Fault Diagnosis Method Of Planetary Gearbox Based On DCGAN Under Unbalanced Samples

The planetary gearbox is a crucial component of rotating machinery,which is often subjected to harsh operating conditions,leading to potential part failure.This can significantly impact the operational efficiency and economic benefits of the equipment.Therefore,researching fault diagnosis theory and methods for the planetary gearbox is of great importance to ensure its security and reliability.In recent times,the utilization of deep learning methods based on big data has led to impressive accomplishments and provided a significant approach for research in fault diagnosis.Nonetheless,the majority of diagnostic techniques reliant on big data require ample training data,and their effectiveness can be easily influenced by the number of samples available.In practice,due to various limitations,it is difficult to collect fault information on planetary gearboxes.It takes a lot of human and material resources to mark the fault information,resulting in insufficient fault information.In other words,the number of erroneous data is much smaller than that of normal data,leading to an uneven distribution of data.This thesis aims to improve the precision of fault diagnosis of planetary gearboxes and the reliability of their diagnostic results,especially when the sample is not balanced.The research specifically examines the planetary gear within the gearbox.The study proposes a fault signal data expansion technique and a diagnosis method that utilizes a generated adversarial network and a convolutional neural network.The key objective is to enhance fault diagnosis accuracy and improve the reliability of diagnosis outcomes.(1)In this thesis,presenting a new model of sample augmentation based on Deep Convolutional Generative Adversarial Networks(DCGAN).The aim is to generate new samples with the same data distribution as the original fault sample to solve the problem of sample imbalance.To capture the periodic signal from the defective gear,using a vibration sensor,which is then subjected to denoising using the wavelet threshold method.The 1-D vibration signal is converted into a 2-D Gram-Angle field encoding scheme using the Gram-Angle field encoding technology.This 2D scheme is then used to produce a sample for DCGAN learning and training,and the sample is input into the DCGAN network.After each training,the network parameters are adjusted,and the DCGAN is used to generate new samples.(2)In this thesis,a fault diagnosis model for a planetary gearbox based on the Alex Net network is proposed.The expanded sample set is input into the Alex Net network for training,and the network parameters are adjusted after each training session.Finally,the best Alex Net fault diagnosis model is obtained.According to the results of fault classification and recognition,the recognition accuracy of the four kinds of gearbox faults is up to 99.2%.In the laboratory,unbalanced samples are collected and processed before being input into the network for classification.The classification results are then compared with the results obtained from the previous network.The findings indicate that using the expanded DCGAN samples can effectively improve the classification accuracy of the neural network.The laboratory bearing data were collected from laboratory and used to verify that the proposed GAF-DCGAN+CNN network model has good generalization.(3)This thesis proposes a planetary gearbox fault diagnosis model based on migration learning and Efficient Net.In order to increase the precision of the neural network,some researchers have simply modified the width,depth,or image resolution of the convolutional neural networks,resulting in a larger network model,more parameters,and slower training speeds.To address this issue,this study first conducts pre-training on the Image Net dataset and then transfers the parameter weights to the Efficient Net network using model-based transfer learning.By loading the weights of the MBConv modules in the original network,the training time is decreased,and the efficiency of network training is improved.The expanded balanced samples are then input into the Efficient Net＿B2 network for training,allowing the network to extract and learn the detailed and depth features of the images.Upon completing the training process,the network can identify and classify the four kinds of faults in the sample with the highest accuracy of 99.49%.Compared with the previously mentioned Alex Net classification network,the accuracy is higher and the training speed is faster.Furthermore,the laboratory bearing data and the planetary gearbox fault dataset of Southeast University are used to verify that the network has good generalization.