| In the last decade or so,deep learning has entered a golden period of rapid development.Research based on deep learning algorithms in various fields has achieved abundant results.Deep learning is widely used in image classification,target detection,scene recognition,natural language processing,intelligent driving,and other fields because of its powerful feature representation capability and good adaptability.With the development of machine learning technologies such as deep learning,methods in the field of fault diagnosis are increasingly moving toward intelligence.Feature extraction is a key part of the fault diagnosis process,and the ability to characterize faulty features directly affects the diagnosis results.Usually,the feature extraction of the conventional fault diagnosis methods which are based on signal processing technology or intelligent fault di-agnosis technology using shallow machine learning method mainly relies on manual work.Once lacking professional signal processing technology or sufficient prior knowledge,it is impossible to extract useful features to ensure the fault diagnosis’s effect.Based on the theory of deep learning,this thesis addresses the above-mentioned problems in the fault diagnosis of vehicle key components,aiming to automatically extract deep features from the raw vibration signal data of key vehicle components to achieve end-to-end fault diagnosis.The main research of this thesis is as follows.Firstly,for the problems such as hyper-parameter selection of convolutional neural net-works,the structure and hyper-parameters of convolutional neural networks are studied.An improved convolutional neural network model is proposed for the fault diagnosis of vehicle key components.The model directly uses the original vibration signal as its input,eliminat-ing the reliance on manual work for feature extraction and feature selection.The improved convolutional neural network model uses leaky modified linear units and parametric modi-fied linear units as activation functions,which have the advantages of the standard modified linear unit activation functions and can overcome its disadvantage at the same time.The effectiveness of the method is verified by experimental data of bearings and gears,and the results show that the improved convolutional neural network has higher diagnosis accuracy,better convergence speed.Secondly,optimization on the top classifier layer of the intelligent diagnosis model is studied to enhance the generalization ability.Combining the advantages of deep learning and classical machine learning,a hybrid intelligent diagnosis model based on convolutional neural networks and support vector regression is proposed.The advantages of both methods are combined to extract the deep features directly from the original signal data for fault diagnosis.A learning rate adjustment strategy is used to adjust the step size of the gradient update,which in turn optimizes the convergence process of the model.The proposed method is validated with rolling bearing data and gearbox data,and since it can automatically mine fault features from the raw time-domain signals,the method does not rely on traditional feature engineering and does not require the prior knowledge of specialized signal processing and diagnosis experts.The diagnosis results show that the proposed intelligent diagnosis model can successfully distinguish the health conditions of key vehicle components and can achieve intelligent fault diagnosis of key vehicle components more efficiently.Thirdly,a generalized convolutional neural network for small samples is proposed for the problem of an unbalanced distribution of fault data and health data.The adaptive moment estimation algorithm is introduced during the training process to accelerate model training and improve the model convergence speed.The proposed network benefits from the au-tomatic feature learning capability of convolutional neural networks and the generalization capability due to jump connections.The method is validated by balanced data experiments and unbalanced data experiments.More importantly,since unbalanced data classification is an inherently difficult task,in this part of the research,six cases with different degrees of imbalance are designed to validate the proposed model.The experimental results show that the method has better learning results than ordinary deep learning methods.The method has good feature representation capability for both a fewcategory fault diagnosis and multi-category class fault diagnosis,and can effectively achieve the diagnosis of bearing faults either using balanced data or different degrees of imbalanced data.Finally,mainly the work of the thesis and the innovation points are summarized,and several prospects for the subsequent research are proposed. |
