[Paper Review] Particle Filter-Based Fault Diagnosis of Nonlinear Systems Using a Dual Particle Filter Scheme

Najmeh Daroogheh, Nader Meskin+1 more|arXiv (Cornell University)|Jun 28, 2016

Fault Detection and Control Systems44 references3 citations

TL;DR

This paper proposes a dual particle filter scheme for simultaneous state and time-varying parameter estimation in nonlinear stochastic systems, enhancing fault diagnosis in gas turbine engines. By decoupling state and parameter estimation into parallel filters, the method achieves higher accuracy (86.29%), lower false positive rate (5.71%), and improved precision compared to RML and Bayesian methods, with reduced computational overhead.

ABSTRACT

In this paper, a dual estimation methodology is developed for both time-varying parameters and states of a nonlinear stochastic system based on the Particle Filtering (PF) scheme. Our developed methodology is based on a concurrent implementation of state and parameter estimation filters as opposed to using a single filter for simultaneously estimating the augmented states and parameters. The convergence and stability of our proposed dual estimation strategy are shown formally to be guaranteed under certain conditions. The ability of our developed dual estimation method is testified to handle simultaneously and efficiently the states and time-varying parameters of a nonlinear system in a context of health monitoring which employs a unified approach to fault detection, isolation and identification is a single algorithm. The performance capabilities of our proposed fault diagnosis methodology is demonstrated and evaluated by its application to a gas turbine engine through accomplishing state and parameter estimation under simultaneous and concurrent component fault scenarios. Extensive simulation results are provided to substantiate and justify the superiority of our proposed fault diagnosis methodology when compared with another well-known alternative diagnostic technique that is available in the literature.

Motivation & Objective

• To address the limitations of single-filter augmented state-parameter estimation in nonlinear systems, particularly high dimensionality and computational burden.
• To improve fault detection, isolation, and identification (FDII) performance in nonlinear, stochastic systems with time-varying parameters.
• To reduce false alarm rates and enhance estimation accuracy in real-time health monitoring of critical systems like gas turbines.
• To develop a dual estimation framework that ensures convergence and stability under formal conditions.
• To validate the method’s superiority over existing particle filter-based FDII techniques using a realistic gas turbine engine model.

Proposed method

• The method employs a dual particle filter structure: one filter estimates system states, and another estimates time-varying parameters in parallel, avoiding state-augmentation.
• It extends the Bayesian parameter estimation framework to enable concurrent estimation without increasing model dimensionality.
• The approach uses particle filtering with resampling and importance sampling to approximate posterior distributions for states and parameters.
• A dual estimation strategy is formally proven to be convergent and stable under certain regularity conditions on the system and noise.
• The method applies a likelihood-based fault detection mechanism using negative log-likelihood, with fault isolation achieved through parameter estimation.
• The framework is evaluated using a gas turbine engine model under multiple simultaneous fault scenarios, including compressor and turbine efficiency and mass flow rate degradation.

Experimental results

Research questions

• RQ1 Can a dual particle filter framework outperform single-filter augmented state-parameter estimation in nonlinear fault diagnosis?
• RQ2 Does decoupling state and parameter estimation improve accuracy and reduce false positive rates in fault detection for nonlinear systems?
• RQ3 To what extent does the proposed method maintain stability and convergence in the presence of time-varying parameters and stochastic noise?
• RQ4 How does the dual filter approach compare to RML and Bayesian particle filtering in terms of fault identification accuracy and computational efficiency?
• RQ5 Can the method effectively handle multiple simultaneous component faults in a real-world nonlinear system like a gas turbine engine?

Key findings

• The proposed dual particle filter method achieved 86.29% accuracy in fault diagnosis, significantly outperforming the RML method (78.86%) and the Bayesian KS-based method (25.95%).
• The false positive rate was 5.71%, which is 5.71 percentage points lower than the RML method (11.43%), indicating fewer false alarms.
• Precision for all four health parameters—compressor efficiency, compressor mass flow, turbine efficiency, and turbine mass flow—was higher than in the RML and Bayesian methods.
• The method required only 50 particles to achieve superior performance compared to 150 particles for RML and 45 for the Bayesian KS-based method, indicating better efficiency.
• The confusion matrix analysis confirmed that the dual estimation method correctly identified 31 out of 33 instances of no fault and 28–31 out of 33 instances per fault category, demonstrating high true positive rates.
• The results validate that the dual filter structure enhances fault identification accuracy and reduces computational burden while maintaining stability and convergence.