💻 computer science

Trust-Aware Predictive Emissions Monitoring for Gas Turbine Fleets with Limited Labelled Data

This paper proposes a trust-aware probabilistic framework that combines multi-head recurrent modeling, uncertainty quantification, and feature-space analysis to generate interpretable reliability scores for gas turbine NOx predictions, enabling effective fleet-wide emissions monitoring even when labeled data is scarce.

Original authors: Rebecca Potts, Aiden Durrant, Rick Hackney, Georgios Leontidis

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rebecca Potts, Aiden Durrant, Rick Hackney, Georgios Leontidis

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 a massive fleet of 57 gas turbines, like a fleet of 57 different delivery trucks. The company wants to know exactly how much pollution (specifically NOx) each truck is spewing out. However, they only have a high-tech, expensive pollution sensor installed on one of those trucks. The other 56 trucks are "blind" to their own emissions; they have no labels telling us what they are actually putting into the air.

The goal of this paper is to build a "smart guesser" (a machine learning model) that can predict the pollution for all 56 blind trucks based on the data from the one truck that has a sensor. But here's the catch: since the trucks might drive differently or be in different weather, the smart guesser might make mistakes. If we don't know when it's guessing wrong, we can't trust the data.

So, the researchers didn't just build a predictor; they built a "Trustworthy Predictor." Think of it like a weather forecaster who doesn't just say "It will rain," but also says, "It will rain, and I'm 90% sure," or "It might rain, but I'm only 50% sure because the conditions are weird."

Here is how their system works, broken down into simple concepts:

1. The "Multi-Head" Brain

The computer model is like a student taking a test. Instead of just answering the main question ("How much pollution?"), it is forced to answer other questions too:

The Main Question: Predict the pollution level.
The Side Questions: Predict what the engine's temperature or speed will be in the next second.
The Confidence Check: The model has to rate its own confidence on a scale of 0 to 1 for every single guess.

By forcing the model to predict other things (like temperature) that it can measure on the blind trucks, it learns how the engines generally behave. This helps it understand the "rules of the road" even when it hasn't seen that specific truck before.

2. The "Panel of Experts" (Ensemble)

To make sure the model isn't just guessing blindly, they train multiple versions of the model (like a panel of 5 different experts).

If all 5 experts agree on the pollution number, the system says, "We are very confident."
If the experts are arguing with each other (one says 10, another says 50), the system knows, "We are confused. This is a risky prediction."

This disagreement is a key signal that tells the system, "Hey, this situation is unfamiliar; be careful."

3. The "Trust Score" Report Card

This is the most important part. For every single prediction the system makes, it generates a Trust Score (from 0 to 100) and a short explanation.

Imagine you get a report card that says:

Trust Score: 65/100 (Medium)
Why? "The engine is running in a way we haven't seen before, and our experts are disagreeing."

If the score is high (e.g., 90), the company can trust the number. If the score is low (e.g., 30), the company knows, "Don't use this number for official reporting; it's too risky."

4. The "Unfamiliar Territory" Detector

The system also checks if the truck is doing something totally new. It uses a "distance meter" (called Mahalanobis distance).

Imagine the "safe zone" is a circle drawn around the driving habits of the one truck with the sensor.
If a blind truck is driving inside that circle, the system is comfortable.
If the blind truck drives way outside that circle (e.g., running at a speed the sensor-truck never reached), the system flags it as "Out of Distribution." It automatically lowers the trust score, saying, "I've never seen this before, so my guess might be wrong."

The Results: Does it Work?

The researchers tested this on the one truck they had data for (the "labelled" truck). They found a clear pattern:

When the model said, "I am very confident," the guess was usually very accurate.
When the model said, "I am not sure," the guess was often wrong.

They proved that by filtering out the "low confidence" guesses, they could make the remaining data much more accurate. For example, if they only looked at the top 10% of "most confident" predictions, the error rate dropped significantly.

The Bottom Line

The paper concludes that for a fleet of gas turbines, you can't just rely on a single number for pollution. You need a Trust Framework.

This framework acts like a "safety net." It tells the operators:

Here is the pollution number.
Here is how much you can trust it.
Here is why you should or shouldn't trust it (e.g., "The engine is running weirdly" or "The experts disagree").

This allows companies to use these predictions safely, knowing exactly when to be cautious, even when they don't have sensors on every single machine.

Technical Summary: Trust-Aware Predictive Emissions Monitoring for Gas Turbine Fleets

Problem Statement
Predictive Emissions Monitoring Systems (PEMS) offer a data-driven alternative to direct emissions measurement (CEMS) for gas turbines. However, deploying PEMS across industrial fleets faces a critical challenge: emissions labels (ground-truth measurements) are typically available for only a small subset of assets, often a single turbine, while the majority of the fleet operates without such labels. Standard machine learning approaches evaluate performance on labelled test sets but fail to assess the reliability of individual predictions when applied to unlabelled turbines operating under different conditions. Consequently, a model may achieve high accuracy on the labelled turbine but produce unreliable predictions on unlabelled units due to distributional shifts, without any mechanism to flag these risks.

Methodology
The authors propose a trust-aware probabilistic framework designed to generate calibrated, per-sample trust scores for NOx predictions on unlabelled turbines. The approach integrates a multi-head recurrent neural network with a suite of reliability diagnostics.

Prediction Model:
- The core architecture is a multi-head probabilistic LSTM. It processes input sequences (64 time steps) to predict the next-step NOx emissions.
- To handle fleet heterogeneity, a learnable turbine-specific embedding is concatenated with the latent LSTM representation.
- The model employs a multi-task learning strategy:
  - Emissions Head: Predicts the mean and variance of the target NOx.
  - Feature Head: Predicts the mean and variance of auxiliary input features for the next time step. This auxiliary task encourages the model to learn general system dynamics even when emissions labels are absent.
  - Confidence Heads: Separate heads predict scalar confidence values ( $c \in [0, 1]$ ) for both emissions and feature predictions. These are trained using a soft target derived from the absolute prediction error ( $c^* = \exp(-e/\tau)$ ), encouraging the network to output high confidence only when errors are low.
Uncertainty Quantification:
- An ensemble of $M$ independently initialized models is trained.
- Aleatoric uncertainty (data noise) is estimated as the mean of individual model variances.
- Epistemic uncertainty (model uncertainty) is estimated as the variance of the ensemble's mean predictions.
- Total uncertainty is the sum of these components.
Trust Score Calibration:
- Four reliability signals are extracted for each sample: Mahalanobis distance ( $D^2$ ) from the training distribution, predictive uncertainty ( $\sigma$ ), epistemic uncertainty, and learned confidence.
- An XGBoost regressor is trained on the labelled validation set to map these signals to the observed absolute prediction error.
- The output is a continuous Trust Score (0–100), which is calibrated to reflect the likelihood of low prediction error.
- Diagnostic Explanations: The system includes a range analysis to detect features operating outside the labelled training bounds. These signals are consolidated into textual explanations (e.g., "Moderate model disagreement," "Low emissions confidence") to aid human interpretation.
Training Procedure:
- Stage 1 (Unlabelled Pretraining): The model is trained on the large unlabelled dataset using only feature prediction and feature confidence objectives to learn general dynamics.
- Stage 2 (Labelled Training): The model is fine-tuned on the single labelled turbine using the full loss function, optimizing emissions prediction, feature prediction, and confidence calibration simultaneously.

Key Contributions

Trust-Aware Framework: A novel system that moves beyond aggregate accuracy metrics to provide interpretable, per-sample reliability assessments for fleet-level deployment where ground truth is unavailable.
Multi-Signal Reliability Assessment: The framework demonstrates that no single signal (e.g., distance or confidence alone) is sufficient. It combines learned confidence, epistemic/aleatoric uncertainty, feature-space distance, and operating-range diagnostics to form a robust trust metric.
Confidence Calibration: The study validates that learned confidence is meaningfully correlated with prediction error, allowing for confidence-based filtering to significantly improve accuracy on high-confidence subsets.
Diagnostic Reporting: The generation of concise, signal-driven textual explanations for trust scores, enabling operators to understand why a prediction is flagged as low-trust.

Results

Predictive Performance: On the labelled turbine, the probabilistic LSTM achieved a Mean Absolute Error (MAE) of 0.202, comparable to an XGBoost baseline (0.313).
Uncertainty Behavior: The model appropriately increased predictive uncertainty for unlabelled data and out-of-distribution (OOD) samples. The Normalized Mean Prediction Interval Width (NMPIW) increased from 1.01 (labelled ID) to 10.75 (unlabelled OOD).
Confidence-Error Correlation: A strong inverse relationship was observed between confidence and absolute error (Spearman correlation: -0.494).
Filtering Efficacy: Confidence-based filtering proved highly effective. By restricting evaluation to the top 10% of highest-confidence predictions, the MAE dropped from 0.202 (full coverage) to 0.070.
Trust Score Validation: On labelled data, lower trust levels corresponded to higher prediction errors. On unlabelled data, lower trust scores were associated with reduced confidence, increased uncertainty, higher feature prediction errors, and more frequent deviations from the training operating range.

Significance and Claims
The paper claims that accurate point estimates are insufficient for industrial fleet deployment; trustworthiness must be treated as a multi-signal problem. The proposed framework provides a mechanism to identify when predictions on unlabelled turbines are likely reliable and when they should be treated with caution. By calibrating trust scores on limited labelled data, the system supports the transparent deployment of PEMS across assets where direct emissions monitoring is unavailable or prohibitively expensive. The authors emphasize that while the framework is model-agnostic, its current evaluation is specific to a probabilistic LSTM backbone, and future work requires validation on additional labelled turbines to further verify generalization.