Abstract: Aiming at the problem of excessive parameters in existing anomaly detection techniques for hydraulic turbines, this paper proposes an anomaly detection algorithm for hydraulic turbines based on an improved time-delay neural network. Firstly, the real-time collected audio data of hydraulic turbines are subjected to basic signal processing, and then Fbank features are calculated as the input data of the model. The ECAPA-TDNN(Emphasized Channel Attention, Propagation and Aggregation in TDNN)algorithm- 引用格式:何梦腾,潘浩文,邓洪兴,等.基于点云多尺度方向一致性的羊只体尺测量方法［J］.河北大学学报(自然科学版),2026,46(2):113-127.引用格式:尹宁,王小伟,魏鹏程,等.基于改进时延神经网络的水轮机异常声纹检测算法［J］.河北大学学报(自然科学版),2026,46(2):215-224.DOI:10.3969/j.issn.1000-1565.2026.02.010基于改进时延神经网络的水轮机异常声纹检测算法尹宁¹,王小伟¹,魏鹏程¹,周九道¹,施琳长弘¹,廖勇²(1.贵州乌江水电开发有限责任公司 构皮滩发电厂,贵州 遵义 564408;2.重庆大学 微电子与通信工程学院,重庆 400044)摘 要:为了解决现有水轮机异常检测技术参数量较大的问题,本文提出了一种基于改进时延神经网络的水轮机异常检测算法.首先将实时采集到的水轮机的音频数据进行基本的信号处理,然后计算Fbank特征,将其作为模型的输入数据,使用简单无参数注意力模块(simple parameter-free attention module,SimAM)优化后的ECAPA-TDNN(emphasized channel attention, propagation and aggregation in TDNN)算法对其进行训练,得到输出特征向量,然后对样本数据采用同样的处理方式,最终得到实时数据和样本数据的输出特征向量,然后计算对应的欧几里得相似度,并依据初始设定的相似度阈值来判断是否异常,如有异常,则与样本集中的典型异常状态进行比较,用于对实时监测到的异常状态进行分类.与主流的卷积神经网络(convolutional neural network, CNN)、ResNet以及单独的ECAPA-TDNN等模型的对比结果可知,本文所建立的模型在参数量、算法准确率和等误率性能上均存在较大的优势和提升,参数量为19.7 M,模型准确率达到99.12%,模型等误率低至0.9%,可以很好地满足水轮机异常检测的需求.关键词:水电设备;异常检测;注意力模块;时延神经网络;深度学习中图分类号:TP391.41;TU714 文献标志码:A 文章编号:1000-1565(2026)02-0215-10DOI:10.3969/j.issn.1000-1565.2026.02.010Anomaly acoustic detection algorithm for hydraulic turbine based on improved time-delay neural networkYIN Ning¹, WANG Xiaowei¹, WEI Pengcheng¹, ZHOU Jiudao¹, SHILIN Changhong¹, LIAO Yong²(1.Goupitan Power Plant, Guizhou Wujiang Hydropower Development Co.,Ltd.,Zunyi 564408,China;2.School of Microelectronics and Communication Engineering, Chongqing University, Chongqing 400044, China)Abstract: Aiming at the problem of excessive parameters in existing anomaly detection techniques for hydraulic turbines, this paper proposes an anomaly detection algorithm for hydraulic turbines based on an improved time-delay neural network. Firstly, the real-time collected audio data of hydraulic turbines are subjected to basic signal processing, and then Fbank features are calculated as the input data of the model. The ECAPA-TDNN(Emphasized Channel Attention, Propagation and Aggregation in TDNN)algorithm- 收稿日期:2025-05-12;修回日期:2025-06-11 基金项目:中国华电集团科技项目(GPTDC/2024-1209) 第一作者:尹宁(1991—),男,贵州乌江水电开发有限责任公司高级工程师,主要从事水轮发电机机械检修方向研究.E-mail:1009683793@99.com 通信作者:廖勇(1982—),男,重庆大学副研究员,博士,主要从事移动通信、人工智能及其应用方向研究.E-mail:liaoy@cqu.edu.cn 第2期尹宁等:基于改进时延神经网络的水轮机异常声纹检测算法河北大学学报(自然科学版) 第46卷optimized by the Simple Parameter-free Attention Module(SimAM)is employed for model training to obtain the output feature vectors. The same processing procedure is applied to the sample data. Subsequently, the Euclidean similarity is calculated between the feature vectors of real-time data and sample data, and anomaly judgment is conducted according to a preset similarity threshold. If an anomaly is detected, it is further classified by comparing with typical abnormal states in the sample set. Compared with mainstream models such as Convolutional Neural Network(CNN), ResNet and the standalone ECAPA-TDNN, the proposed model shows significant advantages and improvements in terms of parameter quantity, algorithm accuracy and equal error rate(EER). The model achieves 19.7M parameters, 99.12% accuracy and 0.9% EER, which can well meet the requirements of anomaly detection for hydraulic turbines.

Key words: hydropower equipment, anomaly detection, attention module, time-delay neural network, deep learning