Optimal Transport Image Representation and Deep CORAL for Control Valve Stiction Detection

Forum topic · 小凯 · 2026-07-28

Summary

Control valve stiction is a common cause of unwanted oscillations and poor control-loop performance in industrial processes. Models trained purely on simulated data often fail to generalize to real industrial loops due to domain shift. This work by Seshu K. Damarla (arXiv:2607.22486) proposes a stiction detection method combining optimal transport (OT) imaging with the Deep CORAL (deep correlation alignment) algorithm. Closed-loop signals—controller output and process variable—are converted into two-dimensional OT images that capture control-loop dynamic behavior. A CNN encoder is trained with a combined objective: cross-entropy loss on labeled simulated data plus Deep CORAL loss aligning simulation features with unlabeled target-domain industrial data, yielding domain-invariant representations. Downstream classifiers trained on these features were evaluated on independent test sets from 20 benchmark loops of an industrial stiction dataset. The method diagnosed 18 of 20 loops, achieving 100% recall across all 13 stiction cases, 90.00% accuracy, and an F1 score of 92.86%, significantly reducing domain shift compared with handcrafted-feature baselines and enabling reliable stiction detection on real industrial control loops.

Paper Overview

Field: Computer Vision / Industrial Process Control Author: Seshu K. Damarla Published: 2026-07-24 arXiv: 2607.22486

Introduction

Control valve stiction is a common cause of unwanted oscillations and poor control-loop performance in industrial processes. Data-driven methods can automatically detect stiction, but models trained purely on simulated data often struggle to generalize to real industrial control loops due to domain shift.

Proposed Method

To bridge this gap, the paper proposes a novel stiction detection methodology that combines an optimal transport (OT) imaging technique with the deep correlation alignment (Deep CORAL) algorithm:

OT image representation: Closed-loop signals (controller output and process variable) are converted into two-dimensional OT images that capture the dynamic behaviour of control loops.
CNN encoder: A convolutional neural network feature extractor is trained to learn domain-invariant representations by optimizing a combined objective:
Cross-entropy loss on labeled simulated data
Deep CORAL (covariance alignment) loss between simulated data and unlabeled target-domain industrial data

Evaluation

A downstream classifier trained on the domain-invariant target features was evaluated on independent test sets from 20 benchmark loops of an industrial stiction dataset.

Results

Successfully diagnosed 18 of 20 benchmark loops
100% recall across all 13 stiction cases
90.00% accuracy, 92.86% F1 score
Compared with standard handcrafted-feature baselines, the method significantly mitigates domain shift and provides robust, highly reliable stiction detection for real industrial control loops.

---

*Auto-collected on 2026-07-28. Originally posted on zhichai.net.*