Research On SVR-based Non-Mechanism Modeling And Fault Prediction

Modeling is an important tool used for analyzing and solving practical problems and have a wide range of applications whether it is in pure science or in the practical applica-tion of engineering fields, and is also an effective means that are applied to conduct fault detection and fault diagnosis, and to realize, reconstruct, retrofit and control the concerned plant at the same time. Considering the fact the complex nature of the studied objective are modeled is intractable in a real applications, non-mechanism model modeling based on the obtained data has become a research hotspots.Support vector machine (SVM) based on the statistical learning theory tries to find the optimal trade-off possessing the ability to simultaneously minimize the empirical risk and the structural complexity of the model by introducing the structural risk minimization which is characterised by the relationship between over-fitting and generation performance. SVM applied to regression or function estimation problem is referred to as support vector regression (SVR). This paper proposes two modeling approaches on non-mechanism model consisted of the deterministic model and the non-deterministic model within the framework of SVR, where non-deterministic model is characterised by interval regression model and the both model are finally applied to fault detection and fault prognosis, respectively. In a nutshell, contents of the paper can be summarized as follows:Aiming at the consequent parameters of the TS fuzzy model, a novel cost function based on decomposing least squares support vector regression (LSSVR) have been pro-posed to identify consequent parameters, in which makes use of the structural risk consid-ering how to control the trade-off between empirical risk and model complexity instead of the conventional empirical risk such as least squares (LS) methods with its variant, Kalman Filtering (KF) methods and locally optimal expectation maximum (EM) etc. And then, a new optimization problem is formulated by treating the obtained cost function as the ob-jective function, TS fuzzy model as constraint condition, and the consequent parameters of TS fuzzy model are derived by applying Lagrange method. In the process of constructing LSSVR, the better generalization performance for LSSVR have not only largely to do with the LSSVR’s model parameters (referred to as support value parameters), but also the selec-tion of the free parameters (hyper-parameters). Consequently, an online hyper-parameters updating method for LSSVR based on unscented kalman filter (UKF) which the system model adopt the LSSVR’s parameter choice as state variable and treat LSSVR model as the measurement equation respectively have been proposed. The proposed method weaken the requirements for the size of training data set and avoid more computational burden against updating parameters timely in the optimization problem.The problem of the traditional LSSVR approach, referred to as the global LSSVR ap-proach, is the incapability of interpreting local behavior of the estimated models, and the local LSSVR approach to deal indeed with local behavior of models has the problem of boundary effects, which may generate a large bias at the boundary and also need more time to calculate. As a result, the fuzzy weighted average LSSVR with a fuzzy partition is proposed for modeling nonlinear system based only on the obtained data and the pro-posed method overcome the problem of the bound effects at the boundary and avoid more computation time.As appose to the above introduced deterministic model, non-deterministic model mod-eling based on linear programming-SVR (LP-SVR) have been proposed, where the non-deterministic model is described as interval regression model consisted of the upper regres-sion model (URM) and lower regression model (LRM). First, the upper and lower l1-Norms and l∞-Norms with respect to upper bound approximation errors are constructed respec-tively, and the both norms subjected to respective constraints are integrated into LP-SVR to form new upper and lower optimization problems. Following that, optimization problem corresponding to URM and LRM are solved by linear programming and interval regres-sion model is thus constructed. The proposed method not only possesses the characteristics of adjusting a flexible interval spread, but also independently constructs URM and LRM, instead of adopting the traditional estimated center model and estimated radius of interval regression model which is the incapability of dealing with asymmetrical interval, and at the same time the complexity of model structure is under control.Combining the previous two modeling research on non-mechanism model, namely de- terministic model and non-deterministic model, the both models are to be applied to fault detection and fault prognosis respectively. For fault detection, an adaptive threshold model characterized by interval regression model has been constructed under the circumstances of measured data with no fault, in which adaptive threshold model is used to express a normal system operation and judge whether a fault occur or not. It is shown that to a large exten-t the proposed method applied to fault detection can overcome the problem of the longer detection latency and derive a better results of the fault detection. In addition, it is usually referred to as fault prognosis to how to evaluate the current or future state of the system be-ing studied and to judge whether the state is normal or not. Consequently, a fault prognosis method has been proposed by combining LSSVR with LP-SVR interval regression model. In order to update LSSVR model timely with new arrived measurements and to reduce more computational time in the process of updating parameters, the superiorities of using sliding-windows to reduce largely computation burden and implementing LSSVR model updating by Unscented Kalman Filter (UKF) have been considered. Accurate multi-step-ahead pre-diction over state with long future horizons are implemented by the updated LSSVR model and fault prognosis is completed by adaptive threshold model.Finally, all discussions in the dissertation are summarized, and several issues worthy of further work are presented.