Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpreting What Typical Fault Signals Look Like via Prototype-matching

Qian Chen, Xingjian Dong, Zhike Peng

TL;DR

The paper addresses the opacity of neural fault diagnosis models by introducing a prototype-matching network (PMN) that fuses autoencoder-based feature learning with a prototype-matching classifier. PMN provides three interpretability paths: clear classification logic, sample-level fault prototypes, and frequency-attribution of prototype similarity, enabled by a loss that jointly optimizes accuracy and interpretability. Empirical results on planetary gearboxes and bevel gear domain generalization show PMN achieves competitive diagnostic accuracy and superior representation learning (lower $R_{rps}$), while reconstructing typical fault signals and revealing key fault-related frequencies. The approach advances AI-for-Science by offering interpretable, denoising-capable fault signals and a gray-box framework for understanding neural classification in IFD.

Abstract

Neural networks, with powerful nonlinear mapping and classification capabilities, are widely applied in mechanical fault diagnosis to ensure safety. However, being typical black-box models, their application is limited in high-reliability-required scenarios. To understand the classification logic and explain what typical fault signals look like, the prototype matching network (PMN) is proposed by combining the human-inherent prototype-matching with autoencoder (AE). The PMN matches AE-extracted feature with each prototype and selects the most similar prototype as the prediction result. It has three interpreting paths on classification logic, fault prototypes, and matching contributions. Conventional diagnosis and domain generalization experiments demonstrate its competitive diagnostic performance and distinguished advantages in representation learning. Besides, the learned typical fault signals (i.e., sample-level prototypes) showcase the ability for denoising and extracting subtle key features that experts find challenging to capture. This ability broadens human understanding and provides a promising solution from interpretability research to AI-for-Science.

Interpreting What Typical Fault Signals Look Like via Prototype-matching

TL;DR

The paper addresses the opacity of neural fault diagnosis models by introducing a prototype-matching network (PMN) that fuses autoencoder-based feature learning with a prototype-matching classifier. PMN provides three interpretability paths: clear classification logic, sample-level fault prototypes, and frequency-attribution of prototype similarity, enabled by a loss that jointly optimizes accuracy and interpretability. Empirical results on planetary gearboxes and bevel gear domain generalization show PMN achieves competitive diagnostic accuracy and superior representation learning (lower ), while reconstructing typical fault signals and revealing key fault-related frequencies. The approach advances AI-for-Science by offering interpretable, denoising-capable fault signals and a gray-box framework for understanding neural classification in IFD.

Abstract

Neural networks, with powerful nonlinear mapping and classification capabilities, are widely applied in mechanical fault diagnosis to ensure safety. However, being typical black-box models, their application is limited in high-reliability-required scenarios. To understand the classification logic and explain what typical fault signals look like, the prototype matching network (PMN) is proposed by combining the human-inherent prototype-matching with autoencoder (AE). The PMN matches AE-extracted feature with each prototype and selects the most similar prototype as the prediction result. It has three interpreting paths on classification logic, fault prototypes, and matching contributions. Conventional diagnosis and domain generalization experiments demonstrate its competitive diagnostic performance and distinguished advantages in representation learning. Besides, the learned typical fault signals (i.e., sample-level prototypes) showcase the ability for denoising and extracting subtle key features that experts find challenging to capture. This ability broadens human understanding and provides a promising solution from interpretability research to AI-for-Science.
Paper Structure (12 sections, 17 equations, 12 figures, 6 tables)

This paper contains 12 sections, 17 equations, 12 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Model Achitecture
  4. Loss Function
  5. Three Interpreting Paths and Diagnosis Framework
  6. Experiments
  7. Conventional Fault Diagnosis of Planetary Gearbox
  8. Domain Generalization Experiment of Bevel Gearbox
  9. Discussion
  10. Discussion of Loss Coefficient and Distance Metric
  11. Discussion of The Number of Prototypes
  12. Conclusion

Figures (12)

  • Figure 1: The classification logic of prototype-matching, a concept that categorizes an object based on its similarity to the prototype. Prototype-matching is a common logic in general human classification, where human can tell the typical sample of each class. In contrast, the classification logic of AI in IFD tasks remains unknown, despite its excellent diagnostic prediction, and AI could not explain "what typical fault signals look like".
  • Figure 2: The Architecture of Proposed PMN.
  • Figure 3: The entire framework for applying PMN to mechanical intelligent fault diagnosis.
  • Figure 4: The experimental setup for the planetary gearbox dataset. a) The Experiment rig. b) The schematic diagram.
  • Figure 5: Ten fault types included in the planetary gearbox dataset. a) SP: Sun gear pitting. b) SC: Sun gear crack. c) Sw: Sun gear partial tooth wearing. d) SW: Sun gear full tooth wearing. e) PC: Planet gear crack. f) PW: Planet gear full tooth wearing. g) BI: Bearing inner race. h) BO: Bearing outer race. Additionally, two other faults, bearing cage (BC) fault rolling (BR) fault, are also considered.
  • ...and 7 more figures