💻 computer science

Optimal Transport Image Representation and Deep Covariance Alignment (CORAL) for Control Valve Stiction Detection

This paper proposes a novel stiction detection methodology that combines optimal transport imaging with deep covariance alignment to bridge the domain gap between simulated and real industrial data, achieving high accuracy and recall in diagnosing control valve stiction across benchmark loops.

Original authors: Seshu K. Damarla

Published 2026-07-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Seshu K. Damarla

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

The Invisible Glitch in the Machine

Imagine a factory floor as a giant, humming orchestra. Every instrument, from the massive pumps to the tiny sensors, plays a specific note to keep the production line running smoothly. But sometimes, a valve—a crucial part of the machinery that controls the flow of liquids or gases—gets "sticky." It's like a door hinge that has been painted over with dried glue; it refuses to move until you push hard enough, and then it suddenly jerks open. This "stiction" causes the whole orchestra to start playing a rhythmic, unwanted wobble instead of a steady tune. These wobbles can ruin product quality, waste energy, and wear out expensive equipment.

To fix this, engineers usually need to listen to the machine's heartbeat: the signals telling the valve what to do (the "Controller Output") and the signals telling them what the machine is actually doing (the "Process Variable"). The challenge is that these signals look different depending on where you are. A computer simulation of a sticky valve looks like a clean, perfect drawing, while a real factory valve looks like a messy, noisy sketch. Trying to teach a computer to spot the problem using only the perfect drawings often fails when it meets the messy reality. This paper tackles that exact problem: how to teach a computer to recognize a sticky valve in the real world, even if it was only trained on perfect simulations.

Turning Time into Pictures

The researchers, led by Seshu K. Damarla, came up with a clever way to bridge the gap between the "perfect world" of simulations and the "messy world" of real factories. Instead of feeding raw numbers into a computer, they decided to turn the time-based signals into pictures. Think of it like taking a long, winding road and folding it up until it looks like a topographic map. They used a mathematical technique called Optimal Transport (OT) to create these images.

Imagine you have two piles of sand, and you want to move the sand from one pile to match the shape of the other. The "cost" of moving the sand tells you how different the piles are. The researchers used this idea to compare chunks of the valve's signals. If the valve is sticky, the "sand" moves in a very specific, jagged pattern. If it's working fine, the sand flows smoothly. By turning these patterns into 50x50 pixel images, they gave the computer a visual puzzle to solve. A picture of a sticky valve looks like a distinct, swirling storm cloud, while a healthy valve looks like a calm, organized grid.

The "Deep CORAL" Magic Trick

Here is where the magic happens. The team trained a smart computer brain (a Convolutional Neural Network, or CNN) to recognize these storm clouds. But they had a problem: the computer was only trained on the "perfect" simulation images. When they showed it the "messy" real-world images, the computer got confused because the lighting and texture were different.

To fix this, they used a technique called Deep CORAL (Deep Correlation Alignment). Think of this like a translator who doesn't just translate words, but also teaches two people to think in the same way. The computer looked at the features of the simulation images and the real factory images side-by-side. It then adjusted its internal settings to make the "statistical shape" of the simulation data match the real data. It's as if the computer learned to ignore the noise and focus on the core shape of the problem, effectively saying, "It doesn't matter if the picture is from a simulation or a real factory; if the storm cloud looks like this, it's a sticky valve."

The Results: A Near-Perfect Score

The team tested their new method on a set of 20 real industrial control loops that the computer had never seen before. These loops came from various industries, including chemicals, mining, and paper manufacturing.

The results were impressive. The new method correctly identified 18 out of the 20 loops. Most importantly, it caught all 13 of the actual sticky valve cases, meaning it had a 100% recall rate. It didn't miss a single real problem. In total, it achieved an accuracy of 90.00% and an F1-score (a measure of overall balance between accuracy and completeness) of 92.86%.

For comparison, they also tried an older method that relied on hand-crafted features (like manually measuring the height and width of the signal waves). That older method only got 13 out of 20 correct, with an accuracy of 65.00% and an F1-score of 72.00%. The new method was significantly better, especially at avoiding "false negatives" (missing a real problem).

What Went Wrong (and What's Next)

The method wasn't perfect, though. It made two mistakes: it flagged two non-sticky paper-mill loops as if they were sticky. The authors suggest this happened because those specific loops had wobbles that looked very similar to sticky valves, tricking the computer. However, the paper emphasizes that the method successfully ruled out the idea that simple, hand-crafted features are enough for the job.

The study concludes that combining these "transport images" with the "Deep CORAL" alignment creates a highly reliable tool for spotting sticky valves. It proves that you can train a computer on simulations and still make it work in the real world, provided you teach it to align the two worlds correctly. While the paper notes that future work will focus on distinguishing those tricky "look-alike" wobbles in paper mills, the current findings suggest this approach is a major step forward in keeping industrial machines humming smoothly.

Technical Summary: Optimal Transport Image Representation and Deep Covariance Alignment for Control Valve Stiction Detection

Problem Statement
Control valve stiction is a prevalent issue in industrial processes, causing sustained oscillations that degrade product quality, increase energy consumption, and accelerate equipment wear. While invasive testing can identify stiction, it requires taking valves offline. Consequently, non-invasive methods utilizing routine controller output (OP) and process variable (PV) data are preferred. However, existing data-driven models trained on simulated data often fail to generalize to real-world industrial loops due to "domain shift"—the statistical discrepancy between simulated and actual industrial data. Furthermore, traditional time-series analysis often relies on pointwise amplitudes or specific frequency bands, which may miss complex distributional changes associated with stiction.

Methodology
The proposed framework addresses these challenges by combining Optimal Transport (OT) imaging with Deep Correlation Alignment (Deep CORAL) for domain adaptation. The methodology proceeds in three main stages:

Optimal Transport Image Generation:
Instead of analyzing raw time-series sequences directly, the method converts OP and PV signals into two-dimensional images. This process involves:
- Normalizing signals and computing first-order derivatives.
- Constructing a feature vector comprising PV, OP, the ratio of derivatives ( $dOP/(dPV + \epsilon)$ ), and control error ($OP - PV$).
- Dividing these feature vectors into overlapping windows.
- Calculating the Sliced-Wasserstein distance between window pairs by projecting feature vectors onto random unit vectors and computing 1D Wasserstein distances.
- Generating a normalized, smoothed distance matrix interpreted as an image. This representation captures the complete distributional dynamics (shape, spread, location) of the control loop rather than just pointwise values.
Deep Domain Adaptation (Deep CORAL):
To bridge the gap between the source domain (simulated data) and the target domain (unlabeled industrial data), a Convolutional Neural Network (CNN) encoder is employed.
- Architecture: The encoder consists of three convolutional blocks (convolution, batch normalization, ReLU, max pooling), followed by global average pooling and a fully connected layer.
- Training Objective: The encoder is trained using a combined loss function:
  - Cross-Entropy Loss ( $\mathcal{L}_{CE}$ ): Computed on labeled simulation data to ensure the features distinguish between stiction and non-stiction.
  - Deep CORAL Loss ( $\mathcal{L}_{CORAL}$ ): Computed between the feature covariance matrices of the source (simulation) and target (industrial) domains. This loss minimizes the difference in second-order statistics, forcing the encoder to learn domain-invariant representations.
- A dynamic weighting strategy is used for the CORAL loss coefficient ( $\lambda$ ), which increases over training epochs.
Downstream Classification:
After training the shared encoder, the temporary classification head is removed. The encoder extracts domain-adapted features from the industrial training data. These features are standardized and fed into a logistic regression classifier, which is trained to predict stiction presence. The decision threshold is determined via cross-validation on the industrial training set to prevent data leakage.

Key Contributions

Novel Representation: This work introduces the first application of Optimal Transport imaging to control loop performance monitoring. Unlike methods relying on recurrence plots or Markov Transition Fields, OT imaging compares complete distributions of signal windows, making it robust to alignment issues and amplitude discretization.
Domain Adaptation Framework: The study proposes a specific pipeline that aligns simulated and industrial data distributions using Deep CORAL, effectively mitigating the domain shift that plagues purely simulation-trained models.
Benchmark Validation: The method is rigorously evaluated against a baseline using hand-crafted features (HOG, LBP, GLCM, and statistical features) on an independent set of 20 industrial loops from the Industrial Stiction Data Benchmark (ISDB).

Results
The proposed methodology was tested on 20 independent industrial control loops (13 stiction, 7 non-stiction).

Performance: The proposed method achieved 90.00% accuracy, 86.67% precision, 100% recall, and an F1-score of 92.86%.
Comparison: In contrast, the baseline hand-crafted feature method achieved 65.00% accuracy and a 72.00% F1-score.
Specific Outcomes: The proposed method correctly diagnosed all 13 stiction cases (zero false negatives), whereas the baseline missed four stiction cases. The proposed method misclassified two non-stiction pulp-and-paper loops as stiction (false positives), likely due to non-stiction oscillations mimicking stiction patterns.
Improvement: The proposed approach improved accuracy by 25 percentage points and the F1-score by approximately 20.86 percentage points over the baseline.

Significance and Claims
The paper claims that the integration of OT imaging with Deep CORAL provides a robust, highly reliable tool for detecting control valve stiction in real-world industrial settings. By learning domain-invariant representations, the framework successfully transfers knowledge from simulated environments to unlabeled industrial data without requiring invasive testing. The authors emphasize that the method's ability to achieve 100% recall (no false negatives) is particularly critical for industrial safety and performance monitoring. While the method shows high reliability, the authors modestly note that future work is required to better discriminate between stiction and specific non-stiction oscillations (such as those found in pulp-and-paper processes) that share similar dynamic patterns.