Deep Learning-Based Fault Identification in Condition Monitoring

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

TL;DR

This paper presents a deep learning approach for real-time bearing fault identification by converting raw vibration signals into 2D images. The authors demonstrate that a simple "pixel strength" encoding method surprisingly outperforms more complex mathematical transformations, achieving 99.16% accuracy with a total inference time of just 6.48 milliseconds.

What Problem Does This Paper Address?

In modern manufacturing, rotating machinery is the backbone of production, and bearing failures account for over 50% of all mechanical defects. While vibration-based condition monitoring is a standard diagnostic tool, the industry faces a "speed vs. accuracy" dilemma.

Most existing research focuses heavily on maximizing classification accuracy, often at the expense of computational complexity. However, for remote monitoring and time-sensitive industrial applications, high latency can lead to catastrophic failures before an alert is even generated. The authors aim to bridge this gap by identifying a deep learning pipeline that minimizes inference time without sacrificing the precision needed to distinguish between different types and sizes of bearing faults.

Methods

The researchers employed a Convolutional Neural Network (CNN) architecture (inspired by AlexNet) consisting of three convolutional layers, three pooling layers, and three fully connected (dense) layers, totaling approximately 3.3 million trainable parameters.

The core of their experimentation lies in Data Encoding. Since CNNs are designed for image processing, the 1D vibration time-series must be converted into 2D images. The authors compared five distinct techniques:

  1. Pixel Strength: A direct mapping where normalized vibration values are placed into a 2D grid.
  2. Gramian Angular Field (GAF): Encodes temporal correlation by representing signals in a polar coordinate system.
  3. Markov Transition Field (MTF): Captures transition probabilities between different states in the time series.
  4. Recurrence Plot: Visualizes the periodic nature and rhythmic patterns of the signal.
  5. GAF-MTF Hybrid: A dual-channel approach combining both GAF and MTF images to capture both static and dynamic features.

Datasets

The study utilizes the Case Western Reserve University (CWRU) bearing dataset, a benchmark in the field of fault diagnosis.

  • Fault Types: Inner raceway (IR), ball, and outer raceway (OR) faults, plus healthy samples.
  • Fault Severities: Diameters of 0.007", 0.014", and 0.021".
  • Operating Conditions: Data recorded at four different motor speeds (1720 to 1797 RPM).
  • Classification Tasks: The authors tested both a 4-category task (identifying the faulty part) and a 10-category task (identifying the part and the specific fault size).

Key Insights and Ideas

  • The "Simplicity" Paradox: The most significant "aha moment" is that Pixel Strength—the most computationally "naive" encoding method—was the top performer. While GAF and MTF provide sophisticated mathematical representations, they seem to introduce noise or obscure patterns that the CNN finds more easily in the raw-mapped pixel data.
  • Latency is the Bottleneck: The study highlights that the encoding time is often much higher than the model inference time. For example, GAF-MTF encoding takes 56.21ms, while the model only needs 3.5ms to process it. This underscores the need for efficient preprocessing in real-time systems.
  • Generalization through Diversity: Combining data from all four motor speeds into a single training set improved accuracy by 1-3% compared to training on single-speed datasets. This suggests that the CNN benefits from the "diversity" of the data, helping it learn features that are invariant to slight changes in RPM.

Key Results

  • Accuracy: The Pixel Strength model achieved 99.16% accuracy for 4-class identification and 96.15% for 10-class identification.
  • Speed: The total processing time for the Pixel Strength approach was 6.48 ms (5.07ms for encoding + 1.41ms for inference). In contrast, the GAF-MTF hybrid took nearly 60ms.
  • Encoding Comparison: GAF performed better than MTF, as the latter tended to hide patterns that were difficult for the model to distinguish, particularly in ball defects.

Limitations and Future Work

  • Environmental Noise: The CWRU dataset is collected in a controlled laboratory environment. Real-world industrial floors are significantly noisier, which may degrade the performance of the "Pixel Strength" method. The authors suggest that future work should incorporate noise-reduction techniques.
  • Model Optimization: While the current CNN is effective, the authors propose exploring more advanced architectures, such as those incorporating attention mechanisms, which could potentially improve feature extraction with minimal increases in computational cost.
  • Cross-Dataset Validation: The study is limited to the CWRU dataset; testing the pipeline on different machinery or newer benchmarks would help confirm the robustness of the pixel-based approach.

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},
}

Loading related content...

Figure