华南理工大学学报(自然科学版) ›› 2026, Vol. 54 ›› Issue (2): 38-51.doi: 10.12141/j.issn.1000-565X.250006

• 计算机科学与技术 • 上一篇    下一篇

基于改进EfficientNetV2的铝液泄漏声音识别与预警机制

梁艳辉1(), 温承杰1, 闫军威1,2(), 周璇1,2, 张洪涛1   

  1. 1.华南理工大学 机械与汽车工程学院/广州现代产业技术研究院,广东 广州 510640
    2.人工智能与数字经济 广东省实验室(广州),广东 广州 510335
  • 收稿日期:2025-01-06 出版日期:2026-02-25 发布日期:2025-07-18
  • 通信作者: 闫军威 E-mail:202211080708@mail.scut.edu.cn;mmjwyan@scut.edu.cn
  • 作者简介:梁艳辉(1990—),男,博士,工程师,主要从事铝加工安全研究。E-mail: 202211080708@mail.scut.edu.cn
  • 基金资助:
    广东省技术委托开发项目(2022440002001110)

Sound Recognition and Early Warning Mechanism for Liquid Aluminum Leakage Based on Improved EfficientNetV2

LIANG Yanhui1(), WEN Chengjie1, YAN Junwei1,2(), ZHOU Xuan1,2, ZHANG Hongtao1   

  1. 1.School of Mechanical and Automotive Engineering/ Guangzhou Modern Industrial Technology Research Institute,South China University of Technology,Guangzhou 510640,Guangdong,China
    2.Artificial Intelligence and Digital Economy Guangdong Provincial Laboratory (Guangzhou),Guangzhou 511442,Guangdong,China
  • Received:2025-01-06 Online:2026-02-25 Published:2025-07-18
  • Contact: YAN Junwei E-mail:202211080708@mail.scut.edu.cn;mmjwyan@scut.edu.cn
  • Supported by:
    the Guangdong Provincial Technology Commission Development Project(2022440002001110)

摘要:

铝液泄漏是导致铝加工深井铸造爆炸事故的直接原因。为解决实际工程中铝液泄漏判断方法滞后性强、准确率低和监测范围受限等问题,该文提出了基于改进EfficientNetV2的铝液泄漏声音识别方法。该方法通过声音特征判断铝液泄漏,以扩大监测范围;同时通过优化堆叠因子、引入高效通道注意力机制改进EfficientNetV2结构,以进一步提升识别速率与准确率。首先,利用拾音器采集不同场景下的声音数据,构建包含7类声音场景的声音数据库;然后,从声音信号中提取对数梅尔语谱图作为特征集,输入到改进的EfficientNetV2模型进行训练与验证,最终得到铝液泄漏声音识别模型。实验结果表明:改进的EfficientNetV2识别准确率达95.48%;与原始EfficientNetV2、ResNet、RegNet及DenseNet相比,改进模型的浮点运算次数分别为上述模型的12.34%、8.64%、11.14%和10.80%,参数量分别为上述模型的11.37%、9.55%、15.95%和17.24%,CPU环境下每秒处理图像帧数分别为上述模型的6.53倍、6.14倍、4.41倍和8.00倍,说明改进的EfficientNetV2具有快速准确的识别性能。此外,基于该文提出的铝液泄漏声音识别方法,构建了铝液泄漏风险预警机制,并将该机制应用于铸造单元的实时风险监测。实践结果验证了所提识别方法与预警机制的有效性,可为铝加工深井铸造爆炸事故的预防提供技术参考。

关键词: 铝加工深井铸造, 铝液泄漏, 声音识别, 风险预警, 改进的EfficientNetV2, 对数梅尔语谱图

Abstract:

Liquid aluminum leakage is the direct cause of explosion accidents in aluminum deep-well casting processes. To address the practical engineering challenges of strong lag, low accuracy, and limited monitoring range in existing leakage detection methods, this paper proposed a sound recognition method for liquid aluminum leakage based on an improved EfficientNetV2 model. This method utilizes acoustic characteristics to identify leaks, thereby expanding the monitoring range. The core enhancement involves optimizing the stacking factor and integrating an efficient channel attention mechanism into the EfficientNetV2 architecture to further improve the recognition speed and accuracy. Firstly, a sound database encompassing seven types of acoustic scenes was constructed by collecting audio data under different scenarios using pickups. Then, log-Mel spectrograms were extracted from the sound signal as the feature set and fed into the improved EfficientNetV2 model for training and validation, finally yielding the liquid aluminum leakage sound recognition model. The experimental results show that the recognition accuracy of the improved EfficientNetV2 reaches 95.48%. Compared to the original EfficientNetV2, ResNet, RegNet and DenseNet, the proposed model requires only 12.34%, 8.64%, 11.14%, and 10.80% of the floating point operations, and 11.37%, 9.55%, 15.95%, and 17.24% of the parameters, respectively. Furthermore, it processes 6.53, 6.14, 4.41, and 8.00 times more frames per second in a CPU environment, confirming its fast and accurate recognition performance. In addition, a risk early-warning mechanism for liquid aluminum leakage was established based on the proposed sound recognition method and deployed for real-time risk monitoring in a casting unit. Practical application results verify the effectiveness of both the identification method and the warning mechanism, providing a valuable technical reference for preventing explosion accidents in aluminum deep-well casting.

Key words: aluminum deep-well casting, liquid aluminum leakage, sound recognition, risk early warning, improved EfficientNetV2, logarithmic Mel-spectrogram

中图分类号: