Theoretic And Methodologic Research On Fault Diagnosis Of Mobile Robots Based On Adaptive Particle Filters

Fault detection and diagnosis (FDD) is critical in motion control, localization, map construction, navigation and safety for mobile robots. When operating in complex hazard environments such as planetary exploration, nuclear waste cleaning, military surveilance, etc, the sensor, control, and mechanical parts of mobile robots are prone to be faulty. Meantime, human intervention is usually limited, delayed or even impossible. Moreover, mobile robot is a kind of highly non-linear/non-Gassian dynamic system, in which priori knowledge for fault diagnosis is sparce. For these reasons, data based and linear model based diagnosis approaches are difficult to be applied in mobile robots.Particle filter (PF) is a sequential Monte Carlo (i. e. choosing randomly) method to monitor dynamic systems, which non-parametrically approximates probabilistic distribution using weighted samples (i. e. particles). PF gives a computationally feasible method for state estimation of hybrid systems. Furthermore, a single particle can represent discrete and continuous states simultaneously and can represent any distribution (including non-Gaussian). PF has been widely used in navigating, locating, tracking and so on in mobile robotics. Most recently, it is employed to handle fault diagnosis problems. However, for high dimensional state space, the contradiction of efficiency and accuracy has to be solved, i. e. it needs large number of particles to reach accurate approximation and it also has to reduce particle numbers to achieve real-time response.In this paper, theoretical studies and/or practical applications of improved adaptive particle filters are discussed for abnormality detection and fault diagnosis in mobile robot navigation control systems. Main contributions of the dissertation are shown as following,An adaptive particle filter framework is put forward, which can integrate various kinds of adaptation schemes such as proposal distribution adaptation, state space adaptation, particle number adaptation, evolutionary operators adaptation, etc. State space and distribution adaptation are key factors to determine particle numbers. The particle number is adjusted according Kullback-Leibler Divergence (KLD) of two distributions approximating with two particle sets of different particle numbers, and theoretical proof is given.A general framework, which integrates domain knowledge into particle filters, is put forward to improve logical inference abilities of particle filters. A fuzzy adaptive particle filter is implemented to dignose failures of dead-reckoning sensors of a mobile robot. The key idea is to employ fuzzy logic to represent domain constraints, which constrains sampling space to a fuzzy subset of the whole state space. Fuzzy subsets can be represented with transistional probability matrix.For 'soft faults' such as slippage of wheels and noise or scale factor faults of sensors, two adaptive particle filters are given to compensate large noise with laser range finder measurement, i. e. PD-PNAPF and PD-EAPF. Five kinds of residual features are extracted and the soft fault detection is reached real-timely with threshold logic. PD-PNAPF integrates proposal distribution adaptation and particle number adaptation. PD-EAPF integrates proposal distribution adaptation and mutation operater adaptation. Key idea of proposal distribution is to adaptively determine the variances of linear speed and angular speed according to residual features. In PD-EAPF, mutation operator is employed to alleviate sample improvishment caused by resampling step.For hybrid dynamic systems with incomplete models, a particle filter, PFIM, is given to detect unknown fault and diagnose known faults. Un-modeled dynamics of complex systems are described with 'unknown-fault' model. When the hybrid system is imperfectly modeled, general particle filter (GPF) may diverse. Firstly, the reasons for the diversity of GPF are discussed. Secondly, two kinds of statistics of particles are put forward, i. e. normalization factor of particles, W, and belief of maximal a-posteriori probability estimation state, B. Based on these, threshold logic is presented to detect unknown-faults. The theoretical proof of provided method is given under some reasonable assumptions.A robust measurement model for laser range finder is presented which is robustness to lader anormality and dynamic environment. Firstly, the noise beams are filtered out with segment analysis. Secondly, a 'reverse ray tracing' technique is put forward to calculate expected distance of an individual beam quickly. Thirdly, the beams with higher likelihood are employed to construct likelihood of the whole scan. The robustness of the provided model is two folds, (1) abnormal beams of laser range finder are filtered out with segment analysis method, (2) beams affected by dynamical objects are also picked out. The effectiveness of the presented method is testified with robust simutaneous locating and mapping for mobile robots in complex dynamic environments.