An Explainable Multi Sensor and Multi Domain Feature Fusion Approach for Bearing Fault Diagnosis Using Wavelet Transform and Transformer Networks

Introduction

Rolling bearings, which are frequently found in motors, gearboxes, turbines, pumps, and manufacturing equipment, are essential parts of rotating machinery. The dependability, effectiveness, and lifespan of mechanical systems are all directly impacted by their state of operation. Localized failures on rolling elements and bearing raceways can be caused by environmental disturbances, fluctuating loading, lubrication deterioration, and recurrent contact stresses during long-term operation [1]-[4]. Abnormal vibration, increased noise, decreased transmission efficiency, and ultimately an unplanned machinery shutdown are all possible outcomes of such faults.

As a result, condition-based maintenance and intelligent machinery health management depend heavily on precise bearing failure identification. Because localized flaws produce mechanical responses that can be detected from the bearing housing, vibration analysis is still one of the most well-established methods for monitoring bearing conditions. Complementary signal representations can be used to characterize fault information. While frequency domain analysis shows spectral features related to bearing dynamics, time domain indicators like root mean square, kurtosis, and crest factor characterize signal amplitude and impulsiveness. Time frequency methods, such wavelet transforms, offer localized representations of transient components for non-stationary fault signals that could not be sufficiently captured by either the time or frequency domain alone.

However, environmental interference, transmission routes, and sensor position can all have an impact on relying solely on vibration data. Because mechanical disturbances can alter the electromechanical response of the drive system, motor-current signals offer an additional and relatively non-intrusive source of condition information. The Paderborn bearing dataset was introduced by Lessmeier et al. [5] to support the development and evaluation of data-driven bearing condition monitoring methods using complementary vibration and motor-current measurements. Automatic feature learning for equipment diagnosis has significantly improved with recent developments in deep learning and machine learning [6]-[7]. While convolutional and recurrent neural networks can automatically extract spatial and temporal characteristics from measured signals, conventional methods like support vector machines and ensemble learning algorithms can provide effective classification when informative handcrafted features are available. Transformer designs have garnered more attention recently because to their ability to model global interdependence across diagnostic features using self-attention methods. In order to handle the heterogeneity and redundancy of multi sensor readings, Li et al. [8] suggested a multi branch feature cross fusion model based on Transformer learning in multi-sensor bearing diagnosis. Their findings showed the possibility of sensor specific feature extraction followed by global information fusion, but they also showed that good diagnostic representation is not always ensured by merely expanding the number of sensing channels.

There are still a number of restrictions in spite of these advancements. First, a lot of diagnostic methods focus on the accuracy of the final classification without adequately identifying the sensing modality or signal representation that yields the observed improvement. This is especially crucial for multi sensor and multi domain diagnosis, as fault discrimination may be influenced differently by vibration, motor current, time domain, frequency domain, and time frequency information. Second, robustness under measurement degradation or noise is not always demonstrated by high accuracy gained from pure laboratory signals. Third, it can be challenging to ascertain whether the predictions made by increasingly intricate deep learning architectures are supported by significant fault features since they can operate as black box classifiers. In situations when black-box artificial intelligence cannot be easily trusted, Magadán et al. [9] stressed the significance of explainable and interpretable bearing diagnosis. In a similar vein, SHAP-based methods have shown that feature attribution can be utilized to increase the transparency of bearing defect classification and discover diagnostically significant factors Brito et al. [10]. Therefore, a bearing diagnostic method that assesses sensor complementarity, multi domain information, classification capability, robustness, and interpretability within a single experimental methodology is still required. The separate contributions of vibration and motor current measurements should be measured instead of viewing feature fusion as intrinsically advantageous. Similarly, controlled ablation studies should be used to determine the contribution of time, frequency, and time frequency information. In order to interpret diagnostic conclusions from both data-driven and mechanical viewpoints, explainability should be further linked to measurable signal features.

This study leverages synchronized vibration and motor-current measurements to propose an explainable multi-sensor and multi domain feature fusion approach for intelligent bearing fault diagnostics in order to overcome these problems. Healthy condition (HH), inner ring fault (IR), and outer-ring fault (OR) are the three bearing health conditions that are taken into consideration. Transformer based learning for fault classification is combined with time domain statistical features, frequency domain data, and wavelet-based time frequency representations. The primary contributions are as follows:

• Vibration and two motor-current signals are used in tandem to explore complementary mechanical and electrical fault information;

• Time, frequency, and wavelet-based time frequency representations are integrated to characterize bearing faults from multiple signal domains;

• ablation experiments at the sensor and domain levels are designed to quantify the actual contributions of individual information sources; and

• explainability and noise robustness analyzes are incorporated to assess model stability and ascertain whether diagnostic choices can be linked to significant bearing signal characteristics.

The remainder of this paper is organized as follows. Section 2 describes the dataset, selected bearing conditions, signal preprocessing, and preliminary signal characteristics. Section 3 presents the proposed multi-sensor and multi-domain diagnostic method. Section 4 describes the experimental design, benchmark models, evaluation metrics, and ablation and robustness tests. Section 5 presents and discusses the experimental results and explainability analysis, while Section 6 summarizes the main conclusions and limitations.

Complete Article

The complete article, including all figures, tables, equations and algorithms, is available in the official publication PDF.

View Full PDF

Conclusion

This study implemented synchronized vibration and motor-current signals from a processed subset of the Paderborn bearing dataset to construct an explainable multi-sensor and multi-domain framework for bearing failure diagnostics. Comparative machine learning and Transformer-based categorization, sensor and feature-domain ablation, noise-sensitivity assessment, and SHAP-based explainability were linked with time domain, frequency domain, and time-frequency domain representations. The MLP obtained 96.30% accuracy and a 96.28% macro-F1 score under clean measurement conditions, while SVM, Random Forest, and the suggested Transformer reached 100% classification accuracy and macro-F1 score. The equivalent performance of many classifiers suggests that substantial class separability was already supplied by the extracted representations and that the assessed subset did not benefit further from increasing model complexity.

Vibration measures independently reached 100% accuracy, while motor-current signals achieved 96.30%, according to sensor-ablation data, demonstrating that electrical measurements preserve significant complementary diagnostic information. While the time, frequency, and time-frequency-domain representations alone were adequate to differentiate the chosen bearing conditions under clean measurements, their combination offers complementing explanations of bearing behavior instead of an extra improvement in clean-data accuracy. However, the noise-sensitivity investigation showed significant performance degradation under signal perturbation, showing that robustness in noisy industrial contexts should not be inferred from high benchmark accuracy under clean laboratory conditions.

These findings were validated by the explainability analysis. According to the revised SHAP analysis, vibration-derived features accounted for 42.01% of the total sensor-level attribution, while phase current 1 and phase current 2 contributed 29.27% and 28.72%, respectively. At the signal-domain level, time-frequency features provided the largest aggregated contribution at 51.56%, followed by frequency-domain features at 26.75% and time-domain features at 21.69%. These findings indicate that vibration remained the largest individual sensor-level contributor, while the two motor-current channels together provided substantial complementary information and time-frequency representations played the dominant role at the domain level. According to the agreement between SHAP attribution and the ablation results, the examined subset’s primary diagnostic source was vibration-based spectrum information, with additional information coming from motor-current and complementary signal representations. This study main limitation is that its conclusions are limited to a processed subset of the Paderborn benchmark; as a result, they cannot be applied to actual industrial settings or unseen physical bearings. Therefore, bearing-wise data separation, extra naturally damaged bearings, changing operating circumstances, noise-aware training, and cross-dataset or cross-domain validation should be used in future work to assess the framework. To ascertain whether the reported diagnostic performance and feature-importance patterns hold true under more practical machinery-monitoring situations, such investigations are required.

Data Availability

The bearing-condition monitoring data used in this study are publicly available from the Paderborn University Bearing Data Center. The processed dataset used for the analyses was derived from the publicly accessible benchmark data. Further details regarding data selection and preprocessing are provided in the Materials and Methods section.

References

1. S. Singh, C. Q. Howard, and C. H. Hansen, “An extensive review of vibration modelling of rolling element bearings with localised and extended defects, ” Journal of Sound and Vibration, vol. 357, pp. 300–330, 2015, doi: 10.1016/j.jsv.2015.04.037
2. Z. Chen, X. Hu, L. Li, C. Ye, D. Wu, S. Liu, Y. Ma, and L. Gu, “Study on the failure mechanism and improvement measures of bearings in the full-ocean-depth ultra-high-pressure seawater pump, ” Scientific Reports, vol. 15, Art. no. 41691, 2025, doi: 10.1038/s41598-025-25518-y
3. K. K. Raj, S. Kumar, and R. R. Kumar, “Systematic review of bearing component failure: Strategies for diagnosis and prognosis in rotating machinery, ” Arabian Journal for Science and Engineering, vol. 50, no. 8, pp. 5353–5375, 2025, doi: 10.1007/s13369-024-09866-x
4. X. Song, L. Bai, Y. Hu, L. Ma, and H. Cao, “A comprehensive review of rolling bearing life prediction: From fatigue life model to data-driven remaining useful life prognostic, ” Lubricants, vol. 14, no. 8, Art. no. 292, 2026, doi: 10.3390/lubricants14080292
5. C. Lessmeier, J. K. Kimotho, D. Zimmer, and W. Sextro, “Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set for data-driven classification, ” PHM Society European Conference, vol. 3, no. 1, 2016, doi: 10.36001/phme.2016.v3i1.1577
6. P. Gangsar, A. R. Bajpei, and R. Porwal, “A review on deep learning based condition monitoring and fault diagnosis of rotating machinery, ” Noise & Vibration Worldwide, vol. 53, no. 11, pp. 550–578, 2022, doi: 10.1177/09574565221139638
7. J. Cen, Z. Yang, X. Liu, J. Xiong, and H. Chen, “A review of data-driven machinery fault diagnosis using machine learning algorithms, ” Journal of Vibration Engineering & Technologies, vol. 10, no. 7, pp. 2481–2507, 2022, doi: 10.1007/s42417-022-00498-9
8. X. Li, S. Xiao, Q. Li, L. Zhu, T. Wang, and F. Chu, “The bearing multi-sensor fault diagnosis method based on a multi-branch parallel perception network and feature fusion strategy, ” Reliability Engineering & System Safety, vol. 261, Art. no. 111122, 2025, doi: 10.1016/j.ress.2025.111122
9. L. Magadán, C. Ruiz-Cárcel, J. C. Granda, F. J. Suárez, and A. Starr, “Explainable and interpretable bearing fault classification and diagnosis under limited data, ” Advanced Engineering Informatics, vol. 62, Art. no. 102909, 2024, doi: 10.1016/j.aei.2024.102909
10. L. C. Brito, G. A. Susto, J. N. Brito, and M. A. V. Duarte, “An explainable artificial intelligence approach for unsupervised fault detection and diagnosis in rotating machinery, ” Mechanical Systems and Signal Processing, vol. 163, Art. no. 108105, 2022, doi: 10.1016/j.ymssp.2021.108105
11. L. Lei, W. Li, S. Zhang, C. Wu, and H. Yu, “Research progress on data-driven industrial fault diagnosis methods, ” Sensors, vol. 25, no. 9, Art. no. 2952, 2025, doi: 10.3390/s25092952
12. W. Ke, S. Wang, L. Zhou, R. Wang, and Y. Dai, “Robot bone cutting tool status recognition method based on multimodal fusion network, ” Biomedical Signal Processing and Control, vol. 121, Art. no. 110197, 2026, doi: 10.1016/j.bspc.2026.110197
13. S. Ullah, M. Umar, and J.-M. Kim, “Physics guided fused image learning with enhanced squeeze excitation for failure analysis of multistage centrifugal pumps, ” Scientific Reports, vol. 16, Art. no. 16179, 2026, doi: 10.1038/s41598-026-47809-8
14. S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions, ” in Advances in Neural Information Processing Systems, vol. 30, pp. 4765–4774, 2017.
15. R. R. Kumar, M. Andriollo, G. Cirrincione, M. Cirrincione, and A. Tortella, “A comprehensive review of conventional and intelligence-based approaches for the fault diagnosis and condition monitoring of induction motors, ” Energies, vol. 15, no. 23, Art. no. 8938, 2022, doi: 10.3390/en15238938
16. I. U. Hassan, K. Panduru, and J. Walsh, “An in-depth study of vibration sensors for condition monitoring, ” Sensors, vol. 24, no. 3, Art. no. 740, 2024, doi: 10.3390/s24030740