ZhangYang,GuoJun,AnLiang,et al.Leakage Noise Detection in Urban Water Supply Pipelines Based on Fine-tuning Qwen Multimodal Large Model[J].China Water & Wastewater,2026,42(17):41-47.
基于Qwen多模态大模型微调的供水管道漏损噪声检测
- Title:
- Leakage Noise Detection in Urban Water Supply Pipelines Based on Fine-tuning Qwen Multimodal Large Model
- Keywords:
- water supply pipeline; leakage detection; large language model; model fine-tuning; adaptive filtering; time-frequency analysis
- 摘要:
- 随着城市化进程加速，供水管道漏损问题日益严峻，传统声学检测方法存在效率低、误判率高、缺乏分析过程等局限。大语言模型具备智能分析和判别能力，因此提出一种基于Qwen多模态大模型微调的供水管道漏损噪声智能检测方法。通过自适应线谱增强算法对原始噪声信号进行滤波预处理，有效抑制环境干扰；将滤波后的信号转换为短时傅里叶变换时频图像，并结合漏损噪声的5个维度特征生成多模态数据集。实验表明，微调后模型准确率达77.75%，特异性从0.36%升至86.12%，显著优化了漏损与环境噪声的分类能力。该方法选用参数量较少的Qwen2-7B多模态大模型，并采用LoRA技术进行微调，可有效提升模型对漏损噪声的分类精度；同时，可输出详细分析过程，为技术人员提供了可解释决策依据。
- Abstract:
- With the acceleration of urbanization, the problem of leakage in urban water supply pipelines has become increasingly severe. Traditional acoustic detection methods suffer from low efficiency, high false-alarm rates, and a lack of transparent analysis procedures. The large language model has intelligent analysis and discrimination capabilities, therefore a water supply pipeline leakage noise intelligent detection method based on fine-tuning the Qwen multimodal large-scale model is proposed. An adaptive line spectral enhancement algorithm is applied to the raw noise signals for filtering and preprocessing, which effectively suppresses environmental interference. The filtered signals are then converted into short-time Fourier transform (STFT) time-frequency spectrograms, and combined with the five-dimensional leakage-noise feature vectors to construct a multimodal dataset. Experimental results demonstrated that the accuracy of model after fine-tuning reached 77.75%, and the specificity was improved from 0.36% to 86.12%, significantly optimizing the discrimination between leakage and environmental noise. This method selected the Qwen2-7B multimodal large model with fewer parameters and employed the LoRA technique for fine-tuning to effectively enhance the classification accuracy of leakage noise. Simultaneously, it could output a detailed analytical process, providing interpretable decision support for technicians.
相似文献/References:
[1]胡群芳,孙,肖瑞,等.城市综合管廊供水管道弯管及其支墩体系力学分析[J].中国给水排水,2020,36(20):19.
HU Qun-fang,SUN Yan,XIAO Rui,et al.Mechanical Analysis of Bends and Supports of Water Supply Pipeline in Urban Utility Tunnel[J].China Water & Wastewater,2020,36(17):19.
[2]黄智刚.长距离曲线岩石顶管技术用于供水管道建设[J].中国给水排水,2020,36(20):39.
HUANG Zhi-gang.Application of Long-distance Curved Pipe Jacking for Crossing Rock Formation in Water Supply Pipeline Construction[J].China Water & Wastewater,2020,36(17):39.
[3]杨振东,曹亚龙,张巧玲,等.含泄漏供水管道瞬变流动特征及泄漏定位[J].中国给水排水,2021,37(15):35.
YANG Zhen-dong,CAO Ya-long,ZHANG Qiao-ling,et al.Transient Flow Characteristics and Leakage Location of Water Supply Pipeline with Leakage Points[J].China Water & Wastewater,2021,37(17):35.
[4]何楠,郭浩,刘菁,等.供水管道中金属污染物富集与释放研究进展[J].中国给水排水,2022,38(2):21.
HE Nan,GUO Hao,LIU Jing,et al.Research Progress on Enrichment and Release of Metal Pollutants in Water Supply Pipelines[J].China Water & Wastewater,2022,38(17):21.
[5]王海涛,葛奇,江安华,等.水听器次声波在大口径供水管道探漏中的应用研究[J].中国给水排水,2024,40(19):39.
WANGHai-tao,GEQi,JIANGAn-hua,et al.Application of Hydrophone Infrasound Waves in Leak Detection of Large-diameter Water Supply Pipelines[J].China Water & Wastewater,2024,40(17):39.
[6]袁海伟,胡露,钟涛,等.铸铁供水管道沉积物污染富集特征[J].中国给水排水,2025,41(11):85.
YUANHai-wei,HULu,ZHONGTao,et al.Characteristics of Sediment Accumulation in Cast Iron Water Supply Pipelines[J].China Water & Wastewater,2025,41(17):85.
[7]徐明,崔红军,曹锋锋,等.球墨铸铁管内衬修复大型预应力钢筋混凝土供水管道中试[J].中国给水排水,2026,42(12):39.
XUMing,CUIHongjun,CAOFengfeng,et al.Pilot Test on Ductile Iron Pipe Lining Rehabilitation Technology for Aging Large Prestressed Reinforced Concrete Water Supply Pipelines[J].China Water & Wastewater,2026,42(17):39.
[8]赵林硕,叶郭煊,申永刚,等.基于时频卷积神经网络的供水管道漏损检测[J].中国给水排水,2023,39(17):53.
ZHAOLin-shuo,YEGuo-xuan,SHENYong-gang,et al.Leakage Detection of Water Supply Pipeline Based on Time-Frequency Convolutional Neural Network[J].China Water & Wastewater,2023,39(17):53.