Internal Contrastive Learning for Generalized Out-of-distribution Fault Diagnosis (GOOFD) Framework

Xingyue Wang, Hanrong Zhang, Ke Ma, Shuting Tao, Peng Peng, Hongwei WangarXiv 2023 · arXiv · OpenReview

Abstract

Fault diagnosis is essential in industrial processes for monitoring the conditions of important machines. With the ever-increasing complexity of working conditions and demand for safety during production and operation, different diagnosis methods are required, and more importantly, an integrated fault diagnosis system that can cope with multiple tasks is highly desired. However, the diagnosis subtasks are often studied separately, and the currently available methods still need improvement for such a generalized system. To address this issue, we propose the Generalized Out-of-distribution Fault Diagnosis (GOOFD) framework to integrate diagnosis subtasks, such as fault detection, fault classification, and novel fault diagnosis. Additionally, a unified fault diagnosis method based on internal contrastive learning is put forward to underpin the proposed generalized framework. The method extracts features utilizing the internal contrastive learning technique and then recognizes the outliers based on the Mahalanobis distance. Experiments are conducted on a simulated benchmark dataset as well as two practical process datasets to evaluate the proposed framework. As demonstrated in the experiments, the proposed method achieves better performance compared with several existing techniques and thus verifies the effectiveness of the proposed framework.

Introduction

In modern industrial environments, fault diagnosis is critical for maintaining operational safety and economic efficiency. However, current diagnostic practices are often fragmented, utilizing distinct methodologies for process monitoring (detecting anomalies) and fault classification (identifying specific known failures). This separation becomes problematic when encountering "Open-Set" scenarios—situations where a system faces a fault type it has never seen before. Standard deep learning models frequently misclassify these novel faults into known categories with high confidence. Furthermore, training and maintaining separate models for every subtask is computationally expensive. The proposed Generalized Out-of-distribution Fault Diagnosis (GOOFD) framework addresses these challenges by unifying diagnostic tasks under a single "out-of-distribution" (OOD) detection paradigm.

Methods

The core of the framework is the Internal Contrastive Learning for Outlier Detection (ICL-OD) method, which rests on two pillars:

  1. Internal Contrastive Learning (ICL): Unlike standard contrastive learning that compares different samples, ICL analyzes the internal structure of a single data vector. By slicing sensor data into sub-vectors, the model learns to recognize that these components belong together (positive pairs) while distinguishing them from other dimensions. This self-supervised approach allows the feature extractor to learn the intricate, intrinsic correlations of "normal" operations without requiring extensive manual labels.
  2. Mahalanobis Distance: To identify anomalies, the system calculates the Mahalanobis distance of extracted features. This metric is significantly more effective for industrial sensor data than standard Euclidean distance because it accounts for the natural correlations between different variables, such as temperature and pressure.
The GOOFD framework and ICL-OD method architecture.

As shown in the framework architecture, the system generates boundaries for outliers. For process monitoring, anything deviating from the "normal" distribution is flagged. For open-set diagnosis, data that does not align with a "known fault" profile is categorized as a new or unknown fault.

Key Results

The GOOFD framework was validated using the Tennessee Eastman (TE) process—a complex chemical plant simulation—and the real-world Multiphase Flow Facility (MFF) dataset.

In the TE process experiments, the proposed method achieved a dominant average Fault Detection Rate (FDR) of 84.55%. This significantly outperforms traditional baselines such as PCA (60.50%) and Kernel PCA (74.30%). The method proved particularly adept at identifying "sneaky" faults (e.g., Faults 3, 9, and 15) that traditional methods often fail to detect.

Fault Detection Rate (FDR) comparison on the Tennessee Eastman process.

On the real-world MFF dataset, the method achieved an AUROC of 1.0 in several unseen fault scenarios, indicating perfect separation between known and unknown data. A critical insight from the study is that ICL generates more "extreme" prediction scores for unknown classes compared to methods like SoftMax or OpenMax, creating a wider gap that makes threshold setting more reliable.

Conclusion

The GOOFD framework demonstrates that process monitoring and open-set recognition are essentially the same mathematical problem: detecting when data drifts away from a known distribution. By leveraging the internal structure of data through ICL, the system builds a robust defense against both known and novel industrial faults. Future work will focus on enhancing the interpretability of these models—moving from simple detection to explaining why a specific sensor behavior indicates a fault—to further improve industrial safety and efficiency.

Summary updated

Abstract

Fault diagnosis is essential in industrial processes for monitoring the conditions of important machines. With the ever-increasing complexity of working conditions and demand for safety during production and operation, different diagnosis methods are required, and more importantly, an integrated fault diagnosis system that can cope with multiple tasks is highly desired. However, the diagnosis subtasks are often studied separately, and the currently available methods still need improvement for such a generalized system. To address this issue, we propose the Generalized Out-of-distribution Fault Diagnosis (GOOFD) framework to integrate diagnosis subtasks, such as fault detection, fault classification, and novel fault diagnosis. Additionally, a unified fault diagnosis method based on internal contrastive learning is put forward to underpin the proposed generalized framework. The method extracts features utilizing the internal contrastive learning technique and then recognizes the outliers based on the Mahalanobis distance. Experiments are conducted on a simulated benchmark dataset as well as two practical process datasets to evaluate the proposed framework. As demonstrated in the experiments, the proposed method achieves better performance compared with several existing techniques and thus verifies the effectiveness of the proposed framework.
BibTeX
@misc{wang2023internal,
  author = {Xingyue Wang and Hanrong Zhang and Ke Ma and Shuting Tao and Peng Peng and Hongwei Wang},
  title = {{Internal Contrastive Learning for Generalized Out-of-distribution Fault Diagnosis (GOOFD) Framework}},
  year = {2023},
  eprint = {2306.15266},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.15266},
}
Figure