Does Your Model Think Like an Engineer? Explainable AI for Bearing Fault Detection with Deep Learning

Thomas Decker, Michael Lebacher, Volker Tresp

Abstract Deep Learning has already been successfully applied to analyze industrial sensor data in a variety of relevant use cases. However, the opaque nature of many well-performing methods poses a major obstacle for real-world deployment. Explainable AI (XAI) and especially feature attribution techniques promise to enable insights about how such models form their decision. But the plain application of such methods often fails to provide truly informative and problem-specific insights to domain experts. In this work, we focus on the specific task of detecting faults in rolling element bearings from vibration signals. We propose a novel and domain-specific feature attribution framework that allows us to evaluate how well the underlying logic of a model corresponds with expert reasoning. Utilizing the framework we are able to validate the trustworthiness and to successfully anticipate the generalization ability of different well-performing deep learning models. Our methodology demonstrates how signal processing tools can effectively be used to enhance Explainable AI techniques and acts as a template for similar problems.

Introduction

In industrial predictive maintenance, deep learning models have demonstrated remarkable accuracy in identifying machinery failures. However, for engineers responsible for high-stakes factory operations, these models often remain "black boxes." When a model flags a bearing failure, a simple binary output is insufficient; engineers need to know if the model is identifying the same physical signatures—such as periodic impacts—that have been the cornerstone of mechanical diagnostics for decades.

This research addresses the "interpretability gap" in rolling element bearing fault detection. By bridging the divide between high-performance deep learning and classical mechanical engineering, the authors propose a framework to determine if a model truly understands the physics of the system or is merely exploiting spurious correlations.

The Interpretability Gap in Vibration Analysis

Standard Explainable AI (XAI) tools, such as LIME or Integrated Gradients, typically provide feature attributions in the raw input domain. For vibration data, this results in heatmaps highlighting specific points in a complex, noisy time-series signal. To a human expert, these highlights are often meaningless.

In contrast, mechanical engineers traditionally use Envelope Analysis to diagnose faults. This process identifies periodic impacts caused by defects (e.g., a crack on an outer ring) which occur at mathematically predictable frequencies, such as the Ball Passing Frequency Outer ring (BPFO). For AI to be trusted in this domain, its explanations must be translated into this frequency-based "language."

Methodology: The Domain Mapping Framework

The core innovation of this work is a Domain Mapping Framework that mathematically "wraps" a deep learning model to allow explanations to be calculated in the frequency domain rather than the time domain.

1. Augmented Model Architecture

The researchers define a mapping using the Hilbert Transform (to extract the signal envelope) and the Fourier Transform. By treating this transformation as a linear input layer, they enable standard XAI methods (like Integrated Gradients or SHAP) to attribute importance directly to specific frequencies in the envelope spectrum.

2. Quantitative Alignment Metrics

To move beyond subjective visual interpretations, the framework introduces two key metrics to quantify how well a model aligns with engineering logic:

  • Informative Attribution Share (IAS): Measures the percentage of the model's total "attention" that falls on the BPFO and its harmonics.
  • Informative Signal to Noise Ratio (ISNR): A decibel-scale ratio comparing the importance assigned to physically relevant frequencies versus irrelevant "noise" frequencies.

Experimental Setup

The framework was validated using the Paderborn University (Kat-DataCenter) dataset, a benchmark in the field. The dataset includes vibration recordings from undamaged bearings, bearings with artificial faults, and bearings with "genuine" faults from accelerated lifetime tests. To test the models' ability to generalize, data was collected across varying operating conditions, such as different rotation speeds (e.g., 1500 RPM vs. 900 RPM).

Results and Discussion

The researchers compared three common architectures—Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Fully-Connected Networks (FCNs)—against a physics-based baseline.

The Generalization Gap

While all models achieved high accuracy on the training data, their performance diverged sharply when tested on new rotation speeds. The FCN’s precision plummeted to 0.487, indicating it had failed to learn the underlying physics. In contrast, the CNN maintained a precision of 0.994, demonstrating robust generalization.

"Clever Hans" Models

The framework revealed that the FCN was a "Clever Hans" model—it appeared intelligent by achieving high training scores but was actually "cheating" by relying on spurious correlations in the data. The CNN showed the highest ISNR, proving its decisions were rooted in characteristic fault frequencies. Visualizations of the envelope spectrum attributions confirmed this: the CNN’s attention clearly peaked at the BPFO, while the FCN’s attention was scattered across irrelevant frequencies.

Conclusion

This research demonstrates that "thinking like an engineer" is not just a matter of interpretability—it is a prerequisite for robustness. Models that align with domain-specific physical principles are significantly more likely to generalize to real-world industrial environments. By using signal processing tools to enhance XAI, this methodology provides a template for building trustworthy AI in complex engineering applications. Future work will look to extend this framework to more complex fault types, such as inner ring or cage defects, and evaluate performance in high-noise industrial settings.

Summary updated

Abstract

Deep Learning has already been successfully applied to analyze industrial sensor data in a variety of relevant use cases. However, the opaque nature of many well-performing methods poses a major obstacle for real-world deployment. Explainable AI (XAI) and especially feature attribution techniques promise to enable insights about how such models form their decision. But the plain application of such methods often fails to provide truly informative and problem-specific insights to domain experts. In this work, we focus on the specific task of detecting faults in rolling element bearings from vibration signals. We propose a novel and domain-specific feature attribution framework that allows us to evaluate how well the underlying logic of a model corresponds with expert reasoning. Utilizing the framework we are able to validate the trustworthiness and to successfully anticipate the generalization ability of different well-performing deep learning models. Our methodology demonstrates how signal processing tools can effectively be used to enhance Explainable AI techniques and acts as a template for similar problems.
BibTeX
@inproceedings{decker2023does,
  author = {Thomas Decker and Michael Lebacher and Volker Tresp},
  title = {{Does Your Model Think Like an Engineer? Explainable AI for Bearing Fault Detection with Deep Learning}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=7adLZUJtPi},
}
Figure