Fault Diagnosis of Hot Rolling Equipment Under Uncertain Conditions Based on Cloud Rough Model, Game Theory, and Improved GRA-TOPSIS

Bo Hu, Yongjun Zhang, Fengwei JingInternational Journal of Computational Intelligence Systems 2025 · DOI

This paper, published in the International Journal of Computational Intelligence Systems, presents a sophisticated approach to a high-stakes industrial problem: diagnosing equipment failure in the chaotic environment of a steel mill. By combining advanced uncertainty modeling with game theory, the authors provide a roadmap for more reliable predictive maintenance in complex manufacturing.

TL;DR

The researchers developed a hybrid fault diagnosis framework that simultaneously accounts for "fuzzy" sensor data and the subjective randomness of expert opinions. By integrating Cloud Rough Models with Game Theory-optimized weighting, the system achieved 90% diagnostic accuracy on real-world hot rolling equipment, significantly outperforming traditional decision-making methods.

What Problem Does This Paper Address?

In the production of hot-rolled strip steel, equipment stability is everything. A single failure in a "looper"—the device that maintains tension between rolling stands—can lead to "steel stacking" (a catastrophic pile-up) or strip narrowing, ruining the product.

The challenge is that industrial environments are inherently "noisy." Diagnosing a fault isn't just about reading a sensor; it involves Multi-Criteria Decision-Making (MCDM) under uncertainty. Traditional methods struggle with two specific types of uncertainty:

  1. Fuzziness: Vague definitions of equipment states (e.g., "moderately high" tension).
  2. Randomness: The stochastic fluctuations in sensor data and the natural variance between different experts' judgments.

Existing models usually handle one or the other, but rarely both, leading to diagnostic errors or "rank reversal" where the wrong fault is prioritized.

Methods

The authors propose a three-stage framework to bridge this gap:

  • Cloud Rough Model (CRN): This is the "uncertainty engine." It combines Cloud Models (which use entropy to describe the randomness and fuzziness of linguistic terms) with Rough Set Theory (which handles group consensus by defining upper and lower boundaries for expert opinions). This allows the system to quantify "interpersonal uncertainty"—the fact that five experts might mean five slightly different things when they say a fault risk is "High."
  • Game Theory Weighting: Determining which sensor indicators (like "Angular deviation" or "Falling time") matter most is difficult. Subjective weights (expert-driven) can be biased, while objective weights (data-driven, like the CRITIC method) can be fooled by noise. The authors treated these different weighting methods as "players" in a game, using game theory to find a Nash-like equilibrium that minimizes the conflict between human experience and raw data.
  • Improved GRA-TOPSIS: Finally, to rank the likely faults, they fused TOPSIS (which measures the Euclidean distance to an "ideal" state) with Grey Relational Analysis (GRA) (which looks at the similarity in the shape of data curves). This "dual-scale" measurement makes the ranking much more robust against outliers.

Datasets

The study is grounded in a massive real-world dataset from a domestic steel mill:

  • Operational Data: 19,306 coils of strip steel looper data.
  • Fault Cases: 4,687 abnormal cases were identified and used for the study.
  • Expert Input: Five experts with over 20 years of experience provided linguistic evaluations for 11 key influencing factors (e.g., stabilization time, hydraulic deviation).
  • Baselines: The model was compared against seven standard MCDM methods, including Fuzzy TOPSIS and Cloud GRA.

Key Insights and Ideas

  • The "Double-Whammy" of Uncertainty: The most important insight is that industrial diagnosis fails when it ignores the randomness within fuzzy intervals. The CRN model captures the "thickness" of a cloud of data, representing how much we should trust a specific judgment.
  • Weighting as Negotiation: By using game theory to balance weights, the authors reduced the average deviation of weights by 27.3% compared to using a single-method approach. It prevents a single noisy sensor or a single biased expert from hijacking the diagnosis.
  • Shape vs. Distance: In industrial signals, the trend (shape) of a sensor reading is often more indicative of a specific fault than the absolute value (distance). Fusing GRA and TOPSIS ensures the model doesn't get "distracted" by geometric anomalies in the data.

Key Results

  • 90% Diagnostic Accuracy (DA): The proposed method correctly identified the primary fault type in 90% of cases.
  • 100% Top-2 Hit Rate: The "true" fault was always ranked in the top two most likely candidates, providing a reliable shortlist for maintenance crews.
  • Superior Robustness: When the researchers added Gaussian noise to simulate real-world sensor interference, their model showed significantly lower "Average Rank Change" (0.32) compared to baselines like Fuzzy TOPSIS (0.85).
  • High Discreteness: The model produced a "high discreteness" in its results (27.5% average deviation), meaning it clearly separated the most likely fault from the "noise," making decision-making easier for on-site engineers.

Limitations and Future Work

The authors acknowledge a few areas for improvement:

  • Expert Pool Size: The study relied on only five experts. They suggest expanding this to a cross-disciplinary team of 15+ to further reduce human bias.
  • Missing Data: Industrial sites often have gaps in sensor logs. The authors plan to integrate Transformer-based neural networks to "repair" missing data features in the future.
  • Indicator Coupling: The current model treats indicators as somewhat independent. Future work will explore the "coupling effect"—how different factors interact and amplify each other during a failure.

Summary updated

Abstract

The hot-rolled strip steel is a critical process in steel production, where equipment stability directly impacts product quality. Traditional evaluation methods struggle to balance uncertainties between influencing factors and expert judgments. This study proposes a novel fault evaluation framework: first, a cloud rough model is adopted to handle multi-source uncertainties in expert evaluations; second, game theory is used to optimize the combined weights of subjective and objective criteria; finally, an improved Grey Relational Analysis-Technique for Order Preference by Similarity to Ideal Solution (GRA-TOPSIS) method constructs a dual-scale measurement model by fusing grey relational degree and Euclidean distance, effectively reducing the influence of geometric sequence similarity and distance on fault prioritization and enhancing the robustness of ranking results under complex working conditions. A case study on hot rolling loopers demonstrates a diagnostic accuracy of 90%. Comparative analyses with seven typical multicriteria decision-making methods indicate the proposed method's feasibility and effectiveness in practical applications, offering a novel reference for equipment status evaluation.
BibTeX
@article{hu2025fault,
  author = {Bo Hu and Yongjun Zhang and Fengwei Jing},
  title = {{Fault Diagnosis of Hot Rolling Equipment Under Uncertain Conditions Based on Cloud Rough Model, Game Theory, and Improved GRA-TOPSIS}},
  journal = {International Journal of Computational Intelligence Systems},
  year = {2025},
  doi = {10.1007/s44196-025-00945-x},
}
Figure