Research On Fault Diagnosis Methods Of Rotating Machinery Based On Deep Learning And Data Fusion

Rotating machinery and equipment are widely used in power plant operation.Modern rotating machinery structure is complex,the operating environment is more severe,prone to failure.Once the failure of rotating machinery and equipment,it may affect the economic benefits of power enterprises,and in serious cases even endanger the lives of site staff.Therefore,the study of rotating machinery fault diagnosis methods,timely diagnosis of rotating machinery to ensure its normal work is of great importance.If a failure occurs during the work of rotating machinery,some signals are bound to change.Collecting some data signals such as vibration signal,motor current signal,acoustic signal and so on during the operation of rotating machinery can carry out fault diagnosis of rotating machinery to a certain extent.In this dissertation,in view of the limitations of the fault diagnosis of rotating machinery based on single signal,the fault diagnosis method of rotating machinery based on deep learning and data fusion technology was studied,the main tasks of this investigation were the following.In this dissertation,the common fault categories of rotating machinery and the fault characteristics were analyzed firtst,Short Time Fourier Transform(STFT)was introduced to address the limitations of signal time domain analysis or frequency domain analysis,the processed signal samples were converted into time-frequency maps for deep learning model recognition,and the capability of deep learning methods for image recognition was fully exploited.A rotating machinery fault diagnosis process based on signal monitoring,shorttime Fourier transform,and multi-signal fusion neural network recognition and classification was designed.The multi-channel fusion neural network models based on convolutional neural network and lightweight Mobile-VIT neural network for mobile were built respectively,and the effectiveness of the proposed method was verified by using the publicly available bearing monitoring datasets from Case Western Reserve University and University of Patten,Germany.The effects of hyperparameters on the training process and results of the network were investigated,and the accuracy of the proposed model could reach up to 98.01%and 95.75%with the addition of-4 dB signal-to-noise ratio noise.Finally,the t-sne algorithm was used to downscale the network output and display it in 3D space and drew the confusion matrix to show the classification results.To verify the noise immunity of the proposed model,the model was fine-tuned by adding Gaussian white noise with different signal-to-noise ratios to the vibration signal.The comparison study found that the accuracy of the fault diagnosis model of the comparison object decreased significantly when different S/N ratios of noise were added,while the accuracy of the two network structures proposed could still reach 81.75%and 89.38%under the strong noise condition of-10 dB S/N ratio,indicating that the model proposed had certain noise immunity.A summary of the full paper and the future research directions were presented at the end of the paper.