Amol S. Naik

On the application of PCA technique to fault diagnosis

In this paper, we briefly address the application of the standard principal component analysis (PCA) technique to fault detection and identification. Based on an analysis of the existing test statistic, we propose a new test statistic, which is similar to the Hawkins T 2h statistic but without the numerical drawback. In comparison with the SPE index, the threshold setting associated with the new statistic is computationally simpler. Our further study is dedicated to the analysis of fault sensitivity. We consider the off-set and scaling faults, and evaluate the test statistic by viewing its sensitivity to the faults. Our final study focuses on identifying off-set and scaling faults. To this end, two algorithms are proposed. This paper also includes some critical remarks on the application of the PCA technique to fault diagnosis.

Observer-based FDI Schemes for Wind Turbine Benchmark

Abstract In this paper, observer-based FDI schemes for wind turbines are proposed. This study is based on the benchmark model presented in Odgaard et al. [2009a]. For residual generation, Kalman filter and diagnostic observer based approaches are employed, and for residual evaluation, generalized likelihood ratio test and cumulative variance index are chosen. The fault isolation issue is solved with the help of dual sensor redundancy. Finally, the performance of the proposed FDI schemes is systematically evaluated by Monte Carlo studies.

Study on modifications of PLS approach for process monitoring

Abstract Partial least squares (PLS) is an efficient approach for multivariate statistical process monitoring. Although it works in many industrial applications, Zhou et al. [2010] revealed that some properties of PLS algorithm may hamper overall efficiency of process monitoring scheme. To solve these problems, a modified approach is proposed in this paper. Compared with the existing PLS approaches, the new approach performs an orthogonal decomposition on regression variable space to eliminate the variations useless for output prediction. Based on the new approach, a complete process monitoring scheme is also developed. Finally, the effectiveness of the proposed approach is verified on an industrial benchmark of Tennessee Eastman process.

An Approach to Data-Driven Adaptive Residual Generator Design and Implementation

Abstract This paper addresses data-driven design and implementation of adaptive observer based residual generators for discrete-time systems. The basic idea behind this study is the application of an one-to-one mapping between a parity vector and the solution of Luenberger equations and the data-driven identification of parity space. For the realization of the adaptive residual generation, standard adaptive technique is applied. The proposed approach is demonstrated on the laboratory three-tank-system.

On PCA-based fault diagnosis techniques

This paper presents the application of standard PCA technique to fault diagnosis system design. Based on the fault detectability analysis of existed test statistics, the joint use of some test statistics is recommended. Our further study is dedicated to develop a fault isolation approach based on likelihood ratio test, on which the off-set and scaling fault can be easily isolated with explicit off-set fault direction and scaling fault classification. The issues of off-set and scaling fault identification will be also discussed and the complete scheme of PCA-based fault diagnosis procedure is proposed.

Adaptive process monitoring based on parity space methods

Abstract Motivated by data-driven design of fault detection system, a recursive algorithm is developed for updating the parity relation within the framework of subspace identification methods (SIM). The presented recursive algorithm benefits from the signal processing methods for noise subspace tracking to reduce the computationally burdensome updating of singular value decomposition (SVD), which is a crucial step in SIM. Based on the recursive updated parity relation, the residual generator for the fault detection purpose is constructed and the issues of residual evaluation and threshold computation are discussed further. The adaptive process monitoring scheme that integrates the aforementioned issues, is proposed and tested on the laboratory three-tank-system.

Subspace based identification of diagnostic observer for Wiener systems

Abstract A subspace identification technique based design of observer is developed in this paper. Verhaegen and Westwick [1995] provided a subspace algorithm to identify state space matrices for LTI system with nonlinearity on the output side, or the so called Wiener systems. Their results were based on Bussgangs theorem on crosscorrelation functions of amplitude distorted Gaussian signals. We have combined these two important results in order to develop an observer based fault detection system that is identified directly from the test data generated by nonlinear dynamic systems.

Development and application of a data-driven fault prediction method to the imperial smelting furnace process

Abstract This paper addresses the development and application of a data-driven method to the fault diagnosis in imperial smelting furnace (ISF). Based on the method of the weighted least squares vector machines regression, a Hammerstein model is constructed and identified for the ISF. This model is used to predict the dynamic behavior of the furnace and the possible faults in the process. The simulation study shows that the identified model well adapts to the changes in the structural parameters and provides accurate prediction.

Recursive Identification Algorithm for Parity Space Based Fault Detection Systems

Abstract The problem of recursively identifying parity space in the framework of subspace technique is studied. Updating the entire singular value decomposition, a crucial step in identification, is computationally burdensome and sometimes not even feasible. Hence a recursive eigenvalue decomposition based identification method is recommended in the literature. The algorithm developed here updates the eigenstructure of covariance matrix of input and output data after every new measurement and gives a new parity space. The method improves the fault detection performance against uncertain parameter variations and in non-stationary noise environment. The proposed algorithm is applied to hybrid simulation platform of continuous stirred tank reactor.