Research On Fault Diagnosis And Related Algorithm Based On Hidden Markov Model

The industry control system tends to be more complex, integrated and intelligent. Once the complicated process fails, it does not only bring enocomic losses but also may result in human casualties. Therefore the fault diagnosis system is particularly important with the meaning of fault detection, isolation and identification. Model based fault diagnosis method will lose effectiveness if strong nonlinear coupling property exists in the complex system. On the contrary, the large-scale data that reflect the operation performance of the system can be collected and stored due to the wide spread of computer network and advanced sensors As a result, the developmet of data-driven fault diagnosis approaches has urgent practical needs and significant meanings.In this paper, industry process with nonlinearity, uncertainty and complexity is taken as the research object of Hidden Markov Model based fault diagnosis. As one of the typical data-driven methods, the performance of Hidden Markov Model based fault diagnosis depends to a large extent on the effectiveness of diagnosis strategy and Hidden Markov Model algorithm accuracy. Thus the main research objectives of this paper are focused on two aspects:the improvements of Hidden Markov Model based fault diagnosis strategy and the modifications of parameters estimation algorithm for Hidden Markov Model.Firstly, the Hidden Markov Model based degradation process modeling and fault diagnosis strategies for industry process are studied. The left-right Hidden Markov Model is modeled with system historical degradation life data and the Expectation Maximization algorithm is used to learn parameters of the modeled Hidden Markov Model. Then a maximum a posteriori current health condition assessment approach and a future health condition prediction approach are proposed based on the modeled Hidden Markov Model. A parameters augmentation method is proposed to make these HMMs with both consistent and inconsistent dimensions situations unified to a fault identification system. It is also pointed out that the proposed parameters augmentation method will not change any original character of Hidden Markov Model.Secondly, in the application of Hidden Markov Model many problems met in practice such as physical considerations, practical engineering requirements and prior knowledge can be formulated as constrained problems. Proper constraints can modify the unsatisfactory observed information and make the estimate of Hidden Markov Model parameters closer to the considered systems. Therefore, we extend the parameters estimation research from general Hidden Markov Model to new Hidden Markov Model where the state transition matrix has nonlinear inequality constraints. Then an approximate parameter estimation algorithm is proposed.In addition, considering that the solution of approximate parameter estimation algorithm is only a feasible one but usually not optimal, though the inequality constraints could be satisfied, an active set based Expectation Maximization algorithm is proposed to estimate the optimal Hidden Markov Model parameters when inequality constraints are proposed and the convergence is also demonstrated.Multiple observation sequences would contribute to make reliable estimates of Hidden Markov Model parameters with more sufficient training data. It would also make the modeled HMM describe the target system more precisely. Based on the above considerations, Hidden Markov Model parameters estimation with independent multiple observation sequences and nonlinear inequality constraints is studied. Based on the proposed new HMM, an expandable on-line dynamical process fault diagnosis framework, which consists of on-line fault detection phase and on-line fault identification phase, is then developed. The effectiveness of the proposed HMM approach for fault diagnosis is validated on a realistic industrial process called Tennessee Eastman Process. The experimental results of Tennessee Eastman Process indicate that the proposed approach provides better performance than Principal Component Analysis, Independent Component Analysis and HMM with single observation sequence based methods. Although the independence assumption of multiple observation sequences can make the considered parameters estimation problem simple, the dependence phenomenon may happen in practice. In order to extend to general situation, we consider HMM parameters estimation with inequality constraints and multiple observations imposing no independence assumptions. The likelihood function is replaced through setting conditional probabilities and constructing weighting function. An auxiliary function is devised for reconstructing the optimization objective function and the convergence of the proposed algorithm is also demonstrated. Two special cases with independence assumption and uniform dependence assumption are also listed. Example on Tennessee Eastman process shows that the diagnosis accuracy can be improved further without the independence assumption.Finally,all discussions in the dissertation are summarized, and several issues worthy of further work are presented.