Adaptive Particle Filtering for Fault Detection in Partially-Observed Boolean Dynamical Systems.

Area of Science:

• Systems Biology
• Control Theory
• Computational Biology

Background:

• Partially-observed Boolean dynamical systems (POBDS) are stochastic, nonlinear, and derivativeless, posing challenges for traditional fault detection.
• Existing methods struggle with the complexity and partial observability inherent in POBDS, particularly in biological networks.

Purpose of the Study:

• To propose a novel methodology for fault detection and diagnosis in POBDS.
• To address scenarios with known normal operation but unknown fault modes, and scenarios with known finite fault models.
• To enhance computational efficiency for large-scale systems.

Main Methods:

• Utilized an innovations filter (IF) for detecting deviations from normal operation.
• Employed multiple model adaptive estimation (MMAE) with a likelihood-ratio (LR) statistic for known fault models.
• Integrated an adaptive expectation-maximization (EM) algorithm for parameter estimation.
• Applied particle filtering techniques to manage computational complexity in large state-spaces.

Main Results:

• Demonstrated the efficacy of the proposed methodology through numerical experiments.
• Successfully identified stuck-at faults in a large gene regulatory network (GRN).
• Validated the approach using a single noisy time series of RNA-seq gene expression data.

Conclusions:

• The novel methodology provides effective fault detection and diagnosis for POBDS.
• The approach is robust and applicable to complex biological systems like GRNs.
• Advanced filtering and estimation techniques offer a viable solution for analyzing noisy, high-dimensional biological data.