Classification And Recognition Of Drivetrain Data Based On Convolutional Neural Network

The drivetrain has been widely used in the new energy vehicles.Whether it can efficiently and reliably provide power for new energy vehicles is a long-term focus at domestic and overseas.The online detection of the drivetrain is the last defense to ensure the quality at the end of the production.Therefore,the focus of drivetrain online detection,in the future,is to find an effective and high-precision fault recognition and classification method.At present,the methods commonly used for online detection mainly include expert-systems dependent feature extraction manually and auxiliary inspection manually.Expert systems need experts who are professional to establish a complete knowledge base for fault detection,relying too much on experience and priori knowledge of the experts;The operator assistive judgment,in the manual assistive detection method,to the running noise during operation is subjective.Besides,the operator is easily to be fatigued.It is not only unable to meet the requirements of high efficiency and high precision,but also prone to false negative and misjudgment cases.As to the problems mentioned above,in this thesis we propose a method for recognizing and classifying the fault data of drivetrain with convolutional neural network,which realizes a direct-output mapping from the original vibration signal to the result of fault identification.This method provides important references for the efficient and high-precision identification and classification of fault data in electric drive assemblies.The main research contents of this thesis are as follows:(1)This thesis mainly focuses on the integrated drivetrain,collects and processes the fault signals during the manufacturing process.Firstly,the vibration signals of the electric drive assembly at the end of the actual production are collected,the origination and characteristics of the fault are analyzed and summarized.Secondly,the preprocessed vibration signals are expanded with the data-enhanced method,then the influence of asynchronous length on the recognition and classification accuracy are analyzed.The experimental results show that choosing a step size of 2000 is good for model training.Finally,the original vibration signal data set of electric drive assembly is established.(2)In order to improve the accuracy of the identification and classification of fault data in the drivetrain,the one-dimensional convolutional neural network model is optimized in three aspects: expanding the receptive field,reducing training parameters,and optimizing the model's operability.The experimental results show that the use of dilated convolutional layers instead of traditional convolutional layers increases the range of localreceptive fields,and makes the network model better in extracting features;Filter response normalization(FRN)eliminates the dependence of batch size(Batch size),which significantly improves the training rate and stability of network model;the introduction of global pooling reduces the parameters required for model training,and results in a significant improvement in training efficiency and parameter amount.At the same time,by combining with the visualization technology,the process of extracting the vibration signal features of the one-dimensional convolutional neural network model is explored,and the change of the weight frequency distribution is analyzed.(3)This thesis gives a comparative analysis of the original vibration signal and its processed signal.The results show that when the original vibration signal is used as the input of the network model,the network model can accurately classify the fault data.Finally,the one-dimensional convolutional neural network model suitable for identification and classification of the fault data set in the drivetrain is obtained,the test accuracy is98.00%.(4)In practice of the on-line testing process,the drivetrain is tested under variable operating conditions in most cases,where the original vibration signal obtained always be different.This thesis proposes a method,by which the attention mechanism is integrated into the one-dimensional convolutional neural network model structure,highlights the expression of important local features.The model is trained by varying combinations of variable operating data and the hyperparameters in the network model are optimized.The experimental results show that the complexity of the data set has a direct impact on the network model training.The average accuracy of fault recognition and classification is93.45%.At the same time,comparative analysis of the recognition and classification capabilities of different algorithms is carried out.The experimental results show that for the fault data set of the drivetrain,the ATT-CNN model can obtain a higher recognition and classification accuracy.Finally,the highest recognition and classification accuracy of the fault data set under the variable operating condition can be 95.19%.