Mingxing Jia

Recursive parameter identification of Hammerstein–Wiener systems with measurement noise

Abstract A recursive algorithm is proposed in this paper to identify Hammerstein–Wiener systems with heteroscedastic measurement noise. Based on the parameterization model of Hammerstein–Wiener systems, the algorithm is derived by minimizing the expectation of the sum of squared parameter estimation errors. By replacing the immeasurable internal variables with their estimations, the need for the commonly used invertibility assumption on the output block can be eliminated. The convergence of the proposed algorithm is also studied and conditions for achieving the uniform convergence of the parameter estimation are determined. The validity of this algorithm is demonstrated with three simulation examples, including a practical electric arc furnace system case.

Multi-kernel learnt partial linear regularization network and its application to predict the liquid steel temperature in ladle furnace

In this study, a novel prediction model, hybrid of mechanism method, Takagi-Sugeno (T-S) fuzzy modeling, regularization network technique and Multi-kernel learning algorithm, is proposed for accurately forecasting the liquid steel temperature in ladle furnace (LF). By virtue of mechanism method and TS fuzzy modeling technique, a partial linear structured mechanism model is firstly obtained, which contains a parametric linear part with unknown coefficients and a non-parametric part with unknown functional expression. Thereafter, it is parameterized and implemented by a modified regularization network, called partial linear regularization network (PLRN), which introduces a parametric linear part into the traditional regularization network. Furthermore, to optimally design the kernel of PLRN and thereby further improve the prediction performance, Multi-kernel learning approach is employed to obtain the so called Multi-kernel learnt PLRN. The principal innovation behind the proposed method is the embedding of the prior knowledge into the model, instead of directly predicting the steel temperature using machine learning techniques which is commonly used in the previous steel temperature prediction models. This innovation leads to better final results in reducing the model complexity, improving the generalization performance and consequently promoting the prediction precise. The experiment results demonstrate that the novel predictor is superior in prediction performance over other black-box based predictors. Furthermore, the prediction accuracy is boosted via Multi-kernel learning.

Intermittent-Chaos-and-Cepstrum-Analysis-Based Early Fault Detection on Shuttle Valve of Hydraulic Tube Tester

To ensure the safety and continuity of production, make a reasonable maintenance plan, and save the cost of maintenance for a hydraulic tube tester, it is needed to quickly identify an assignable cause of a fault. This paper is concerned with early fault detection of shuttle valves, which are the key components in hydraulic tube tester. An intermittent-chaos-and-cepstrum-analysis-based method is proposed to detect this early fault on the hydraulic tube tester. The presented approach is based on the insight that the phase transition of Duffing oscillator is very sensitive to a periodic weak signal having tiny angular frequency difference with the reference signal in this oscillator. Thus, to determine the angular frequency of the reference signal, cepstrum analysis is used to determine the eigenfrequency of the fault signal, and then, the angular frequency of the referential signal is computed according to this eigenfrequency. Finally, a Lyapunov exponent is introduced to realize computer observation of the shuttle valve early fault. The presented method is experimented with data simulated from an AMESim model of hydraulic tube tester. The results indicate that the proposed approach is capable of detecting the signal of shuttle valve early fault in the hydraulic tube tester.

Recursive parameter estimation for Hammerstein-Wiener systems using modified EKF algorithm

This paper focuses on the recursive parameter estimation for the single input single output Hammerstein-Wiener system model, and the study is then extended to a rarely mentioned multiple input single output Hammerstein-Wiener system. Inspired by the extended Kalman filter algorithm, two basic recursive algorithms are derived from the first and the second order Taylor approximation. Based on the form of the first order approximation algorithm, a modified algorithm with larger parameter convergence domain is proposed to cope with the problem of small parameter convergence domain of the first order one and the application limit of the second order one. The validity of the modification on the expansion of convergence domain is shown from the convergence analysis and is demonstrated with two simulation cases.

Hybrid modelling for real-time prediction of the sulphur content during ladle furnace steel refining with embedding prior knowledge

Abstract Real-time prediction of the sulphur content of steel is of great importance for operation guidance during ladle furnace (LF) steel refining. For seeking an accurate prediction, this paper proposes to establish sulphur content prediction model in a hybrid way, where a simplified first principle model is introduced and fine tuned by data-driven modelling methods. The derived hybrid model employs optimization approach to optimize its data representation part, while prior knowledge is embedded in the form of linear constraints. An innovation of the proposed methodology is the full exploitation of prior knowledge about the process for determining reasonable process parameters. Moreover, a novel optimization approach is developed for ensuring accuracy and improving solution efficiency by the integration of genetic algorithm and successive approximation method. The proposed hybrid model possesses flexible interpretable structure and adaptive learning ability. As a result, it ensures the extrapolation property for real-time prediction and is able to provide an in-depth understanding of practical desulphurization process, making it very suitable for process monitoring and operations optimization during LF steel refining. Finally, this hybrid model is validated on recorded data from an industrial LF plant.

A sub-stage moving window GRNN quality prediction method for injection molding processes

For injection molding process, a typical multistage batch process, the final product qualities are usually available at the end of the batch, which make it difficult for on-line quality control. A sub-stage moving window generalized regression neural network (GRNN) is proposed for dedicating to reveal the nonlinearly and dynamic relationship between process variables and final qualities at different stages. Firstly, using an clustering arithmetic, PCA P-loading matrices of time-slice matrices is clustered and the batch process is divided into several operation stages, the most relevant stage to the quality variable is defined, and then applying moving windows to un-fold stage data according to time, and sub-stage GRNN models are developed for every windows for on-line quality prediction. For comparison purposes a sub-MPLS quality model of every moving window was establish. The results prove the effectiveness of the proposed quality prediction method is supervior to sub- MPLS quality prediction method.

A Modeling Method for Time Series in Complex Industrial System

The data of complex industrial system were usually arrayed in the form of time series. This paper put forward the multivariate time-delayed principal component regression (MTPCR) method, which utilized the historical time series in the production process so as to establish a systematic prediction model. This method can calculate the delayed time of each input and output tunnel by which the modeling data were selected. The model established can predict the production outcome and product quality accurately in accordance with real-time input. With the aid of Simulink data and Matlab arithmetic, this paper concludes that MTPCR method possesses higher precision compared with other method

A stage-based quality prediction and control method for batch processes

A quality prediction method is proposed for multistage batch process. Firstly, using clustering methods based on the stage-based sub-PCA, the process is divided into several operation stages according to the change of process correlation, the most relevant stage to the quality variable is defined, and then sub-MPLS models are developed for that stage for online quality prediction. In addition, a closed-loop quality control system is proposed. The proposed method is finally applied to the injection molding process, a typical multistage batch process, and the results prove the effectiveness of the proposed quality prediction method and quality control scheme.

PCA-SDG based fault diagnosis on CAPL furnace temperature system

PCA-SDG based fault diagnosis method and its application on Continuous Annealing Process Line (CAPL) furnace temperature system are mainly discussed. Principle component analysis (PCA) method is applied to build the process monitoring model with a large number of historical data under normal operation conditions. High-dimension process data with noise and linear correlation are projected onto low-dimension and orthogonal sub-space. Real-time monitoring of furnace temperature system is carried out through online calculating T2 and SPE statistics of PCA model. When a fault is detected, the signed directed graph (SDG) model of furnace temperature system is used to interpret the residual contributions of PCA model, and then perform fault diagnosis with the rules of SDG. PCA-SDG method combines the advantages of both PCA and SDG methods. The effectiveness and reliability of the proposed PCA-SDG method are verified by the simulations.

Corrigendum to ''Multi-kernel learnt partial linear regularization network and its application to predict the liquid steel temperature in ladle furnace'' (Knowl.-Based Syst. 36 (2012) 280-287)

Models Comparison items We published a hybrid model for forecasting the ladle furnace (LF) liquid steel temperature in Knowledge-Based Systems [1] as an extension of the predictor presented in Steel Research International [2]. The fundamental of LF thermal modeling and partial linear regularization network (PLRN) algorithm are the basis of the completed model development and therefore introduced in both [1,2] with minor amendment. As one of the factors affecting the temperature variation, temperature change DTArc caused by arc heating was initially described in a nonparametric way in [2], while in [1] it was parameterized by T-S fuzzy method with embedding its prior knowledge, thereby improving the prediction accuracy. Moreover, Section 3 in [1] proposed a novel partial linear fitting method, termed multi-kernel learnt partial linear regularization network (MKL-PLRN), to fit the hybrid model. This MKL-PLRN employed multi-kernel learning (MKL) method to optimize the kernel function of PLRN presented in [2]. In summary, compared with the prediction model in [2], the hybrid predictor proposed in [1] exploited both structural information and the prior knowledge of the thermal system for boosting performance. There exist some advantages of the hybrid predictor: