Research And Application Of Filtering Methods For Fault Diagnosis Of Non-linear Systems

Along with the increasing requirement of reliability and safety of weapon equipments and mechantronic systems, fault detection, isolation(FDI) are playing very important role. While modern weapon equipments are becoming more and more complex, and one trait is that many systems and processes are nonlinear and disturbed by random noises and many kinds of uncertainty, which cause the accurate fault diagnosis to be very difficult. Among various techniques, nonlinear filtering method is an important method for the fault diagnosis of nonlinear systems. However, traditional filtering methods are mainly based on linearization or Gaussian hypothesis, which may influence the filtering precision and lead to low diagnosis accuracy rate(DAR), thus block their engineering application. How to break through the drawbacks of traditional methods and improve the DAR of nonlinear systems has been a key research content.Supported by the National Defense Advanced Research Project and required by practical engineering project, this dissertation aims at the four influence factors that are linearization error, modeling error, Gaussian hypothesis error and low information usage rate. The new nonlinear filtering based on the"probability approximation"concept and multi-source information fusion technology are adopted to improve the accuracy rate. The main contents are as follows:1. As to the fault diagnosis problem of nonlinear systems in Gaussian noise, the Unscented Kalman Filter(UKF) and fault decision methods are deeply studied and improved. By adding the orthogonal restriction to the residuals of nonlinear systems with modeling error, two different strong tracking UKF methods are proposed in order to enhance the tracking ability of faulty state and improve the precision of fault identification, thus lay a foundation to parameter bias type fault diagnosis for nonlinear systems: (1) By transforming the forgetting matrix to multi-dimensional nonlinear optimization problem with no restriction, the DFP(Daviden-Fletcher-Powell) strong tracking UKF is proposed; (2) The quick approximate strong tracking algorithm with multiple forgetting factors is proposed to reduce the calculation cost, which is suitable for high dimension nonlinear system. Furthermore, the fault detection and diagnosis strategy and realization flow are presented based on UKF and its improved algorithms, the filter selection and application range are analyzed and discussed. Simulation results show that the UKF and strong tracking UKF can reduce the influence of linearization error and modeling error compared to the traditional method, the diagnosis and identification accuracy rate is consequently improved.2. The fault diagnosis problem of nonlinear systems in non-Gaussian noise is deeply studied. In order to detect the fault of multi-dimensional systems in non-Gaussian noise with independent and dependent observation elements respectively,by defining independent and identical condition accumulative distribution function of observations, two new detection statistics variable and detection methods are proposed based on the Parzen window smoothing and Chi-Square detection. In order to increase the detection speed, by using mixing particle sets to express the posterior probability distribution and adopting the Monte Carlo numerical methods to calculate the mixing weights, the fast fault detection algorithm is proposed based on the estimate window method. Furthermore, the fault diagnosis strategy and realization flow are presented by joint estimation and likelihood ratio methods. Simulation results demonstrate the effectiveness of the new methods.3. In order to improve the DAR of nonlinear systems by multi-source information fusion besides the input and output information, the multi-knowledge fusion method is deeply studied.(1) A multi-knowledge fusion model is proposed, and a new dynamic case expression methods is then proposed in order to realize the case based reasoning for the fusion realization. In the following, a case representation method is proposed by treat the system characteristics set, symptom set and feature set as the case conditional attributes. Then the synthetic similarity measure method for case retrieve is proposed.(2) The three basic probability assignment methods of model, case and rule evidence are proposed for multi-knowledge fusion model.(3) As to the problem of evidence confliction of D-S evidence theory for decision level fusion, the"evidence importance"concept is firstly defined to import the partiality adjustment, then the conflict modification term is defined, thus a new combination method is proposed with the advantage of simpleness and practicality.4. The application of DAR improvement technologies are applied to the helicopter flight control systems. The small perturbation linearization model and nonlinear model are established, and the application results show that disturbed by Gaussian and non-Gaussian noises respectively, UKF and particle filters can improve the accuracy rate of fault detection. Then the parallel rudder loop experiment sub-system are established, the fault diagnosis systems are designed and realized by the studied techniques in this thesis, the experiment results indicate that the DAR and the diagnosis performance are improved effectively.