TLDR: FISHER is a new foundation model designed to analyze diverse industrial signals (sound, vibration, etc.) for anomaly detection and fault diagnosis. It addresses the “M5 problem” (multi-modal, multi-sampling-rate, multi-scale, multi-task, mini-fault) by processing signals in sub-bands and using a teacher-student self-supervised learning approach. Evaluated on the new RMIS benchmark, FISHER outperforms existing models, especially in fault diagnosis, demonstrating superior performance and efficient scaling, making it highly suitable for industrial health management.
In the world of modern manufacturing, where machines are constantly monitored by a multitude of sensors, effectively analyzing the vast amounts of industrial signals to detect anomalies and diagnose faults is a critical challenge. These signals, ranging from sound and vibration to voltage and temperature, are incredibly diverse. Researchers refer to this complex challenge as the “M5 problem,” encompassing multi-modal, multi-sampling-rate, multi-scale, multi-task, and mini-fault characteristics.
Traditionally, addressing the M5 problem has involved using specialized models for each specific sub-problem, like a model just for sound-based anomaly detection or another for vibration-based fault diagnosis. While these specialized models perform well in their narrow domains, they often fail to leverage the combined insights from different types of signals and struggle to scale up efficiently. This approach also adds significant burden during development and deployment, as each unique issue requires its own dedicated model.
However, a new perspective suggests that despite their apparent differences, these industrial signals share fundamental similarities. For instance, various signals might describe the same mechanical event, sound and vibration are closely related forms of observation, and many signals can be analyzed using similar spectral methods. Recognizing these intrinsic connections, a team of researchers including Pingyi Fan, Anbai Jiang, and Jia Liu, has proposed a groundbreaking solution: FISHER, a Foundation model for multi-modal Industrial Signal compreHEnsive Representation.
FISHER is designed to overcome the limitations of previous approaches by modeling diverse industrial signals in a unified manner. A key innovation of FISHER is its ability to handle arbitrary sampling rates, which is crucial given the varied rates at which sensors collect data. It achieves this by treating the increment of sampling rate as the concatenation of sub-band information. Specifically, FISHER converts raw signals into Short-Time Fourier Transform (STFT) spectrograms, then splits these into smaller “sub-bands” for individual processing. This allows the model to adaptively utilize the full signal bandwidth, preserving crucial information that might be lost if signals were downsampled.
The model is pre-trained using a teacher-student self-supervised learning framework. In this setup, a “student” model learns by mimicking the representations generated by a more stable “teacher” model, which is an exponentially moving average of the student. This method helps FISHER learn robust and comprehensive representations of industrial signals without needing extensive manual labeling, which is particularly beneficial given the scarcity of fault data (the “mini-fault” aspect of the M5 problem).
To rigorously evaluate FISHER’s capabilities, the researchers also developed a new benchmark called RMIS (Representation of M5 Industrial Signals). The RMIS benchmark includes a wide array of datasets covering different modalities and supports two primary health management tasks: anomaly detection (identifying unusual signals without prior examples of anomalies) and fault diagnosis (pinpointing the specific type of fault when labeled data is available). What’s remarkable is that FISHER is evaluated on this benchmark without any fine-tuning on downstream datasets, instead using a simple K-nearest neighbor (KNN) approach for inference. This demonstrates the model’s inherent versatility and generalization ability.
The results are compelling. FISHER consistently outperforms multiple top self-supervised learning models on the RMIS benchmark. While it performs very well in anomaly detection, it truly excels in fault diagnosis tasks, significantly outperforming all baselines. Even the smallest version of FISHER, FISHER-tiny, with only 5.5 million parameters, surpasses larger baseline models. This indicates that FISHER not only achieves superior performance but also scales much more efficiently, meaning it can deliver strong results with a smaller model size, making it more practical for real-world industrial deployment.
Also Read:
- Advancing Industrial Fault Diagnosis with Privacy-Preserving Semi-Supervised Federated Learning
- Time-RA: Enhancing Anomaly Detection with AI-Powered Explanations
The researchers believe FISHER’s success stems from its unique sub-band modeling approach, which allows it to effectively utilize the full signal bandwidth. They also highlight that scaling up data volume with more unique and non-duplicated signals, along with adaptive test-time scaling, are promising avenues for future research in developing even more powerful foundation models for industrial signals. For those interested in exploring this innovative model further, the full research paper is available here: FISHER Research Paper.
