Research On Gearbox Fault Diagnosis Based On Convolutional Neural Network And Gated Recurrent Unit

Gearbox is a common mechanical equipment used to transmit power,has been widely used in agricultural production,wind power generation and other fields,once the gearbox failure,will directly affect the industrial production and daily life,and in serious cases even cause personal injury and economic loss.Therefore,the gearbox and its key components to carry out fault diagnosis,timely detection and troubleshooting to ensure the healthy operation of machinery and equipment and reduce equipment maintenance costs play a key role.In this paper,we take gears,a key component in gearboxes,as the research object,build a gearbox fault experiment platform,collect vibration signals,and carry out intelligent diagnosis research on gearbox faults by using machine learning and deep learning methods.The main contents of the paper include the following aspects:1.First,a brief analysis of the common failure forms in gearboxes and the reasons for their generation is carried out,a gearbox failure experimental platform is built by selecting suitable experimental equipment according to the existing experimental conditions,setting up the gear failure forms and conducting experiments to collect the vibration data required for subsequent experimental verification and analysis.2.For the characteristics of nonlinearity of gearbox vibration signal and the problem of interference of redundant components,a feature extraction method combining adaptive variational mode decomposition(AVMD)and refined composite multiscale dispersion entropy(RCMDE)is proposed.First,the AVMD decomposition is used to denoise the obtained vibration signal,and for the problem that the VMD needs to artificially set the decomposition layer K during the decomposition,a comprehensive evaluation index is proposed for the adaptive selection of the decomposition layer K value,and the components containing the main information are selected for reconstruction according to the double threshold of IMF energy entropy and correlation coefficient to complete the denoising pre-processing of the vibration signal;second,the RCMDE is used to extract Finally,the intelligent diagnosis of gearbox faults is completed by the kernel extreme learning machine(KELM)which automatically obtains the kernel parameters and penalty coefficients by the particle swarm optimisation algorithm(PSO),and the effectiveness of the method is proved by experimental analysis.3.To address the problem that traditional machine learning methods require human extraction of features and the recognition results are easily affected by human factors and expert experience,based on the AVMD decomposition method for vibration signal denoising,a method combining AVMD with convolutional neural network(CNN)-gated recurrent unit(GRU)is proposed to directly input the denoised vibration signal into the CNN-GRU model,using CNN is used to mine the spatial information of vibration signal,and GRU is used to further mine the temporal information to obtain the spatio-temporal features that can characterize the spatial and temporal information of vibration signal,and finally the effective recognition of different states of gearboxes is completed by softmax layer,and the average accuracy of 10 tests is 98.08%,which can effectively complete the recognition of different fault forms of gearboxes,and the recognition The results are more stable.4.To address the problem that the vibration signals collected by a single sensor are easily disturbed by noise and cannot effectively characterize the operating status of gearboxes,and to obtain a more effective and stable fault diagnosis method for gearboxes,a parallel convolutional neural network(PCNN)-GRU fusion multi-sensor information fault diagnosis model is proposed based on the CNN-GRU model,which is used to fuse the multi-sensor acquired The results show that the average accuracy of 10 tests without denoising is 99.92%.The results show that the average accuracy of 10 tests is 99.92% without denoising,and it has higher recognition accuracy and stability and lower loss rate compared with other methods.