Efficient Gearbox Fault Diagnosis Based on Improved Multi-Scale CNN with Lightweight Convolutional Attention

Abstract

As a core transmission component of modern industrial equipment, the operation status of the gearbox has a significant impact on the reliability and service life of major machinery. In this paper, we propose an intelligent diagnosis framework based on Empirical Mode Decomposition and multimodal feature co-optimization and innovatively construct a fault diagnosis model by fusing a multi-scale convolutional neural network and a lightweight convolutional attention model. The framework extracts the multi-band features of vibration signals through the improved multi-scale convolutional neural network, which significantly enhances adaptability to complex working conditions (variable rotational speed, strong noise); at the same time, the lightweight convolutional attention mechanism is used to replace the multi-attention of the traditional Transformer, which greatly reduces computational complexity while guaranteeing accuracy and realizes highly efficient, lightweight local-global feature modeling. The lightweight convolutional attention is adaptively captured by the dynamic convolutional kernel generation strategy to adaptively capture local features in the time domain, and combined with grouped convolution to enhance the computational efficiency further; in addition, parameterized revised linear units are introduced to retain fault-sensitive negative information, which enhances the model's ability to detect weak faults. The experimental findings demonstrate that the proposed model achieves an accuracy greater than 98.9%, highlighting its exceptional diagnostic accuracy and robustness. Moreover, compared to other fault diagnosis methods, the model exhibits superior performance under complex working conditions.

Keywords: EMD; dynamic convolutional kernel; gearbox fault diagnosis; lightweight convolutional attention; multi-scale CNN.

Figure 1

Multi-scale convolutional neural network.

Figure 2

The LCA module generates dynamic convolutional kernels and performs grouped convolution operations.

Figure 3

Learning rate strategy comparison. (a) Exponential decay dynamic learning rate; (b) fixed learning rate.

Figure 4

Structure of MSCNN-LCA-Transformer.

Figure 5

Driveline simulator.

Figure 6

Examples of failures of different gears, with fault characteristics highlighted in red boxes. (a) Broken tooth; (b) root crack; (c) missing teeth; (d) tooth wear.

Figure 7

Signal segmentation example. (The red box represents the first 1024 points, the blue box indicates the last 1024 points, with a 50% overlap).

Figure 8

Diagnostic process.

Figure 9

The training accuracy curves for the Southeast University gearbox dataset. (a) Condition 1; (b) condition 2.

Figure 10

Confusion matrices of test set diagnostic results. (a) Condition 1; (b) condition 2.

Figure 11

Feature distribution visualization through t-SNE. (a) Raw data for condition 1; (b) raw data for condition 2; (c) data after classification for condition 1; (d) data after classification for condition 2.

Figure 12

The training accuracy curves for the WT planetary gearbox dataset.

Figure 13

Confusion matrix of test set diagnostic results.

Figure 14

Feature distribution visualization through t-SNE.

Figure 15

Comparison of accuracy of different models.