TL;DR
This paper introduces GOOFD, a novel framework that unifies multiple fault diagnosis tasks (process monitoring, fault classification, and out-of-distribution detection) into a single system. It achieves this by employing Internal Contrastive Learning (ICL) for robust feature extraction and Mahalanobis distance for effective outlier detection, demonstrating superior performance over single-task methods.
What Problem Does This Paper Address?
Industrial processes are becoming increasingly complex, demanding sophisticated fault diagnosis systems. Traditionally, different fault diagnosis tasks like detecting anomalies (process monitoring), identifying known faults (fault classification), and recognizing entirely new, unseen faults (open-set fault diagnosis or OSFD) are addressed by separate, specialized methods. This fragmented approach leads to inefficiencies, requires separate model training, and struggles to adapt to the dynamic nature of industrial environments. The paper aims to bridge this gap by proposing a unified framework that can handle these diverse tasks simultaneously, improving efficiency and diagnostic accuracy.
Methods
The core of the proposed solution lies in the Generalized Out-of-distribution Fault Diagnosis (GOOFD) framework and its associated ICL-OD (Internal Contrastive Learning for Outlier Detection) method.
- GOOFD Framework: This framework integrates three key tasks:
- Process Monitoring: Detecting deviations from normal operating conditions.
- Fault Classification: Identifying and categorizing known types of faults.
-
Open-Set Fault Diagnosis (OSFD): Detecting and flagging faults that were not part of the training data (unseen faults). The framework unifies these tasks by treating them as variations of an outlier detection problem, where the definition of an "outlier" shifts depending on the specific subtask.
-
ICL-OD Method: This method is the engine driving the GOOFD framework.
- Internal Contrastive Learning (ICL): Instead of relying on external data augmentation, ICL extracts features by learning to make sub-vectors within a single data sample similar while pushing dissimilar sub-vectors apart. This internal focus makes it particularly adept at learning robust representations for outlier detection. The paper uses two networks, F and G, to learn these feature maps.
- Mahalanobis Distance: After ICL extracts features and generates a score for each sample, the Mahalanobis distance is used to measure how far a sample's score is from the distribution of known (normal or known fault) data. This distance metric is chosen for its ability to account for correlations between variables, making it more effective than simpler distance measures for outlier detection.
- Unified Outlier Detection: For process monitoring, outliers are simply abnormal data points. For OSFD, outliers are the unknown faults. The ICL-OD method uses the Mahalanobis distance to identify these outliers based on a learned threshold derived from the training data.
Datasets
The proposed GOOFD framework and ICL-OD method were evaluated on two types of datasets:
- Simulated Benchmark Dataset: The Tennessee Eastman (TE) process dataset [27] was used. This dataset is a well-established benchmark for fault diagnosis research, simulating complex industrial processes with 21 distinct faults.
- Practical Process Datasets: The Multiphase Flow Facility (MFF) Dataset [28] was utilized. This dataset is derived from a real-world three-phase flow facility and includes normal operating conditions and six fault cases, tested under varying operational conditions.
Key Insights and Ideas
- Unification is Key: The most significant insight is that diverse fault diagnosis tasks can be effectively unified under a single framework by reframing them as variations of outlier detection. This challenges the traditional approach of developing separate solutions for each task.
- Internal Contrastive Learning for Anomaly Detection: ICL proves to be a powerful feature extractor for anomaly detection, especially when dealing with limited data or complex internal data structures. Its ability to learn representations from within a sample without external augmentation is a novel application in this domain.
- Mahalanobis Distance for Robust Outlier Scoring: The Mahalanobis distance is shown to be a superior metric for distinguishing between known and unknown classes compared to other distance measures. Its ability to handle correlated features makes it more sensitive to subtle deviations indicative of faults.
- Improved Discriminability of Prediction Scores: The GOOFD framework, particularly with its ICL-based feature extraction, generates prediction scores with greater disparity between known and unknown classes, leading to more accurate classifications and fewer misclassifications.
Key Results
- Process Monitoring: The GOOFD method consistently outperformed several baseline methods (PCA, KPCA, ICA, KICA, DPCA, SJCCA, OSA, DSVDD) in terms of Fault Detection Rate (FDR) on the TE process dataset. It also showed superior performance on the MFF dataset.
- Open-Set Fault Diagnosis (OSFD): On both the TE and MFF datasets, the GOOFD framework achieved significantly better F1-scores and AUROC values compared to various baseline methods (SoftMax, OpenMax, CenterLoss, EOW-SoftMax, GEN, PULSE).
- For the TE dataset, the method demonstrated a higher recognition rate for unknown faults while maintaining accuracy for known faults.
- In the MFF dataset, for unknown fault 2, the GOOFD method improved the F1-score by 12% compared to the ICL+ baseline.
- Threshold Selection: An ablation study on the TE dataset showed that an outlier threshold (θu) of 98% yielded the best performance for OSFD, indicating the importance of careful threshold tuning.
Limitations and Future Work
- Limitations:
- The paper acknowledges that "OT" (impossible to determine) cases occurred in process monitoring on the TE dataset, where indicator values exceeded thresholds before fault introduction.
- While the framework is generalized, the paper focuses on process monitoring, fault classification, and OSFD. Other diagnostic subtasks are not explicitly covered.
- Future Work:
- The authors intend to expand the GOOFD framework to encompass more diagnostic subtasks, further enhancing its versatility.
- A key future direction is to incorporate interpretability into the framework to improve transparency and user trust, ultimately contributing to safer industrial operations and increased machine reliability.
Summary updated