Convolutional neural network intelligent diagnosis method using small samples based on SK-CAM

Liang Chen, Simin Li, Peijun Li, Yutao Liu, R.c-j. ChangJournal of Vibroengineering 2024 · DOI

TL;DR

Researchers have developed a new fault diagnosis method called SK-CAM that combines Spectral Kurtosis (SK) preprocessing with a Channel Attention Module (CAM) inside a Convolutional Neural Network (CNN). This approach achieves near-perfect accuracy (99.8%) in identifying mechanical failures even when training data is scarce, effectively overcoming the "data hunger" typically associated with deep learning in industrial settings.

What Problem Does This Paper Address?

In the world of industrial maintenance, identifying a failing bearing or gear before it breaks is critical for safety and cost-saving. While Convolutional Neural Networks (CNNs) are incredibly powerful at recognizing patterns in vibration data, they have two major "Achilles' heels":

  1. Data Dependency: They typically require massive amounts of labeled training data to perform well. In real-world factories, fault data is rare because machines (hopefully) don't fail that often.
  2. Noise Interference: Industrial environments are loud and messy. Standard CNNs often struggle to distinguish the subtle "click" of a microscopic crack from the background hum of a factory floor.

The authors aim to bridge this gap by creating a model that is both "data-efficient" (works with small samples) and "noise-robust."

Methods

The proposed SK-CAM framework uses a two-step punch to process vibration signals:

  1. Spectral Kurtosis (SK) Preprocessing: Instead of feeding raw vibration signals directly into the AI, the authors use the SK algorithm. Kurtosis is a statistical measure of "peakiness." In vibration analysis, a fault (like a crack) creates transient impacts—sharp spikes in the data. SK maps these spikes across different frequencies to create a "Fast Kurtosis Graph." This graph is then flattened into 1D samples, essentially highlighting the most suspicious parts of the signal before the AI even looks at it.
  2. Channel Attention Module (CAM): Once the data is inside the CNN, the authors insert a CAM. Think of this as a "smart volume knob" for the network’s internal features. It calculates which "channels" (feature sets) are most relevant to the specific fault and turns their "volume" up, while muting channels that contain mostly noise or irrelevant information.

By combining these, the model doesn't have to "learn" how to find impacts from scratch—the SK gives it a head start, and the CAM keeps it focused.

Datasets

The authors validated their model using two distinct sources:

  • Case Western Reserve University (CWRU) Bearing Dataset: The "gold standard" benchmark in the fault diagnosis community. It includes data for normal bearings and three types of faults (inner ring, outer ring, and ball faults) at various damage diameters.
  • QPZZ-II Rotating Machinery Test Bench: A custom experimental setup used to collect real-world gearbox fault data. This dataset included 10 different gear states, including complex "pitting" and "crack" faults on both single and double teeth.

Key Insights and Ideas

  • Domain Knowledge Matters: The "Aha!" moment here is that deep learning doesn't have to be a "black box." By injecting traditional signal processing (Spectral Kurtosis) into the pipeline, the authors reduced the complexity of the task for the CNN. This is why the model performs so well with fewer samples.
  • Attention as a Filter: The use of the Channel Attention Module effectively acts as an adaptive filter. In traditional CNNs, every feature is treated with equal importance initially. CAM allows the network to realize that "Channel A" might be great for detecting gear cracks while "Channel B" is just recording the motor's cooling fan, allowing it to ignore the latter.
  • Robustness to Data Scarcity: One of the most impressive findings was the model's stability. While standard deep learning models see a sharp drop in performance as you remove training data, SK-CAM remained highly accurate even when trained on only a fraction of the original dataset.

Key Results

  • Top-Tier Accuracy: On the CWRU bearing dataset, the method reached an identification accuracy of 99.8%.
  • Small Sample Performance: When the training data was slashed to just 20% of its original size, the model still maintained a 96% accuracy rate, significantly outperforming standard CNNs and SVM-based methods.
  • Generalization: In cross-load experiments (training on a machine under 1HP load and testing it on 3HP), the model maintained an average accuracy of 98.33%, proving it isn't just "memorizing" one specific operating condition.
  • Gearbox Success: In the more complex gearbox experiments, the proposed method averaged 98.96% accuracy, compared to just 95.41% for a standard CNN.

Limitations and Future Work

The primary limitation noted is that the current study focuses on machines running at a constant speed. In many industrial applications, machines speed up and slow down constantly (variable speed conditions).

The authors suggest that future research will integrate order tracking analysis—a technique specifically designed for variable speeds—with the SK-CAM model to make the system more "universal" for complex engineering environments.

Summary updated

Abstract

In order to solve the dependence of convolutional neural networks (CNN) on large samples of training data, an intelligent fault diagnosis method based on spectral kurtosis (SK) and attention mechanism is proposed. Firstly, the SK algorithm is used to obtain two-dimensional fast kurtosis graphs from vibration signals, and the two-dimensional fast spectral kurtosis graphs are converted into one-dimensional kurtosis time-domain samples, which are used as the input of CNN. Then the channel attention module (CAM) is added to CNN, and the weight is increased in the channel domain to eliminate the interference of invalid features. The accuracy of fault identification can reach 99.8 % by applying the proposed method on the fault diagnosis experiment of rolling bearings. Compared with the traditional deep learning (DL) method, the proposed method not only has higher accuracy, but also has lower dependence on the number of samples.
BibTeX
@article{chen2024convolutional,
  author = {Liang Chen and Simin Li and Peijun Li and Yutao Liu and R.C-J. Chang},
  title = {{Convolutional neural network intelligent diagnosis method using small samples based on SK-CAM}},
  journal = {Journal of Vibroengineering},
  year = {2024},
  doi = {10.21595/jve.2023.23384},
}
Figure