Adversarial Algorithm Unrolling Network for Interpretable Mechanical Anomaly Detection.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• High-accuracy mechanical anomaly detection algorithms, often neural networks, lack interpretability, hindering trust in their results.
• Black-box models present challenges in understanding their decision-making processes for mechanical systems.

Purpose of the Study:

• To propose an interpretable mechanical anomaly detection method using an adversarial algorithm unrolling network (AAU-Net).
• To enhance the credibility of anomaly detection results through interpretable network architecture and feature visualization.

Main Methods:

• Developed an adversarial algorithm unrolling network (AAU-Net), a generative adversarial network (GAN).
• Designed the generator using algorithm unrolling of a sparse coding model for vibration signal feature encoding/decoding.
• Implemented a multiscale feature visualization approach for post hoc interpretability.

Main Results:

• AAU-Net demonstrated a mechanism-driven and interpretable network architecture.
• Feature visualization confirmed that AAU-Net encodes meaningful signal features aligned with mechanical system dynamics.
• AAU-Net achieved superior overall anomaly detection performance compared to other methods.

Conclusions:

• AAU-Net offers a solution for interpretable mechanical anomaly detection, addressing the black-box nature of traditional methods.
• The network's ability to learn system dynamics and provide interpretable results enhances user trust and detection accuracy.