Research On Fault Diagnosis Method Of Machine Tool Rotary Parts Based On Time-frequency Dual-source Data

Rotary moving parts are the main components in CNC machine tools,and the quality of their working conditions can directly affect the quality of the entire machine tool production and processing,the service life of the machine tool and the probability of production accidents.Mechanical fault diagnosis technology has evolved from the traditional diagnosis method that relies too much on expert knowledge to the modern diagnosis method of artificial intelligence based on data,which greatly improves the accuracy of the fault diagnosis process and the robustness and universality of the diagnosis model.Introduced a new problem that the model generalization performance is too dependent on the data volume.Based on the scene with less data,the paper analyzes the good and bad situation of the different feature spaces of the signal to the representation of the fault information,and proposes a new data-driven artificial intelligence fault diagnosis method that combines the signal time-frequency signal space information.By characterizing the signal in the image in the time domain and frequency domain respectively,two time-frequency feature spaces of the signal are constructed.In the time domain,the time-amplitude intensity axis is established to draw the waveform image of the signal in the time domain;in the frequency domain,the frequency-amplitude intensity axis is established to draw the frequency spectrum of the signal in the frequency domain.The signal-image time-frequency feature space representation method can effectively represent information and compress information,and can also artificially generate more new data through data enhancement technology.Based on the constructed time-frequency dual-source data,three dual-source data convolutional neural network schemes are proposed,and three dual-source data input models the CNN of being applied image stitching,Concat-2CNN and Gate-2CNN are constructed.The trainability of the model also uses a weight sharing's regularization strategy and parameter initialization strategy in the model.The introduced strategy can accelerate the convergence of the model.In a scenario with less data,based on the well-trained source fault diagnosis model,a specific transfer learning scheme for a dual source data convolutional neural network is designed,and two fixed and fine-tuning styles‘s model parameter training algorithm introduced for this scheme.And in the transfer learning function verification experiment,the fine-tuning scheme can improve the accuracy of about 13%Finally,based on the bearing failure data provided in the CWRU bearing database,it is verified that the proposed model has a high stability in the convergence process and a strong generalization performance on a small data set;based on the Hust CNC Center bearing failure The bearing failure data collected by the experimental platform verifies that the introduced transfer learning method can effectively improve the fault diagnosis performance.