2025年01期 v.46 36-40+46页

林朝晖1  廖奕校2  招智铭1  万智勇2  周松斌2

（1.广州市自来水有限公司石门水厂，广东 广州 510000

2.广东省科学院智能制造研究所，广东 广州 510070）

摘要：水泵是供水系统的重要加压设备，对其进行异常检测并及时发现运行异常，对保障供水安全具有重要意义。针对现有人工智能方法在水泵异常检测中存在的异常样本获取困难、检测精度低等问题，提出一种基于特征分布对齐与多传感器融合的水泵异常检测方法。该方法以多通道传感信号的对数梅尔谱为输入；利用卷积自编码网络来融合多传感器信息；以最小化信号重构损失和特征分布损失为目标，训练卷积自编码网络；基于重构损失计算样本的异常分数，实现水泵的异常检测。实验结果表明，该方法有效提高了水泵异常检测的性能。

关键词：水泵异常检测；数据重构；特征分布对齐；多传感器融合

中图分类号：TP311.5           文献标志码：A          文章编号：1674-2605(2025)01-0005-06

DOI：10.3969/j.issn.1674-2605.2025.01.005                    开放获取

Water Pump Anomaly Detection Method Based on Feature Distribution Alignment and Multi-sensor Fusion

LIN Zhaohui1  LIAO Yixiao2  ZHAO Zhiming1  WAN Zhiyong2  ZHOU Songbin2

(1.Shimen Waterworks, Guangzhou Water Supply Co., Ltd., Guangzhou 510000, China

2.Institute of Intelligent Manufacturing, Guangdong Academy of Sciences, Guangzhou 510070, China) 

Abstract: The water pump is an important pressurization device in the water supply system. Conducting anomaly detection and timely detection of operational abnormalities is of great significance for ensuring water supply safety. A water pump anomaly detection method based on feature distribution alignment and multi-sensor fusion is proposed to address the problems of difficulty in obtaining anomaly samples and low detection accuracy in existing artificial intelligence methods for water pump anomaly detection. This method takes the logarithmic Mel spectrum of multi-channel sensing signals as input; Using convolutional autoencoder networks to fuse multi-sensor information; Train a convolutional autoencoder network with the goal of minimizing signal reconstruction loss and feature distribution loss; Calculate the anomaly score of samples based on reconstruction loss to achieve anomaly detection of water pumps. The experimental results show that this method effectively improves the performance of water pump anomaly detection.

Keywords: anomaly detection of water pump; data reconstruction; feature distribution alignment; multi-sensor fusion