Fast Bayesian equipment condition monitoring via simulation based inference: applications to heat exchanger health

Peter Collett; Alexander Johannes Stasik; Simone Casolo; Signe Riemer-Sørensen

✨ TL;DR

This paper proposes a Simulation-Based Inference (SBI) framework using neural networks for fast Bayesian condition monitoring of industrial equipment, specifically heat exchangers. The approach achieves 82× speedup over MCMC while maintaining comparable diagnostic accuracy and uncertainty quantification.

01 · Problem

Industrial equipment condition monitoring requires inferring hidden degradation parameters from indirect sensor measurements while quantifying uncertainty. Traditional Bayesian methods like MCMC provide rigorous uncertainty quantification but are computationally expensive, making them impractical for real-time process control and digital twin applications where fast inference is critical.

02 · Approach

The authors develop an AI-driven framework using amortized neural posterior estimation, a simulation-based inference technique. Neural density estimators are trained on simulated thermal-fluid data from heat exchanger models to learn a direct, likelihood-free mapping from observations to the posterior distribution of degradation parameters. This eliminates the need for expensive likelihood evaluations required by traditional MCMC sampling.

03 · Key insights

What the paper shows.

01Amortized neural posterior estimation enables near-instantaneous inference after training, making real-time condition monitoring feasible

02SBI achieves comparable diagnostic accuracy to MCMC while providing reliable uncertainty quantification without likelihood computation

03The framework successfully handles challenging scenarios including low-probability sparse-event failures and multiple failure modes (fouling and leakage)

04Likelihood-free inference is particularly valuable for complex physical systems where likelihood functions are intractable or expensive to evaluate

04 · Results

The SBI framework achieves 82× acceleration in inference time compared to MCMC baselines while maintaining comparable diagnostic accuracy across synthetic fouling and leakage scenarios. The method provides reliable uncertainty quantification and successfully diagnoses complex failure modes, including rare, low-probability events.

05 · Limitations

The approach relies on the quality and representativeness of the simulated training dataset; performance depends on how well the simulator captures real-world heat exchanger physics. The paper focuses on synthetic scenarios and does not demonstrate validation on real industrial data. Generalization to other equipment types or failure modes beyond heat exchangers is not explored. The computational cost of generating the training simulation dataset is not thoroughly discussed.

✨ Generated by Claude · Apr 25, 2026 · Read the PDF for authoritative content.

What the paper shows.

↘ Related papers