Fault Diagnosis Of Ship Air Compressor Based On HHT And Deep Learning

Ship reciprocating two-stage air compressor is one of the key power equipment of a ship.According to the requirements of “made in China 2025 “ in the era of big data,fault diagnosis of ship air compressor based on one-dimensional vibration signal in time domain has the practical significance.HHT algorithm has the advantage of processing nonlinear and non-stationary signals,the CNN network has the ability to extract spatial features,and the improved RNN network LSTM has the characteristics of long-term and short-term memories of the time series signals.In this paper,four kinds of deep learning neural network models are proposed to diagnose the faults of air compressor for Vibration signals measured under laboratory conditions with added noise.In this paper,the data set enhancement technology,HHT algorithms and the processing of vibration signal with added noise are used to obtain the eigenmode components with large correlation coefficient,and then the data sets of five states are made.A 3-layer CNN network is constructed to verify the effectiveness of the CNN network in fault diagnosis of ship air compressor.Batch normalization(BN layer)and Dropout technology are used to optimize the CNN network,which effectively reduces the overfitting and improves the generalization ability and recognition rate.However,the recognition rate is low on data sets with large SNR.LSTM has the ability to remember the long-term and short-term characteristics of time series signals.An LSTM network is proposed for one-dimensional time-domain signal fault diagnosis of ship reciprocating two-stage air compressor,and the result is 97% recognition,which proves the effectiveness of LSTM in fault diagnosis of ship air compressor.An improved EMD algorithm is proposed for mixing vibration signals with different SNR to simulate the vibration signals of ship air compressor under real conditions.For such signals,a CNN network combined with LSTM network(RCNN)is proposed.The mixed matrix was used to view the specific test results,and t-SNE visible technology was used to observe the aggregation and separation of input characteristics in the neural network every layer to evaluate the performance of the model.