基于归一化流概率模型的水电机组异常声音检测

Abnormal Sound Detection of Hydropower Units Based on Normalized Flow Probability Model

为推进水电机组智能化运维的发展,提出了一种自注意多阶统计量池化(SAMOSP)归一化流条件概率模型(NFCPM)用于水电机组的无监督异常声音检测。文中首次提出了自注意多阶统计量池化模块。该模块首先用一维压缩卷积层和瓶颈压缩激励部分自注意到时间帧的权重向量。权重向量用来计算多阶统计池化向量。然后再分频段的自注意到多阶统计池化量的不同权重,并按其提取不同频段的重要统计量信息,从而得到时频图的自注意统计池化特征向量。接着用归一化流条件概率模型对从水轮机组正常音频信号中提取到的自注意统计池化特征向量进行正常数据的概率建模。不同时间段的测试样本在该正常数据概率模型中进行测试,并得到相应的分数。分数越低表示概率密度越低,则异常程度越大,从而实现水电机组音频信号的无监督异常检测。

To promote the development of intelligent operation and maintenance of hydropower units,this paper proposes a robust multi-or⁃der statistic pooling(SAMOSP)normalized flow conditional probability mode(l NFCPM)for unsupervised abnormal sound detection of hydro⁃power units.A self-attentive multi-order statistic pooling module(SAMOSP)is proposed for the first time in the paper.The module starts with a one-dimensional compressed convolutional layer and a bottleneck compressing the weight vector of the excitation part of the self-atten⁃tive to time frame.The weight vectors are used to compute the multi-order statistical pooling vectors.Then the different weights of the self-at⁃tentive to multi-order statistical pooling quantities of the frequency bands are subdivided and information about the significant statistics of the different frequency bands is extracted by them to obtain the self-attentive statistical pooling feature vector of the time-frequency map.Test samples of different time periods are tested in this normal data probability model,and the corresponding scores are obtained.The lower score indicates the lower probability density,the greater the degree of abnormality,thus realizing the unsupervised abnormality detection of the audio signal of the hydropower units.