电子测量与仪器学报

张昶安,吴佳轩,陈世利.边缘云端协同的阀门内漏检测方法[J].电子测量与仪器学报,2025,39(12):34-42

边缘云端协同的阀门内漏检测方法

Edge-cloud collaboration for valve internal leakage detection

DOI：

英文关键词:acoustic emission detection valve internal leakage edge-cloud collaboration compressed sensing neural network

基金项目:

作者	单位
张昶安	天津大学精密测试技术及仪器国家重点实验室天津300072
吴佳轩	内蒙古工业大学化工学院呼和浩特010051
陈世利	天津大学精密测试技术及仪器国家重点实验室天津300072

Author	Institution
Zhang Chang′an	State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University, Tianjin 300072, China
Wu Jiaxuan	College of Chemical Engineering,Inner Mongolia University of Technology, Hohhot 010051, China
Chen Shili	State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University, Tianjin 300072, China

摘要点击次数: 721

全文下载次数: 421

中文摘要:

传统的声发射阀门内漏检测场景中，便携巡检式仪器检测存在实时性不足、数据存储与管理效率低、环境适应性受限等问题，而无线采集云端处理的方式受到电池寿命、和云端算力成本的制约。针对以上问题提出一种边缘-云端协同的阀门内泄漏声发射信号识别方法。首先构建轻量化的识别模型，在复频域内引入残差块和多头注意力机制，自适应关注不同频率成分之间的全局关系，增强模型对关键特征的聚焦能力。在残差结构中使用深度卷积，在注意力机制中对K、V做了维度分割实现压缩注意力机制，以此保证模型轻量化。映射回时域后，将原始输入与频域重构信号相加，避免频域处理过程中的信息丢失，同时缓解梯度消失问题。在训练阶段将编码器、解码器与识别模型一起训练，部署阶段将编码器部署在无线检测设备中，降低无线传输的功耗，解码器与识别模型部署在云端。实验结果表明，所提神经网络仅需10.1×103参数即可取得较好的表现，该方法在压缩比为8，准确率由99.5%降至98.9%的情况下，可将设备每次工作的功耗由0.49 mAh降低至0.15 mAh，有效延长电池使用寿命或提高检测频率。为阀门内泄漏长期在线监测和阀门内漏识别模型的低成本部署提供了解决方案。

英文摘要:

In the traditional acoustic emission valve internal leakage detection scenario, the portable inspection type instrument detection exists problems such as lack of real-time, low efficiency of data storage and management, and limited environmental adaptability, while the wireless acquisition and cloud processing are constrained by the battery life, and the cost of cloud computing power. To address the above problems, we propose an edge-cloud collaborative acoustic emission signal recognition method for valve internal leakage. Firstly, a lightweight recognition model is constructed, and residual blocks and multiple attention mechanisms are introduced in the complex frequency domain to adaptively focus on the global relationship between different frequency components and enhance the model’s ability to focus on key features. Deep convolution is used in the residual structure, and dimensional splitting of K and V is done in the attention mechanism to realize the compressed attention mechanism, so as to ensure the model lightweight. After mapping back to the time domain, the original input is summed with the reconstructed signal in the frequency domain to avoid information loss during frequency domain processing and to alleviate the problem of gradient vanishing. The encoder, decoder and recognition model are trained together in the training phase, the encoder is deployed in the wireless detection device to reduce the power consumption of wireless transmission in the deployment phase, and the decoder and recognition model are deployed in the cloud. The experimental results demonstrate that the proposed neural network model requires a mere 10.1×103 parameters to achieve optimal performance. This method, when implemented with a compression ratio of 8, reduces the accuracy from 99.5% to 98.9%, while concurrently reducing the energy consumption of the device from 0.49 mAh to 0.15 mAh. This enhancement not only prolongs the battery’s operational lifespan but also facilitates the enhancement of the detection frequency. This solution offers a cost-effective approach for the online monitoring and identification of leakage in valves.

查看全文查看/发表评论下载PDF阅读器