Research On Inertial Sensors Fault Diagnosis Of Intelligent Vehicle

With the application of automation technology in all walks of life, the automation systems have become more and more complex. As the main information acquirement device, the sensors guarantee a reliable, secure and stable work. However, if the sensors have low performance, failure, and even become useless, it will have serious effect on system monitor and control, and cause incalculable damage. Thus, the sensor fault diagnosis is very important.This dissertation is supported by the National Natural Science Foundation of China under Grant90820302, Research on the Key Scientific Problems of Intelligent Vehicle Driving on the Highway. As one part of the project, the dissertation focuses on anomaly detection and fault diagnosis of inertial sensors and navigation systems in unmanned vehicle. Three aspects are researched, including digital signal process, fault diagnosis and fault prediction methods. Firstly, the original sensor data is processed by using digital signal process methods to remove noise, reduce uncertainties and improve the data accuracy. Then, the sensor fault diagnosed by using different precision sensors in redundancy way, which combines the advantages of software redundancy method and hardware redundancy method; Finally, the fault prediction technology based on particle filter is researched to improve the real-time of the fault diagnosis.Main contributions of the dissertation are shown as following:1) A real-time Kalman filter based on the non-decimated Haar algorithm is proposed to avoid the problem of large computation and long delays in existing multi-scale Kalman filter. The simple addition, subtraction, and shift operations are used to complete multi-scale transformation at time t, and the signal is reconstructed after de-noising by wavelets soft-threshold and Kalman filter in each scale. To verify the validity of this method, the experiments with low-precision acceleration sensor for intelligent vehicle are carried out. The experimental results show that the repeated computation is reduced and the speed of the algorithm is improved by using the non-decimated Haar algorithm. The method can effectively improve the sensor performance. Meanwhile, the de-noising performance of this method is superior to Kalman filter when the status can not be estimated correctly.2) A fault diagnosis method with different precision redundant sensors is proposed in this thesis. The method uses the principle of minimizing uncertainties of the dynamic model and maximizing the impact of fault to preprocess the low-precision sensor data, taking turns to use a sensor data as input, and the other sensor as output to establish the Kalman filter equations, and the fault diagnosis can be made by the obtained new information. The experiments show that, the proposed method can not only effectively suppresses the noise in low-precision sensors, but also can reduce costs and complexity of system modeling. Meanwhile, in the moving process of the intelligent vehicle, the background noise changes badly, which causes the diagnosis for multiple-precision sensor failure more difficult. Thus, the wavelet-based noise estimation of the two stage Kalman filter fault diagnosis method is proposed. The relationship between the fault detection rate and noise intensity is also discussed, and the result shows that the method improves the accuracy of fault diagnosis and has relatively good robustness.3) An adaptive particle filter based on posterior distribution is proposed. The prior knowledge is used to set the confidence interval of likelihood, and the number of particles is adjusted by the posterior estimation in the confidence interval. Then the particle filter is used in fault diagnosis and fault prediction. An algorithm of fault prediction base on particle filter is proposed. After k-step iterative, the algorithm uses k groups particles to estimate the system status and get the fault status distribution and the probability of failure, then the change rate of the failure probability is obtained to predict the fault type and the time of the failure. The experimental results of the mobile robot dead-reckoning system demonstrate the effectiveness and feasibility of the method.