🤖 machine learning

Developing Distance-Aware Physics-Constrained Probabilistic Frameworks for Industrial Prognostics

This paper introduces two sampling-free, distance-aware physics-constrained probabilistic frameworks, PC-SNGP and PC-SNER, which utilize spectral normalization and dynamic loss weighting to achieve robust, well-calibrated uncertainty estimates and improved prognostic accuracy for industrial rolling-element bearings under distributional shifts and adversarial perturbations.

Original authors: Waleed Razzaq, Yun-Bo Zhao

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Waleed Razzaq, Yun-Bo Zhao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a mechanic trying to predict when a machine part (like a bearing in a car engine) is about to break. You have a computer program that watches the machine's vibrations and guesses its remaining life.

The problem with most current computer programs is that they are overconfident. If you show them a machine they have never seen before, or if the machine starts acting strangely, the program might still give a very specific answer with a tiny "error bar," as if it knows exactly what's happening. In reality, it should be saying, "I'm not sure, this looks weird to me."

This paper introduces two new computer programs designed to be humble and aware of their own limits. They are built to know when they are in "familiar territory" and when they are in "unknown territory."

Here is how the authors built these smarter programs, explained through simple analogies:

1. The "Distance-Aware" Map

Think of the training data (the examples the computer learns from) as a well-mapped city.

Old Programs: If you drive 100 miles outside the city into the wilderness, the old program still acts like you are on Main Street. It gives you a precise address, even though it has no idea where you are.
New Programs (PC-SNGP and PC-SNER): These programs have a special "distance sensor." As you drive further away from the city (the training data), the program's confidence drops. It starts saying, "I know the city well, but out here in the wilderness, my map is fuzzy, so I'm going to give you a wide range of possibilities."

2. The "Physics" Rulebook

To make sure the program doesn't just guess randomly, the authors gave it a Rulebook of Physics.

Imagine the machine part is a balloon slowly losing air. Physics tells us the air must go down, never up, and it usually speeds up as it gets smaller.
The new programs are forced to follow these rules. If the computer tries to predict that the balloon is suddenly inflating, the "Physics Rulebook" slaps its hand and says, "No, that violates the laws of nature." This keeps the predictions realistic even when the data is messy.

3. The "Stretchy Rubber Band" (Spectral Normalization)

This is the technical secret sauce that makes the "distance sensor" work.

Imagine the computer's internal brain is a stretchy rubber sheet. If you poke it in one spot, the whole sheet shouldn't warp wildly; the movement should be proportional.
The authors used a technique called Spectral Normalization to ensure the rubber sheet behaves nicely. It guarantees that if two inputs are slightly different, the computer's internal understanding of them will also be slightly different. This prevents the computer from getting confused and giving wild answers when the input changes just a little bit.

4. Two Different Ways to Be Uncertain

The paper proposes two specific versions of this system, which use different math tricks to measure uncertainty:

PC-SNGP (The "Gaussian" Approach): This version replaces the final guess with a "Gaussian Process." Think of this as a cloud of possible answers. The further you get from the training data, the bigger and fluffier the cloud gets, visually showing that the answer is less certain.
PC-SNER (The "Evidence" Approach): This version asks the computer to bet on how much "evidence" it has. If the input is weird, the computer says, "I don't have enough evidence to be sure," and it automatically widens its safety margin.

5. The "Dynamic Weight" Balancer

Training these programs is like balancing a scale. On one side is Data (what the sensors actually saw), and on the other is Physics (what the laws of nature say should happen).

Sometimes the sensors are noisy, and sometimes the physics model is a simplification.
The authors created a Dynamic Weighting system. It's like an automatic pilot that constantly adjusts the scale. If the data looks shaky, it leans more on the physics. If the physics model seems too rigid for the current situation, it leans more on the data. This keeps the program stable.

6. The "Stress Test" Results

The authors tested these programs on real-world bearing data (rolling-element bearings) from three different public datasets.

Accuracy: They predicted the remaining life of the bearings very accurately.
Honesty: When tested on bearings they had never seen before (or bearings with strange faults), these programs correctly increased their "uncertainty bands." They didn't pretend to know the answer.
Toughness: They even tested the programs by adding "noise" (fake, chaotic signals) to the data, like trying to hear a whisper in a hurricane. The new programs stayed calm and didn't break, whereas older programs got confused and gave wild predictions.

Summary

In short, the authors built two new tools for predicting machine failure. These tools are physically grounded (they follow the laws of nature), distance-aware (they know when they are out of their depth), and uncertainty-calibrated (they tell you how sure they are). This makes them much safer and more reliable for industrial use than the "overconfident" tools currently in use.

Technical Summary: Developing Distance-Aware Physics-Constrained Probabilistic Frameworks for Industrial Prognostics

Problem Statement
Industrial prognostics requires reliable uncertainty quantification (UQ) to prevent unsafe maintenance decisions caused by overconfident predictions. While existing literature offers physics-based models (PBMs), data-driven models (DDMs), and physics-constrained models (PCMs), significant gaps remain. PBMs often lack robustness in complex systems and struggle to quantify uncertainty. DDMs, even probabilistic extensions, typically fail to capture distributional shifts, producing overconfident predictions when inputs move away from the training manifold (Out-of-Distribution or OOD). Furthermore, existing probabilistic PCM approaches (e.g., Bayesian methods, Monte Carlo dropout, Deep Ensembles) are often computationally expensive due to sampling-based inference and lack "distance awareness"—the ability to reliably increase predictive uncertainty as inputs deviate from training data.

Methodology
The authors propose two sampling-free, distance-aware physics-constrained probabilistic frameworks: PC-SNGP (Physics-Constrained Spectral Normalization Gaussian Process) and PC-SNER (Physics-Constrained Spectral Normalization Evidential Regression).

Distance-Preserving Representations: Both frameworks enforce bi-Lipschitz distance-preserving mappings from the input space to the latent space. This is achieved by applying Spectral Normalization to the hidden layer weights, ensuring that the distance between inputs is approximately maintained in the hidden representation. This mechanism is critical for enabling distance-aware uncertainty.
Probabilistic Output Layers:
- PC-SNGP: Replaces the standard dense output layer with a Gaussian Process (GP) using an RBF kernel. The posterior variance of the GP increases as the Euclidean distance between the test input's latent representation and the training manifold increases.
- PC-SNER: Modifies the output layer to predict parameters of a Normal-Inverse-Gamma (NIG) distribution. This allows for the joint modeling of aleatoric (data noise) and epistemic (model uncertainty) uncertainties without sampling, utilizing a Student-t marginal distribution for predictions.
Dynamic Physics-Constrained Loss: To balance data fidelity with physical consistency, the authors introduce a dynamically weighted loss function. The weights for the data loss ( $L_{data}$ ) and physics loss ( $L_{phys}$ ) are adaptively determined based on the variability (standard deviation) of the input indicators and the physics-based degradation estimate.
Unified Physics-Based Degradation Model: A unified model is developed for Rolling-Element Bearings (REBs) that couples fatigue, abrasive wear, lubrication breakdown, and stochastic contamination. This model serves as the physical constraint ( $L_{phys}$ ) during training, ensuring the network adheres to known degradation dynamics.
Evaluation Metric: A Distance-Aware Coefficient (DAC) is introduced, based on the Pearson correlation coefficient, to quantitatively measure a model's sensitivity to distributional shifts.

Key Contributions

Development of two sampling-free frameworks (PC-SNGP and PC-SNER) that provide calibrated, distance-aware uncertainty estimates, reducing computational costs compared to sampling-based methods.
Introduction of a unified multicomponent physics-based degradation model incorporating fatigue, wear, debris, and lubrication dynamics.
Proposal of a dynamically weighted physics-constrained loss strategy to adaptively balance training objectives.
Definition of the DAC metric to quantify sensitivity to distributional shifts.
Comprehensive validation on three benchmark datasets (PRONOSTIA, XJTU-SY, HUST) demonstrating improved accuracy and robustness.

Experimental Results
The frameworks were evaluated on rolling-element bearing (REB) prognostics against baselines including MC-Dropout (PC-MC) and Deep Ensembles (PC-DE).

In-Distribution (XIND) Performance: On the PRONOSTIA dataset, PC-SNER (with regularization $\lambda=0.5$ ) achieved the lowest Mean Squared Error (MSE = 0.0007) and the highest DAC (0.4788), outperforming PC-MC and PC-DE in uncertainty calibration.
Out-of-Distribution (XOOD) Performance: On unseen bearings, PC-SNER maintained superior robustness (MSE = 0.0016, DAC = 0.4744) compared to baselines, which showed degraded performance or lack of distance awareness.
Adversarial Robustness: Under Fast Gradient Sign Method (FGSM) perturbations, both frameworks maintained stable predictive means and well-bound uncertainty bands, attributed to the stability provided by Spectral Normalization and the regularization effect of the physics-constrained loss.
Cross-Validation: The models demonstrated strong generalization on the XJTU-SY and HUST datasets, preserving distance awareness in zero-shot scenarios.
Ablation Studies: Removing physics guidance, dynamic weighting, or spectral normalization significantly degraded performance or eliminated distance awareness, confirming that all components are complementary and essential.

Significance
The paper claims that these frameworks address the critical limitation of existing probabilistic models: the inability to reliably increase uncertainty when inputs deviate from the training manifold. By integrating spectral normalization with physics-constrained learning, the proposed methods achieve a balance between data-driven accuracy and physical interpretability without the computational burden of sampling-based inference. The results suggest that these frameworks offer a robust solution for industrial prognostics, providing auditable performance and resilience against extreme perturbations and distributional shifts.