高级检索
    张海燕, 袁明帅, 蒋琦, 孙钰, 崔剑, 任利利, 骆有庆. 面向钻蛀振动实时侦听的深度学习模型压缩[J]. 北京林业大学学报, 2021, 43(6): 92-100. DOI: 10.12171/j.1000-1522.20200100
    引用本文: 张海燕, 袁明帅, 蒋琦, 孙钰, 崔剑, 任利利, 骆有庆. 面向钻蛀振动实时侦听的深度学习模型压缩[J]. 北京林业大学学报, 2021, 43(6): 92-100. DOI: 10.12171/j.1000-1522.20200100
    Zhang Haiyan, Yuan Mingshuai, Jiang Qi, Sun Yu, Cui Jian, Ren Lili, Luo Youqing. Deep learning model compression for real-time listening of boring vibration[J]. Journal of Beijing Forestry University, 2021, 43(6): 92-100. DOI: 10.12171/j.1000-1522.20200100
    Citation: Zhang Haiyan, Yuan Mingshuai, Jiang Qi, Sun Yu, Cui Jian, Ren Lili, Luo Youqing. Deep learning model compression for real-time listening of boring vibration[J]. Journal of Beijing Forestry University, 2021, 43(6): 92-100. DOI: 10.12171/j.1000-1522.20200100

    面向钻蛀振动实时侦听的深度学习模型压缩

    Deep learning model compression for real-time listening of boring vibration

    • 摘要:
        目的   基于深度学习的钻蛀振动识别模型结合边缘计算可实现林业蛀干害虫钻蛀振动长期监测和实时预警,但要求大幅压缩智能识别模型的参数量和运算量。本研究采用深度学习模型压缩算法,在不损失精度的前提下,对已有的钻蛀振动识别模型进行压缩,减小模型的体积并提升模型在嵌入式平台的识别速度。
        方法   首先采集双条杉天牛钻蛀振动和背景噪声两类信号训练人工设计的5层卷积神经网络BoringNet,得到钻蛀振动识别模型;然后分别使用不同裁剪率的滤波器裁剪、模型量化、多目标知识蒸馏对钻蛀振动识别模型进行压缩;最后设计上述压缩算法的组合策略,联合使用3种算法对蛀振动识别模型进行压缩,探究多种组合的模型压缩效果。
        结果   3种模型压缩算法组合,裁剪率为60%时模型达到最优,此时模型计算量和参数量分别从原模型的18.06 ×106次和0.54 ×106个降低为3.01 ×106次和0.09 × 106个,模型体积从2 200 kB压缩至134.9 kB,树莓派3B+上的识别时间由原模型的9.04 ms降低至1.65 ms,而模型精度仍能达到99.29%,提升了0.5%。
        结论   本研究的深度学习模型压缩方法,可以针对钻蛀振动侦听场景大幅压缩模型参数量和运算量,在保证准确率的前提下实现嵌入式平台的实时识别,促进钻蛀振动识别模型从工作站试验到野外实地部署的转变,为钻蛀振动识别的边缘计算奠定基础。

       

      Abstract:
        Objective   The boring vibration recognition model based on deep learning combined with edge computing can realize the long-term monitoring and real-time warning of boring vibration of forest stem borers, but it needs to greatly reduce the parameters and computation of the intelligent recognition model. In this study, the deep learning model compression algorithms were adopted to compress the existing boring vibration recognition model without losing accuracy, so as to reduce the volume of the model and improve the recognition speed of the model in the embedded platform.
        Method   Firstly, two kinds of signals, i.e. the boring vibration of Semanotus bifasciatus and the background noise were collected, and used to train five-layer convolution neural network BoringNet to obtain the boring vibration recognition model. Then, the convolution filter pruning with different rates, model quantization and multi-target knowledge distillation were used to compress the boring vibration recognition model. Finally, the combination strategies of the above-mentioned compression algorithms were designed and three algorithms were combined to compress the boring vibration recognition model to explore the model compression effect of various combinations.
        Result   When the pruning rate was 60% under the combination of 3 model compression algorithms, the model was the best. At this time, the amount of calculation and parameters of the model were reduced from 18.06 ×106 and 0.54 ×106 to 3.01 ×106 and 0.09 ×106, respectively. The volume of the model was compressed from 2200 kB to 134.9 kB. The recognition time of raspberry pie 3B + was reduced from 9.04 ms to 1.65 ms, and the accuracy of model was still 99.29%, which was improved by 0.5%.
        Conclusion   The deep learning model compression method in this study can greatly compress the model parameters and computation for the boring vibration interception scene, realize the real-time recognition of the embedded platform on the premise of ensuring the accuracy, promote the transformation of boring vibration recognition model from workstation test to field deployment, and lay the foundation for edge computing of boring vibration recognition.

       

    /

    返回文章
    返回