Deep learning-based fault identification in condition monitoring

Hariom Dhungana, Suresh Kumar Mukhiya, Pragya Dhungana, Benjamin Karic

Abstract

Vibration-based condition monitoring techniques are commonly used to identify faults in rolling element bearings. Accuracy and speed of fault detection procedures are critical performance measures in condition monitoring. Delay is especially important in remote condition monitoring and time-sensitive industrial applications. While most existing methods focus on accuracy, little attention has been given to the inference time in the fault identification process. In this paper, we address this gap by presenting a Convolutional Neural Network (CNN) based approach for real-time fault identification in rolling element bearings. We encode raw vibration signals into two-dimensional images using various encoding methods and use these with a CNN to classify several categories of bearing fault types and sizes. We analyse the interplay between fault identification accuracy and processing time. For training and evaluation we use a bearing failure CWRU dataset.

Introduction

In modern industrial manufacturing, rotating machinery serves as the critical infrastructure for production. Among the various components of these machines, rolling element bearings are particularly vulnerable, accounting for over 50% of all mechanical defects. Consequently, vibration-based condition monitoring has become a standard diagnostic tool. However, a significant challenge remains: the trade-off between diagnostic accuracy and computational speed.

While high classification accuracy is essential for preventing false alarms, the latency of the diagnostic pipeline is equally critical for remote monitoring and time-sensitive applications. High latency can delay alerts, potentially leading to catastrophic failures. This research investigates a deep learning-based pipeline designed to minimize inference time while maintaining high precision in identifying the type and severity of bearing faults.

Methodology

The proposed approach utilizes a Convolutional Neural Network (CNN) architecture inspired by AlexNet. The model consists of three convolutional layers, three pooling layers, and three fully connected (dense) layers, totaling approximately 3.3 million trainable parameters.

Data Encoding Techniques

Because CNNs are optimized for two-dimensional image data, the primary challenge lies in converting one-dimensional vibration time-series signals into 2D representations. This study evaluates five distinct encoding methods:

  1. Pixel Strength: A direct mapping technique where normalized vibration values are arranged into a 2D grid.
  2. Gramian Angular Field (GAF): Represents signals in a polar coordinate system to preserve temporal correlations.
  3. Markov Transition Field (MTF): Captures the transition probabilities between different states in the time series.
  4. Recurrence Plot: Visualizes periodic patterns and rhythmic structures within the signal.
  5. GAF-MTF Hybrid: A dual-channel approach that combines GAF and MTF images to capture both static and dynamic features.

Experimental Setup

The study utilizes the Case Western Reserve University (CWRU) bearing dataset, a benchmark for fault diagnosis. The dataset includes:

  • Fault Types: Inner raceway (IR), ball, and outer raceway (OR) faults, alongside healthy samples.
  • Fault Severities: Damage diameters of 0.007", 0.014", and 0.021".
  • Operating Conditions: Data recorded at four different motor speeds (1720 to 1797 RPM).

The performance was evaluated across two classification tasks: a 4-category task (identifying the faulty component) and a 10-category task (identifying both the component and the specific fault size).

Results and Discussion

The Simplicity Paradox

A key finding of this research is that the most computationally simple method—Pixel Strength—outperformed more complex mathematical transformations. While GAF and MTF offer sophisticated representations, they may introduce noise or obscure features that the CNN can more easily extract from raw-mapped pixel data. The Pixel Strength model achieved an accuracy of 99.16% for 4-class identification and 96.15% for 10-class identification.

Latency and Real-Time Performance

The analysis reveals that the primary bottleneck in the diagnostic pipeline is the encoding time rather than the model inference time. For instance, the GAF-MTF hybrid encoding requires 56.21 ms, whereas the CNN inference takes only 3.5 ms. In contrast, the Pixel Strength approach is highly efficient, with a total processing time of 6.48 ms (5.07 ms for encoding and 1.41 ms for inference). This efficiency is vital for real-time industrial applications.

Generalization and Data Diversity

The study also found that training the model on a combined dataset featuring all four motor speeds improved accuracy by 1-3% compared to models trained on single-speed data. This suggests that the CNN benefits from data diversity, allowing it to learn features that are invariant to minor fluctuations in operating speed.

Conclusion

This research demonstrates that a CNN-based approach using simple Pixel Strength encoding provides an optimal balance between accuracy and speed for bearing fault identification. By achieving sub-10ms processing times with over 99% accuracy, this method is well-suited for real-time industrial condition monitoring.

Future work will focus on improving the robustness of the model against environmental noise typical of industrial floors and exploring advanced architectures, such as attention mechanisms, to further refine feature extraction without significantly increasing computational overhead.

Summary updated

Abstract

Vibration-based condition monitoring techniques are commonly used to identify faults in rolling element bearings. Accuracy and speed of fault detection procedures are critical performance measures in condition monitoring. Delay is especially important in remote condition monitoring and time-sensitive industrial applications. While most existing methods focus on accuracy, little attention has been given to the inference time in the fault identification process. In this paper, we address this gap by presenting a Convolutional Neural Network (CNN) based approach for real-time fault identification in rolling element bearings. We encode raw vibration signals into two-dimensional images using various encoding methods and use these with a CNN to classify several categories of bearing fault types and sizes. We analyse the interplay between fault identification accuracy and processing time. For training and evaluation we use a bearing failure CWRU dataset.
BibTeX
@inproceedings{dhungana2024deep,
  author = {Hariom Dhungana and Suresh Kumar Mukhiya and Pragya Dhungana and Benjamin Karic},
  title = {{Deep learning-based fault identification in condition monitoring}},
  booktitle = {Applepies},
  year = {2024},
  doi = {10.1007/978-3-031-84100-2\_50},
}
Figure