Research On Fault Diagnosis Method Of Reciprocating Compressor Valve Based On SE-MSCNN And GRU Fusion

Reciprocating compressor is a kind of general equipment widely used in petroleum and petrochemical,metallurgy,and other industrial fields.In the production of heavy production tasks,its working medium often has flammable and explosive,high-temperature,and high-pressure characteristics.Once the failure,it will inevitably cause economic losses of production stoppage,and even lead to major accidents of machine destruction.Valve is a reciprocating compressor that works key components and high-frequency wear parts.Its working condition directly affects the operating efficiency and reliability of the reciprocating compressor.Therefore,it is important to carry out research on the fault diagnosis method of reciprocating compressor valves to ensure the continuous and safe operation of the equipment.Because of the complex structure of reciprocating compressors and many excitation sources,the vibration signals show strong nonlinear,non-stationary,and non-Gaussian complex characteristics.Although fault diagnosis methods based on machine learning are capable of adaptive classification and identification of the extracted features,they still need to rely on experts’ empirical knowledge,and it is difficult for these methods to make the adaptive diagnosis for different operating conditions,fault types,and structural parameters.In particular,when the massive vibration data is collected in the field,the normal samples are much larger than the fault samples,and the diagnosis efficiency and accuracy of intelligent algorithms do not reach the industrial application conditions.To solve the above-mentioned difficulties,this paper proposes a deep learning algorithm based on the fusion of SE-MSCNN and GRU,which integrates two key parts,feature extraction and pattern recognition.It can extract potential fault features from a large amount of data and get rid of the dependence on expert empirical knowledge and manually extracted features to achieve intelligent diagnosis with adaptive feature extraction.The main research contents and results of this paper are as follows:(1)To address the problems of traditional convolutional neural networks with a single scale of convolutional kernel and redundant network layer structure features,the SE-MSCNN fault diagnosis model is proposed.It is based on convolutional neural networks and makes an improvement.The multi-scale features extracted by the added multi-scale convolutional kernel not only consider the global fundamental features of the signal but also extract local features.The multiscale features extracted by MSCNN contain a small amount of redundant information or irrelevant information.Adding SE-Net to MSCNN to recalibrate the multiscale features can squeeze the irrelevant information and excite the important information to enhance the characterization ability of the model.The results show that the method achieves a more than 99% recognition rate on the simulation and bearing datasets.(2)To address the problem that the frequency distribution and structure of reciprocating compressor vibration signals change continuously with time.Especially with the deepening of the fault degree,its frequency complex characteristics also change,which eventually leads to the problem that the features extracted by the model are not regular.GRU fault diagnosis model is proposed.The method is good at extracting temporal features,achieving a recognition rate of 100% on the simulation data set,and 93.78% on the bearing experimental dataset.The network performance is significantly better than RNN and LSTM,and solves the problem of gradient disappearance.(3)To address the problem that the SE-MSCNN model cannot extract temporal features,a fusion fault diagnosis model of SE-MSCNN and GRU is proposed,which makes up for the deficiency that SE-MSCNN cannot extract temporal features and enhances the robustness and feature extraction ability of the model.The results show that the fusion model achieves an accuracy of 99.20% in the simulation dataset and a diagnosis rate of 98.98% in the bearing dataset with variable operating conditions experiments.(4)To further validate the performance of the SE-MSCNN and GRU fusion methods in the reciprocating compressor experimental data and the migration dataset,the SE-MSCNN and GRU fusion models were applied to the reciprocating compressor valve fault diagnosis.The results show that the recognition rate reaches up to 100% on the reciprocating compressor valve dataset,and improves by 2.39%,0.56%,and 0.07% compared with the GRU model,MSCNN model,and MSCNN-GRU model.The best performance is achieved in datasets of different sizes,with an average accuracy of 97.60%.In the migration experiments,the fine-tuned SE-MSCNN and GRU models have a 95.80% recognition rate in new fault datasets.