摘要详情

ID / 提交时间

35 / 2025-05-11 22:52:20

标题

Abnormal Sound Detection Method for Microfans Based on Lightweight Attention Capsule Network

关键字

Abnormal sound detection, Microfans, EfficientNet, Capsule network, Adaptive attention mechanism

主题及专题

6、Artificial Intelligence for EOM

状态

全文待审

作者

Wenlong Dong / Kunming university of science and technology

Yu Guo / Faculty of Mechanical and Electrical Engineering Kunming University of Science Technology

jiawei fan / Kunming University of Science and Technology

Tingwei Liu / Ltd；P&R Measurement Technology Co.

Chuanhui Wu / Ltd；P&R Measurement Technology Co.

摘要

Abnormal sound detection of microfans is key to ensuring product quality. Traditional manual auditory detection methods are subjective, while deep learning-based approaches can improve detection accuracy but often require large datasets, which are difficult to obtain in industrial applications. To address this, an abnormal sound detection method is proposed based on a lightweight attention capsule network (LACNet). First, A-weighting filtering is applied to the audio signals, and they are converted into time-frequency maps using Short-Time Fourier Transform (STFT). In the feature extraction phase, LACNet uses an optimized EfficientNet as the backbone, removing certain MBConv layers to achieve network lightweighting. Moreover, an adaptive attention mechanism is used to dynamically select key features from multi-scale representations. In the feature fusion phase, the dynamic routing mechanism of capsule networks is utilized to enhance the spatial hierarchical expression of features. Experimental results demonstrate that LACNet achieves excellent detection performance on both the MIMII dataset and our self-built microfans abnormal sound dataset, validating its effectiveness.