Research On Automobile Sensor Fault Detection And Classification Based On ICA

With the continuous development and innovation of autonomous driving technology,automobile has gradually become an essential means of transportation in people's daily life.In order to ensure the reliability and safety of automobile system operation and reduce the frequent occurrence of traffic accidents,the process monitoring of automobile system has become the main research content at present.The data-driven fault diagnosis method does not need to build accurate analytical model and requires less prior knowledge.It only needs to collect the process data of the automobile system to establish an intelligent fault diagnosis model.Therefore,compared with the analytical model and knowledge-based method,this method is more suitable for the process monitoring of today's automobile system.In the scene of automobile faults,actuator and sensor faults account for a high proportion.Therefore,it is of great theoretical and practical significance to conduct research on automobile sensor fault diagnosis.In this paper,the research work of fault detection and classification for automobile sensors mainly includes:(1)The traditional fault diagnosis method based on Principal Component Analysis(PCA)assumes that the process variables conform to Gaussian distribution.The process data of automobile sensor has strong non-Gaussian characteristics.Although Fast Independent Component Analysis(FastICA)algorithm can be applied to the fault diagnosis of non-Gaussian process data,it is easy for FastICA algorithm to fall into local optimal value using Newton iteration method.Therefore,this paper introduces Adam optimization algorithm to propose an improved ICA algorithm.Through the simulation experiment of automobile sensor,it is verified that the improved ICA algorithm is more effective than FastICA algorithm in fault detection.(2)Automobile process data is nonlinear and non-Gaussian.Sparse denoising autoencoder(SDAE)can reduce the feature dimension and improve the nonlinearity and robustness of fault detection system.In order to further improve the detection effect,this paper proposes an ICA-SDAE automobile fault detection method which combines FastICA algorithm and SDAE.This method first uses FastICA algorithm to extract the independent components(non-Gaussian components)of process data to obtain the feature space,and then uses SDAE to reduce the dimension of independent components and reduce the interference of irrelevant signals;using SDAE in the residual space to extract the Gaussian information of process data;the corresponding statistical calculation method is proposed for automobile fault detection.The validity of the algorithm is verified by the simulation experiment of automobile sensor and comparing the detection effect of this method with the ICA-PCA method.(3)The above two methods have good fault detection capabilities for automobile systems,but they cannot classify the faults effectively.Support vector machine(SVM)has the problem that they cannot effectively use the time series characteristics of automobile process data.Considering that long short term memory(LSTM)network can fit sequence data effectively,this paper proposes the application of LSTM network in automobile fault classification.Because the data distribution between different layers is easy to change in the process of neural network training,this paper introduces batch normalization method and combines Adam optimization algorithm to improve the accuracy of fault classification.Through the simulation experiment of automobile sensor,the classification effect of this method is compared with PCA-SVM method.The results show that this method can significantly improve the accuracy of automobile system fault classification.