accuracy of the proposed method reaches 99.4%. In addition, even under strong noise, the method of this work can still maintain an accuracy of 90.2%. Compared with other typical methods, the method suggested in this work not only holds a higher recognition accuracy, but also a stronger ability of anti-noise.