WO2023245990A1 - 一种多元信息融合的发电厂电气设备分类与异常检测方法 - Google Patents

一种多元信息融合的发电厂电气设备分类与异常检测方法 Download PDF

Info

Publication number
WO2023245990A1
WO2023245990A1 PCT/CN2022/135716 CN2022135716W WO2023245990A1 WO 2023245990 A1 WO2023245990 A1 WO 2023245990A1 CN 2022135716 W CN2022135716 W CN 2022135716W WO 2023245990 A1 WO2023245990 A1 WO 2023245990A1
Authority
WO
WIPO (PCT)
Prior art keywords
electrical equipment
fault
sound
neural network
classification
Prior art date
Application number
PCT/CN2022/135716
Other languages
English (en)
French (fr)
Other versions
WO2023245990A8 (zh
Inventor
彭煜民
冯志延
张豪
孙建超
李壮
沈燕
赵增涛
付建美
刘士刚
陈艳鑫
高靖昂
徐子韬
陆哲贤
李毅
周天添
Original Assignee
南方电网调峰调频发电有限公司储能研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 南方电网调峰调频发电有限公司储能研究院 filed Critical 南方电网调峰调频发电有限公司储能研究院
Publication of WO2023245990A1 publication Critical patent/WO2023245990A1/zh
Publication of WO2023245990A8 publication Critical patent/WO2023245990A8/zh

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
    • G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
    • G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks

Definitions

  • the invention relates to the field of electric power, and in particular to a multi-information fusion sound classification and abnormality detection method for electrical equipment in a power plant.
  • Computer hearing technology mainly targets the recognition and processing of non-human voices, including music signal processing combined with knowledge in the music field and audio information processing with knowledge in other fields.
  • Computer auditory technology is a discipline that understands and analyzes the content of digital sound and music based on audio signal processing and machine learning. It is an interdisciplinary subject that uses computational methods to understand and analyze the content of digital sound and music. The main basic discipline is audio Signal processing and machine learning.
  • audio Signal processing and machine learning is Currently, the development direction of computer hearing technology is mainly divided into two categories: music-oriented computer hearing technology and environmental sound-oriented computer hearing technology (which can be called computer hearing technology based on general audio).
  • the application characteristics of computer hearing technology are mainly reflected in: based on traditional audio signal processing, audio features are extracted, and combined with machine learning algorithms (mainly pattern recognition methods) to complete status monitoring and fault diagnosis.
  • machine learning algorithms mainly pattern recognition methods
  • Fault diagnosis is based on equipment operating noise, and the main application is acoustic target recognition.
  • electrical equipment sound abnormality detection methods include traditional electrical equipment sound monitoring methods and electrical equipment sound monitoring methods based on machine learning.
  • the target acoustic signal is mixed with a large amount of noise signals.
  • computer hearing technology When computer hearing technology is used to monitor the status of rotating equipment, it is mostly combined with traditional audio processing methods such as spectrum analysis or wavelet transform to judge the equipment. Whether a fault exists is then processed through complex audio feature extraction algorithms such as WPT, EMD or MFCC, and finally traditional classifiers such as HMM or SVM are used to complete status monitoring and fault warning tasks.
  • Acoustic signal processing of rotating equipment is generally combined with other monitoring parameters such as traditional vibration signals. The results of audio signal analysis are used to assist judgment, making the results based on vibration signal analysis more reliable.
  • there is still no unified and reliable equipment acoustic signal processing algorithm that can be applied to all equipment in the power system. Different acoustic signal feature extraction methods need to be selected based on the equipment's own operating characteristics and working status.
  • Traditional audio signal processing is mostly used in applications. technology.
  • non-rotating equipment mainly includes small equipment in power plants or primary and secondary equipment in substations. Since there is no operation of high-power rotating devices, non-rotating equipment has the characteristics of small vibration amplitude and low environmental noise. It can be seen from the above analysis that non-rotating equipment that performs status monitoring through equipment acoustic signals does not need the help of other auxiliary information to complete the status monitoring task; at the same time, since most non-rotating equipment operates in an open environment, the processing difficulty is relatively simple, so Audio processing of non-rotating equipment only requires traditional time-frequency domain algorithms such as FFT, STFT or WA to extract features, and perform status classification through mathematical statistical algorithms such as VQ, autocorrelation coefficient or fuzzy clustering to meet basic monitoring requirements. However, the application of the current algorithm reduces the integration of audio processing technology and machine learning algorithms, which is not conducive to improving the recognition rate and stability of condition monitoring.
  • the sound monitoring method of electrical equipment based on machine learning is as follows:
  • Sound-based online monitoring is a simple and reliable non-invasive monitoring method that will not interfere with the normal operation of electrical equipment and can well reflect the working status and abnormal conditions of electrical equipment.
  • electrical equipment There are many types of electrical equipment, complex structures, and various fault types, making it difficult to directly infer health conditions based on sound.
  • machine learning methods to monitor acoustic anomalies in electrical equipment has received widespread attention.
  • Deep learning technology usually requires more data for training than traditional machine learning, but it will not significantly increase the training time because the data scale and feature dimensions are too large. Difficulty, only the forward propagation of the neural network is required when running online, and it has high computational efficiency.
  • the application of deep learning technology to unsupervised/semi-supervised anomaly detection is still in its infancy.
  • Current deep learning anomaly detection methods include autoencoders, variational autoencoders, single-target generative adversarial active learning, and multi-target generative adversarial active learning.
  • the principles of principal component analysis and autoencoders are similar.
  • Principal component analysis is a linear combination of vectors using linear algebra technology, while autoencoders are a nonlinear combination of vectors using deep neural network technology.
  • Variational autoencoders use probabilistic methods when calculating anomaly scores, which are more interpretable than autoencoders.
  • Multi-objective generative adversarial active learning uses multiple generators to improve performance based on single-objective generative adversarial active learning.
  • electrical equipment sound classification algorithms include traditional sound classification algorithms and machine learning-based sound classification algorithms.
  • Sound classification algorithms have been widely used in speech recognition, environmental sound recognition and other fields, but are currently rarely used in the classification of operating sounds of electrical equipment.
  • the existing technology fuses Mel spectrum features and inverted Mel spectrum features as sound data sets, and uses multi-class support vector machines to classify power plant sounds, without diagnosing and identifying faults. ([2] Zhai Yongjie, Peng Yani, Yang Xu, Hu Dongyang, Wang Xinyin. Power plant equipment sound recognition algorithm integrating MFCC and IMFCC features)
  • This invention combines the process of fault discovery and subsequent manual analysis to propose a multi-information fusion power plant electrical equipment classification and anomaly detection method to solve the problem that the current algorithm only detects a single object, can only detect known anomalies, and does not understand the equipment operating status. Classification problem.
  • the object of the present invention is achieved by at least one of the following technical solutions.
  • a multi-information fusion method for classifying and anomaly detecting electrical equipment in power plants solves the problem that the current sound monitoring algorithm only detects a single object, can only detect known anomalies, and does not classify the operating status of the equipment. It includes the following steps:
  • step S5 If the state identification confidence ⁇ the confidence threshold, it is considered that there is a fault in the electrical equipment, and step S6 is executed. Otherwise, the sound classification of the electrical equipment status is completed;
  • step S7 If the fault identification confidence score ⁇ the confidence threshold, it is considered that there is an unknown fault in the electrical equipment, and step S8 is executed. Otherwise, the abnormal identification of known faults in the electrical equipment is completed;
  • step S1 the multiple information includes rotation speed, active output, reactive output, guide vane opening and water head.
  • step S2 the operating status of the electrical equipment and the corresponding multivariate information set are counted in advance, and the corresponding relationship between the operating status and the multivariate information is refined and used as a comprehensive criterion, and based on the comprehensive criterion and the information of the electrical equipment obtained in step S1 Multiple information to determine the operating status of electrical equipment.
  • step S3 by presetting the sound sensor position, the operating sound of the electrical equipment of the specific object is collected and stored.
  • step S4 the deep neural network is first trained through the electrical equipment status sound sample library, and the deep neural network in the training process adopts the LeNet convolutional neural network structure;
  • the training process includes sound preprocessing, time-frequency transformation and convolutional neural network training; sound preprocessing is duration cutting, and time-frequency transformation is to convert time domain signals into spectrograms;
  • the current operating status of the electrical equipment is obtained.
  • the trained deep neural network is used to identify the status of the collected audio to obtain the confidence of specific operating status identification.
  • step S5 the confidence of the identification of the specific operating status of the electrical equipment is compared with the preset confidence threshold. If it is greater than the confidence threshold, it is deemed that the identification of the operating status of the electrical equipment is normal; if it is less than or equal to the confidence threshold, then This audio is regarded as abnormal audio.
  • step S6 the deep neural network is first trained through the electrical equipment fault sound sample library, and the deep neural network in the training process adopts the LeNet convolutional neural network structure;
  • the trained deep neural network is used to identify known faults in abnormal audio, and the abnormal fault identification results and fault identification confidence are obtained.
  • step S7 the fault identification confidence is compared with the preset confidence threshold. If it is greater than the confidence threshold, it is deemed that the known fault recognition of the electrical equipment is normal, and a known fault occurs in the electrical equipment; if it is less than or equal to the confidence threshold threshold, the audio is considered to be unknown abnormal audio.
  • the method proposed by the present invention simultaneously realizes the effects of electrical equipment operating status classification, known fault audio identification, and unknown fault audio identification. It solves the problem that the current algorithm only detects a single object, can only detect known anomalies, and does not perform analysis on the equipment operating status. Classification problem.
  • Figure 1 is a step flow chart of a multi-information fusion power plant electrical equipment classification and anomaly detection method in an embodiment of the present invention
  • Figure 2 is a schematic diagram of the training process of the convolutional neural network in the embodiment of the present invention.
  • the operating status of the electrical equipment and the corresponding multi-dimensional information set are counted in advance, and the corresponding relationship between the operating status and the multi-dimensional information is extracted and used as a comprehensive criterion. Based on the comprehensive criterion and the multi-dimensional information of the electrical equipment obtained in step S1, the operation of the electrical equipment is determined. state.
  • the operating sound of the electrical equipment of a specific object is collected and stored in the form of a single-channel WAV.
  • the deep neural network is trained through the electrical equipment status sound sample library.
  • the deep neural network in the training process adopts the LeNet convolutional neural network structure
  • the training process includes sound preprocessing, time-frequency transformation and convolutional neural network training; sound preprocessing is duration cutting, and time-frequency transformation is to convert time domain signals into spectrograms;
  • the current operating status of the electrical equipment is obtained.
  • the trained deep neural network is used to identify the status of the collected audio to obtain the confidence of specific operating status identification.
  • step S5 If the state identification confidence ⁇ the confidence threshold, it is considered that there is a fault in the electrical equipment, and step S6 is executed. Otherwise, the sound classification of the electrical equipment status is completed;
  • the deep neural network is trained through the electrical equipment fault sound sample library.
  • the deep neural network in the training process adopts the LeNet convolutional neural network structure.
  • the fault identification confidence Compare the fault identification confidence with the preset confidence threshold. If it is greater than the confidence threshold, it is deemed that the known fault recognition of the electrical equipment is normal and a known fault occurs in the electrical equipment; if it is less than or equal to the confidence threshold, it is deemed that the electrical equipment has a known fault.
  • the audio is unknown abnormal audio.
  • the LeNet network includes an input layer, a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a flattening layer, a first fully connected layer, a second Fully connected layer (i.e. output layer).
  • STFT-1.WAV, STFT-2.WAV, and STFT-3.WAV are normal operating state sound data
  • STFT-4.WAV and STFT-5.WAV It is the sound data of the normal operating state
  • STFT-6.WAV, STFT-7.WAV and STFT-8.WAV are the abnormal sound data, realizing the classification of the normal state of the equipment.
  • the first convolution layer sets the convolution kernel to 5*5 and the number of convolution kernels to 20; the first maximum pooling layer sets the window size to 2*2 and downsamples; the second convolution layer sets the convolution kernel to 5*5, the number of convolution kernels is 50; the second pooling layer sets the window size to 2*2, downsampling; the first fully connected layer sets the list length to 500; the second fully connected layer sets the list length to The number of electrical equipment statuses to be monitored.
  • the first convolution layer sets the convolution kernel to 5*5 and the number of convolution kernels to 20; the first maximum pooling layer sets the window size to 2*2 and downsamples; the second convolution layer sets the convolution kernel to 5*5, the number of convolution kernels is 50; the second pooling layer sets the window size to 2*2, downsampling; the first fully connected layer sets the list length to 500; the second fully connected layer sets the list length to The number of known fault conditions of electrical equipment to be monitored.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
  • Testing And Monitoring For Control Systems (AREA)

Abstract

本发明公开了一种多元信息融合的发电厂电气设备分类与异常检测方法。所述方法包括以下步骤：获取电气设备的多元信息；判断电气设备所处的状态；预置声音传感器获取电气设备运行的音频；通过训练好的深度神经网络进行状态识别，得到状态识别置信度分数；若状态识别置信度≤置信度阈值，则认为电气设备存在故障，否则完成电气设备状态声音分类；通过训练好的深度神经网络进行已知故障异常识别，得到故障识别置信度分数；若故障识别置信度分数≤置信度阈值，则认为电气设备存在未知故障，否则完成电气设备已知故障异常识别；进行电气设备未知故障识别，并作为新异常样本加入电气设备故障声音样本库。本发明解决了目前算法存在的问题。

Description

一种多元信息融合的发电厂电气设备分类与异常检测方法 技术领域
本发明涉及电力领域，具体涉及一种多元信息融合的发电厂电气设备声音分类与异常检测方法。
背景技术
计算机听觉技术主要针对非人声进行识别与处理，包括结合音乐领域知识的音乐信号处理和其他领域知识的音频信息处理。计算机听觉技术是一个基于音频信号处理和机器学习、对数字声音与音乐的内容进行理解和分析的学科，是使用计算方法对数字化声音与音乐内容进行理解和分析的交叉学科，主要基础学科是音频信号处理和机器学习。当前，计算机听觉技术的发展方向主要分为2个类别：面向音乐的计算机听觉技术和面向环境声的计算机听觉技术(可称为基于一般音频的计算机听觉技术)。
计算机听觉技术的应用特点主要体现在：在传统音频信号处理的基础上，提取音频特征，结合机器学习算法(主要是模式识别方法)完成状态监测与故障诊断。在工业领域，计算机听觉技术也有很多的研究与应用，根据设备运转噪声进行故障诊断，主要进行声目标识别的应用。
在电力领域，电力设备在运行过程中，由于机械振动会产生声音，正常运行下的声音一般具有一定的规律性，但当设备发生某种故障后，由于运行状态或运行结构改变，其声音也会随之改变，比如出现机械故障时，其振动特性或部分频段内的振动能量将发生改变，同时会伴随刺耳或尖锐的噪声。此外，设备的超负荷运行或其他故障也会引起异常的声音变化。因此，电力设备的声音信号包含许多运行状态信息，具有丰富经验的工程师可以根据现场设备的异常声音，通过音色、音量、音高等音频特征的变化判断出设备是否处于不正常运行状态，甚至判别出故障的类型和严重程度。
现有技术中，电气设备声音异常检测方法包括传统电气设备声音监测方法和基于机器学习的电气设备声音监测方法。
传统电气设备声音监测方法，具体如下：
由于旋转设备的工作状态与运行环境等条件导致的目标声信号夹杂大量的噪声信号，当计算机听觉技术用于旋转设备的状态监测时，大多结合频谱分析或小波变换等传统音频处理方法来判断设备故障是否存在，再通过WPT、EMD或MFCC等音频特征提取复杂算法进行处理，最后采用HMM或SVM等传统分类器完成状态监测与故障预警任务。旋转设备的声信号处理一般会与传统的振动信号等其他监测参数结合，音频信号分析的结果用于辅助判断，使基于振动信号分析的结果更加可靠。但目前仍没有统一、可靠的设备声信号处理算法能够应用于电力系统的全部设备，需要结合设备自身运行特点与工作状态选择不同的声信号特征提取方法，在应用中大多采用传统的音频信号处理技术。
与旋转设备相比，非旋转设备主要包括电厂的小型设备或变电站的一、二次设备，由于没有高功率旋转装置的运行，具有振动幅度小、环境噪音小等特点。由上述分析可以看出，通过设备声信号进行状态监测的非旋转设备不需要其他辅助信息的帮助即可完成状态监测任务；同时由于非旋转设备大多在空旷环境中运行，处理难度相对简单，故非旋转设备的音频处理仅需要FFT、STFT或WA等传统时频域算法提取特征，并通过VQ、自相关系数或模糊聚类等数学统计算法进行状态分类就能够满足基本监测要求。但当前算法的应用降低了音频处理技术与机器学习算法的结合度，不利于状态监测识别率与稳定性的提高。
基于机器学习的电气设备声音监测方法，具体如下：
基于声音的在线监测是一种简单可靠的非侵入性监测方法,不会干扰电气设备的正常运行,并可以良好地反映电气设备的工作状态和异常情况。电气设备种类多、结构复杂、故障类型多样,难以直接根据声音推断出健康状况。采用机器学习方法对电气设备进行声学异常监测受到了广泛的关注，深度学习技术比传统机器学习通常霁要更多数据进行训练,但也不会因为数据规模、特征维度过大而显著增大训练难度,在线运行时仅需要神经网络的前向传播,具有很高的计算效率。然而,将深度学习技术应用于无/半监督的异常监测,还处于 刚起步阶段。目前深度学习异常检测方法包括自编码器、变分自编码器、单目标生成对抗式主动学习和多目标生成对抗式主动学习。主成分分析和自编码器原理类似,主成分分析是使用线性代数技术的向量线性组合，而自动编码器是使用深度神经网络技术的向量非线性组合。变分自编码器在计算异常分数时使用了概率方法,相对于自编码器更具有可解释性。多目标生成对抗式主动学习在单目标生成对抗式主动学习的基础上使用了多个生成器以提高性能。
现有技术中，电气设备声音分类算法包括传统声音分类算法和基于机器学习的声音分类算法。
传统声音分类算法，具体如下：
电气设备运行往往产生特定的声响，传统声音分类算法主要利用人工特征提取方法，再利用特征的差别对声音进行分类。提取短时能量、短时平均振幅、短时过零率、短时自相关函数等时域特征，结合傅里叶变换、声谱图、倒频谱、色谱图等频域特征，通过相似度矩阵等方法综合判断实现声音分类，
基于机器学习的声音分类算法，具体如下：
基于机器学习的声音分类算法大部分运用在语音识别、环境声分类等方面，应用在电力设备运行声音的深度学习分类算法较少，传统机器学习则是首先通过提取声音特征(如：梅尔频谱倒谱xishu)，再利用k-最近邻算法、支持向量机、字典学习([4]孙玉伟,罗林根,陈敬德,王辉,盛戈皞,江秀臣.基于声音特征与改进稀疏表示分类的断路器机械故障诊断方法)、BP神经网络([3]王剑强,彭涛.基于变压器运行声音识别的智能运维辅助决策技术[J].电子测试)等传统机器学习算法进行分类，在保证数据量的情况下，分类效果较好
基于深度学习的电气设备运行声音分类算法，目前还处于起步阶段，已有利用卷积神经网络、长短时记忆网络对开关柜局部放电识别以及变压器的声纹识别与分类([5]连玮琪.基于超声信号的高压开关柜局部放电识别方法研究)。现有技术已经实现通过卷积神经和循环神经网络对摩托车发动机声音进行故障诊断([1]聂慧兰.基于深度神经网络的设备声音故障诊断模型研究[D].湘潭大学)，
现有电气设备声音监测技术多数采用传统声音监测方法，效率低、难度大，利用机器学习方法需要数据规模大、训练次数多，还处于起步阶段。
声音分类算法已普遍利用到语音识别、环境声识别等领域，目前还较少运用在电气设备运行声音分类中。现有技术通过融合梅尔频谱特征与倒梅尔频谱特征作为声音数据集，并利用多分类支持向量机对电厂声音进行分类，未对故障进行诊断识别。([2]翟永杰,彭雅妮,杨旭,胡东阳,王新颖.融合MFCC和IMFCC特征的电厂设备声音识别算法)
然而目前的分类算法往往先固定了待检测对象，而且是检测先验已有的异常，无法充分发挥机器学习的强大学习力。
发明内容
本发明结合故障发现与事后人工分析的过程，提出了一种多元信息融合的发电厂电气设备分类与异常检测方法，解决目前算法只检测单一对象、只能检测已知异常、未对设备运行状态进行分类的问题。
本发明的目的至少通过如下技术方案之一实现。
一种多元信息融合的发电厂电气设备分类与异常检测方法，解决目前声音监测算法只检测单一对象、只能检测已知异常、未对设备运行状态进行分类的问题，包括以下步骤：
S1、通过发电厂监控系统，获取电气设备的多元信息；
S2、根据综合判据，判断电气设备所处的状态；
S3、预置声音传感器获取电气设备运行的音频；
S4、采用电气设备状态声音样本库训练深度神经网络，并通过训练好的深度神经网络进行状态识别，得到状态识别置信度分数；
S5、若状态识别置信度≤置信度阈值，则认为电气设备存在故障，执行步骤S6，否则完成电气设备状态声音分类；
S6、采用电气设备故障声音样本库训练深度神经网络，并通过训练好的深度神经网络进行已知故障异常识别，得到故障识别置信度分数；
S7、若故障识别置信度分数≤置信度阈值，则认为电气设备存在未知故障， 执行步骤S8，否则完成电气设备已知故障异常识别；
S8、进行电气设备未知故障识别，并作为新异常样本加入电气设备故障声音样本库。
进一步地，步骤S1中，所述多元信息包括转速、有功出力、无功出力、导叶开度和水头。
进一步地，步骤S2中，提前统计电气设备运行状态与对应的多元信息集，提炼得到运行状态与多元信息的对应关系，作为综合判据，并根据综合判据和步骤S1中得到的电气设备的多元信息，确定电气设备运行状态。
进一步地，步骤S3中，通过预置声音传感器位置，收集特定对象的电气设备运行声音并进行存储。
进一步地，步骤S4中，首先通过电气设备状态声音样本库对深度神经网络进行训练，训练过程的深度神经网络采用LeNet卷积神经网络结构；
训练过程包括声音预处理、时频变换和卷积神经网络训练；声音预处理为时长切割，时频变换为将时域信号转换成声谱图；
其次基于对多元信息的综合判据，获取电气设备当前运行状态，同时利用训练好的深度神经网络对采集的音频进行状态识别，得到具体运行状态识别的置信度。
进一步地，步骤S5中，将电气设备具体运行状态识别的置信度与预设的置信度阈值对比，若大于置信度阈值，则视为电气设备运行状态识别正常；若小于等于置信度阈值，则视为该音频为异常音频。
进一步地，步骤S6中，首先通过电气设备故障声音样本库对深度神经网络进行训练，训练过程的深度神经网络采用LeNet卷积神经网络结构；
训练过程包括声音预处理、时频变换和卷积神经网络训练；声音预处理为时长切割，时频变换为将时域信号转换成声谱图；
利用训练好的深度神经网络对异常音频进行已知故障识别，得到异常故障识别结果及其故障识别置信度。
进一步地，步骤S7中，将故障识别置信度与预设的置信度阈值对比，若 大于置信度阈值，则视为电气设备已知故障识别正常，该电气设备出现已知故障；若小于等于置信度阈值，则视为该音频为未知异常音频。
进一步地，步骤S8中，识别出未知异常音频，通过分析电气设备此时异常状态下的多元信息数据，以人工方式进行对该未知异常音频进行定性分析，将其打上故障标签，加入电气设备故障声音样本库。
相比与现有技术，本发明的优点在于：
1本发明提出的方法解决了传统声音监测方法需要深度挖掘不同电气设备运行状态声音特征从而导致难度大、效率低的问题；
2本发明提出的方法同时实现了电气设备运行状态分类、已知故障音频识别、未知故障音频识别的效果，解决了目前算法只检测单一对象、只能检测已知异常、未对设备运行状态进行分类的问题。
附图说明
图1为本发明实施例中一种多元信息融合的发电厂电气设备分类与异常检测方法的步骤流程图；
图2为本发明实施例中卷积神经网络的训练过程示意图；
图3为本发明实施例中LeNet网络的结构示意图。
具体实施方式
上述识别和跟踪方法组合为本发明较佳的实施方式，但本发明的实施方式并不受上述实施例的限制，其他任何未背离本发明的精神实质和原理下所作的修改、修饰、替代、组合、简化，均应为等效的置换方式，都应包含在本发明的保护范围之内。
实施例：
一种多元信息融合的发电厂电气设备分类与异常检测方法，解决目前声音监测算法只检测单一对象、只能检测已知异常、未对设备运行状态进行分类的问题，如图1所示，包括以下步骤：
S1、通过发电厂监控系统，获取电气设备的多元信息；
所述多元信息包括转速、有功出力、无功出力、导叶开度和水头。
S2、根据综合判据，判断电气设备所处的状态；
提前统计电气设备运行状态与对应的多元信息集，提炼得到运行状态与多元信息的对应关系，作为综合判据，并根据综合判据和步骤S1中得到的电气设备的多元信息，确定电气设备运行状态。
S3、预置声音传感器获取电气设备运行的音频；
通过预置声音传感器位置，收集特定对象的电气设备运行声音，以单通道WAV形式存储。
S4、采用电气设备状态声音样本库训练深度神经网络，并通过训练好的深度神经网络进行状态识别，得到状态识别置信度分数；
首先通过电气设备状态声音样本库对深度神经网络进行训练，训练过程的深度神经网络采用LeNet卷积神经网络结构；
训练过程包括声音预处理、时频变换和卷积神经网络训练；声音预处理为时长切割，时频变换为将时域信号转换成声谱图；
其次基于对多元信息的综合判据，获取电气设备当前运行状态，同时利用训练好的深度神经网络对采集的音频进行状态识别，得到具体运行状态识别的置信度。
S5、若状态识别置信度≤置信度阈值，则认为电气设备存在故障，执行步骤S6，否则完成电气设备状态声音分类；
将电气设备具体运行状态识别的置信度与预设的置信度阈值对比，若大于置信度阈值，则视为电气设备运行状态识别正常；若小于等于置信度阈值，则视为该音频为异常音频。
S6、采用电气设备故障声音样本库训练深度神经网络，并通过训练好的深度神经网络进行已知故障异常识别，得到故障识别置信度分数；
首先通过电气设备故障声音样本库对深度神经网络进行训练，训练过程的深度神经网络采用LeNet卷积神经网络结构
训练过程包括声音预处理、时频变换和卷积神经网络训练；声音预处理为时长切割，时频变换为将时域信号转换成声谱图；
利用训练好的深度神经网络对异常音频进行已知故障识别，得到异常故障识别结果及其故障识别置信度。
S7、若故障识别置信度分数≤置信度阈值，则认为电气设备存在未知故障，执行步骤S8，否则完成电气设备已知故障异常识别；
将故障识别置信度与预设的置信度阈值对比，若大于置信度阈值，则视为电气设备已知故障识别正常，该电气设备出现已知故障；若小于等于置信度阈值，则视为该音频为未知异常音频。
S8、进行电气设备未知故障识别，通过分析电气设备此时异常状态下的多元信息数据，以人工方式进行对该未知异常音频进行定性分析，并作为新异常样本加入电气设备故障声音样本库；
通过分析电气设备此时异常状态下的多元信息数据，以人工方式进行对该未知异常音频进行定性分析，将其打上故障标签，加入电气设备故障声音样本库。
通过监控系统得到电气设备运行数据，以wav格式存储，根据综合判据得到电气设备运行状态标签。如图3所示，LeNet网络包括输入层、第一卷积层、第一最大池化层、第二卷积层、第二最大池化层、压平层、第一全连接层、第二全连接层(即输出层)。第一卷积层设置卷积核为5*5，卷积核个数为20；第一最大池化层设置窗口大小为2*2，向下采样；第二卷积层设置卷积核为5*5，卷积核个数为50；第二池化层设置窗口大小为2*2，向下采样；第一全连接层设置列表长度为500；第二全连接层设置列表长度为所需监测的电气设备状态数。
通过声音传感器得到设备声音数据，利用预训练好的LeNet网络对采集到的声音数据进行分类，得到分类标签及其置信度如下表所示：
表1 LeNet声音分类结果及其置信度
声音数据 LeNet分类状态 置信度
STFT-1.WAV 正常运行状态一 97.132％
STFT-2.WAV 正常运行状态一 97.816％
STFT-3.WAV 正常运行状态一 94.382％
STFT-4.WAV 正常运行状态二 97.715％
STFT-5.WAV 正常运行状态二 96.937％
STFT-6.WAV 正常运行状态二 71.956％
STFT-7.WAV 正常运行状态二 54.096％
STFT-8.WAV 正常运行状态二 70.289％
通过设置正常状态识别置信度阈值85％，可以得知STFT-1.WAV、STFT-2.WAV、STFT-3.WAV为正常运行状态一声音数据，STFT-4.WAV和STFT-5.WAV为正常运行状态二声音数据，而STFT-6.WAV、STFT-7.WAV和STFT-8.WAV为异常声音数据，实现了设备正常状态分类。
实施例2：
本实施例中，通过上一阶段的设备正常状态分类，得到了异常声音数据STFT-6.WAV、STFT-7.WAV和STFT-8.WAV。LeNet网络包括输入层、第一卷积层、第一最大池化层、第二卷积层、第二最大池化层、压平层、第一全连接层、第二全连接层(即输出层)。第一卷积层设置卷积核为5*5，卷积核个数为20；第一最大池化层设置窗口大小为2*2，向下采样；第二卷积层设置卷积核为5*5，卷积核个数为50；第二池化层设置窗口大小为2*2，向下采样；第一全连接层设置列表长度为500；第二全连接层设置列表长度为所需监测的电气设备已知故障状态数。
利用预训练好的LeNet网络对上一阶段得到异常声音数据进行已知故障分类，得到分类标签及其置信度如下表所示：
表2 LeNet声音故障分类结果及其置信度
声音数据 LeNet分类状态 置信度
STFT-6.WAV 敲击故障 94.935％
STFT-7.WAV 敲击故障 95.375％
STFT-8.WAV 敲击故障 52.363％
通过设置故障状态识别置信度阈值80％，可以得知STFT-6.WAV、STFT-7.WAV为敲击故障(已知故障)声音数据，STFT-8.WAV为未知故障异常声音数据，实现了设备在已知故障下的声音分类，并发现在采集STFT-8.WAV 时间段内电气设备出现了未知故障。
实施例3：
本实施例中，以MP3格式存储的电气设备运行声音数据。LeNet网络包括输入层、第一卷积层、第一最大池化层、第二卷积层、第二最大池化层、压平层、第一全连接层、第二全连接层(即输出层)。第一卷积层设置卷积核为5*5，卷积核个数为20；第一最大池化层设置窗口大小为2*2，向下采样；第二卷积层设置卷积核为5*5，卷积核个数为50；第二池化层设置窗口大小为2*2，向下采样；第一全连接层设置列表长度为500；第二全连接层设置列表长度为所需监测的电气设备状态数。
通过声音传感器得到设备声音数据，利用预训练好的LeNet网络对采集到的声音数据进行分类，得到分类标签及其置信度如下表所示：
表3 LeNet声音分类结果及其置信度
声音数据 LeNet分类状态 置信度
STFT-1.MP3 正常运行状态一 96.258％
STFT-2.MP3 正常运行状态一 98.658％
STFT-3.MP3 正常运行状态二 93.499％
STFT-4.MP3 正常运行状态二 97.957％
STFT-5.MP3 正常运行状态二 73.602％
STFT-6.MP3 正常运行状态二 70.934％
STFT-7.MP3 正常运行状态二 81.759％
通过设置正常状态识别置信度阈值90％，可以得知STFT-1.MP3、STFT-2.MP3为正常运行状态一声音数据，STFT-3.MP3、STFT-4.MP3为正常运行状态二声音数据，而STFT-5.MP3、STFT-6.MP3和STFT-7.MP3为异常声音数据，实现了设备正常状态分类。
将STFT-5.MP3、STFT-6.MP3和STFT-7.MP3进行故障分类识别。LeNet网络包括输入层、第一卷积层、第一最大池化层、第二卷积层、第二最大池化层、压平层、第一全连接层、第二全连接层(即输出层)。第一卷积层设置卷积核为5*5，卷积核个数为20；第一最大池化层设置窗口大小为2*2，向下采 样；第二卷积层设置卷积核为5*5，卷积核个数为50；第二池化层设置窗口大小为2*2，向下采样；第一全连接层设置列表长度为500；第二全连接层设置列表长度为所需监测的电气设备已知故障状态数。
利用预训练好的LeNet网络对上一阶段得到异常声音数据进行已知故障分类，得到分类标签及其置信度如下表所示：
表4 LeNet声音故障分类结果及其置信度
声音数据 LeNet分类状态 置信度
STFT-5.MP3 敲击故障 86.688％
STFT-6.MP3 敲击故障 87.843％
STFT-7.MP3 敲击故障 89.090％
通过设置故障状态识别置信度阈值80％，可以得知STFT-5.MP3、STFT-6.MP3为敲击故障(已知故障)声音数据，STFT-7.MP3为未知故障异常声音数据，实现了设备在已知故障下的声音分类，并发现在采集STFT-7.MP3时间段内电气设备出现了未知故障。

Claims (10)

  1. 一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，包括以下步骤：
    S1、通过发电厂监控系统，获取电气设备的多元信息；
    S2、根据综合判据，判断电气设备所处的状态；
    S3、预置声音传感器获取电气设备运行的音频；
    S4、采用电气设备状态声音样本库训练深度神经网络，并通过训练好的深度神经网络进行状态识别，得到状态识别置信度分数；
    S5、若状态识别置信度≤置信度阈值，则认为电气设备存在故障，执行步骤S6，否则完成电气设备状态声音分类；
    S6、采用电气设备故障声音样本库训练深度神经网络，并通过训练好的深度神经网络进行已知故障异常识别，得到故障识别置信度分数；
    S7、若故障识别置信度分数≤置信度阈值，则认为电气设备存在未知故障，执行步骤S8，否则完成电气设备已知故障异常识别；
    S8、进行电气设备未知故障识别，并作为新异常样本加入电气设备故障声音样本库。
  2. 根据权利要求1所述的一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，步骤S1中，所述多元信息包括转速、有功出力、无功出力、导叶开度和水头。
  3. 根据权利要求1所述的一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，步骤S2中，提前统计电气设备运行状态与对应的多元信息集，提炼得到运行状态与多元信息的对应关系，作为综合判据，并根据综合判据和步骤S1中得到的电气设备的多元信息，确定电气设备运行状态。
  4. 根据权利要求1所述的一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，步骤S3中，通过预置声音传感器位置，收集特定对象的电气设备运行声音并进行存储。
  5. 根据权利要求1所述的一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，步骤S4中，首先通过电气设备状态声音样本库对 深度神经网络进行训练，训练过程的深度神经网络采用卷积神经网络结构；
    训练过程包括声音预处理、时频变换和卷积神经网络训练；声音预处理为时长切割，时频变换为将时域信号转换成声谱图；
    其次基于对多元信息的综合判据，获取电气设备当前运行状态，同时利用训练好的深度神经网络对采集的音频进行状态识别，得到具体运行状态识别的置信度。
  6. 根据权利要求1所述的一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，步骤S5中，将电气设备具体运行状态识别的置信度与预设的置信度阈值对比，若大于置信度阈值，则视为电气设备运行状态识别正常；若小于等于置信度阈值，则视为该音频为异常音频。
  7. 根据权利要求1所述的一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，步骤S6中，首先通过电气设备故障声音样本库对深度神经网络进行训练，训练过程的深度神经网络采用卷积神经网络结构；
    训练过程包括声音预处理、时频变换和卷积神经网络训练；声音预处理为时长切割，时频变换为将时域信号转换成声谱图；
    利用训练好的深度神经网络对异常音频进行已知故障识别，得到异常故障识别结果及其故障识别置信度。
  8. 根据权利要求5或7所述的一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，卷积神经网络包括输入层、卷积层、池化层和输出层。
  9. 根据权利要求1所述的一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，步骤S7中，将故障识别置信度与预设的置信度阈值对比，若大于置信度阈值，则视为电气设备已知故障识别正常，该电气设备出现已知故障；若小于等于置信度阈值，则视为该音频为未知异常音频。
  10. 根据权利要求1所述的一种多元信息融合的发电厂电气设备分类与异常检测方法，其特征在于，步骤S8中，识别出未知异常音频，通过分析电气设备此时异常状态下的多元信息数据，以人工方式进行对该未知异常音频进行 定性分析，将其打上故障标签，加入电气设备故障声音样本库。
PCT/CN2022/135716 2022-06-24 2022-11-30 一种多元信息融合的发电厂电气设备分类与异常检测方法 WO2023245990A1 (zh)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
CN202210727865 2022-06-24
CN202210727865.2 2022-06-24
CN202211214141.4A CN115602193A (zh) 2022-06-24 2022-09-30 一种多元信息融合的发电厂电气设备分类与异常检测方法
CN202211214141.4 2022-09-30

Publications (2)

Publication Number Publication Date
WO2023245990A1 true WO2023245990A1 (zh) 2023-12-28
WO2023245990A8 WO2023245990A8 (zh) 2024-01-25

Family

ID=84845469

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/135716 WO2023245990A1 (zh) 2022-06-24 2022-11-30 一种多元信息融合的发电厂电气设备分类与异常检测方法

Country Status (2)

Country Link
CN (1) CN115602193A (zh)
WO (1) WO2023245990A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117914003A (zh) * 2024-03-19 2024-04-19 沈阳智帮电气设备有限公司 基于云边协同的箱式变压器智能监测辅助方法及系统

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116359642B (zh) * 2023-03-10 2024-01-23 湖南金烽信息科技有限公司 一种变压器运行状态5g智能监测系统与方法
CN116754934B (zh) * 2023-05-22 2024-02-23 杭州轨物科技有限公司 一种高压断路器机械特性故障诊断方法
CN117711436B (zh) * 2024-02-05 2024-04-09 中国电子科技集团公司第十五研究所 一种基于多传感器融合的远场声音分类方法和装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200233397A1 (en) * 2019-01-23 2020-07-23 New York University System, method and computer-accessible medium for machine condition monitoring
CN111666982A (zh) * 2020-05-19 2020-09-15 上海核工程研究设计院有限公司 一种基于深度神经网络的机电设备故障诊断方法
US20210272580A1 (en) * 2020-03-02 2021-09-02 Espressif Systems (Shanghai) Co., Ltd. System and method for offline embedded abnormal sound fault detection
US20220099527A1 (en) * 2020-09-29 2022-03-31 Aktiebolaget Skf Method and system for performing fault diagnosis by bearing noise detection
CN114530166A (zh) * 2022-01-29 2022-05-24 国网福建省电力有限公司电力科学研究院 基于背景声纹理的变压器有载分接开关故障诊断方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200233397A1 (en) * 2019-01-23 2020-07-23 New York University System, method and computer-accessible medium for machine condition monitoring
US20210272580A1 (en) * 2020-03-02 2021-09-02 Espressif Systems (Shanghai) Co., Ltd. System and method for offline embedded abnormal sound fault detection
CN111666982A (zh) * 2020-05-19 2020-09-15 上海核工程研究设计院有限公司 一种基于深度神经网络的机电设备故障诊断方法
US20220099527A1 (en) * 2020-09-29 2022-03-31 Aktiebolaget Skf Method and system for performing fault diagnosis by bearing noise detection
CN114530166A (zh) * 2022-01-29 2022-05-24 国网福建省电力有限公司电力科学研究院 基于背景声纹理的变压器有载分接开关故障诊断方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117914003A (zh) * 2024-03-19 2024-04-19 沈阳智帮电气设备有限公司 基于云边协同的箱式变压器智能监测辅助方法及系统
CN117914003B (zh) * 2024-03-19 2024-05-24 沈阳智帮电气设备有限公司 基于云边协同的箱式变压器智能监测辅助方法及系统

Also Published As

Publication number Publication date
WO2023245990A8 (zh) 2024-01-25
CN115602193A (zh) 2023-01-13

Similar Documents

Publication Publication Date Title
WO2023245990A1 (zh) 一种多元信息融合的发电厂电气设备分类与异常检测方法
CN111626153B (zh) 一种基于集成学习的局放故障状态识别方法
CN102426835B (zh) 一种基于支持向量机模型的开关柜局部放电信号识别方法
CN112201260B (zh) 一种基于声纹识别的变压器运行状态在线检测方法
CN108896878B (zh) 一种基于超声波的局部放电检测方法
CN109034046B (zh) 一种基于声学检测的电能表内异物自动识别方法
CN108169639A (zh) 基于并行长短时记忆神经网络的识别开关柜故障的方法
CN107993648A (zh) 一种无人机识别方法、装置及电子设备
CN115035913B (zh) 一种声音异常检测方法
CN112268615B (zh) 一种机电设备振动信号特征提取方法
CN116778964A (zh) 一种基于声纹识别的变电设备故障监测系统及方法
Zhu et al. Complex disturbances identification: A novel PQDs decomposition and modeling method
CN116776075A (zh) 风机叶片健康监测系统及其监测方法
WO2023245991A1 (zh) 融合频带自向下注意力机制的电厂设备状态听觉监测方法
CN114352486A (zh) 一种基于分类的风电机组叶片音频故障检测方法
CN114137410A (zh) 一种基于声纹检测技术的液压机构断路器故障识别方法
CN111933186B (zh) 用于对有载分接开关进行故障识别的方法、装置及系统
Marchi et al. Audio onset detection: A wavelet packet based approach with recurrent neural networks
Sharma et al. Voice activity detection using windowing and updated K-means clustering algorithm
CN114492543A (zh) 一种基于时频域特征的局部放电分类识别方法
Shao et al. Multi-signal Joint HVCB Fault Diagnosis Research Based on Adaptive Framing MFCC Feature Extraction Method
Yu et al. Automatic Condition Monitoring and Fault Diagnosis System for Power Transformers Based on Voiceprint Recognition
Li et al. GIS insulation defect diagnosis method based on improved MFCC and PCA-SVM model
CN112466322A (zh) 一种机电设备噪声信号特征提取方法
Mengxi et al. Research on Failure Identification of Partial Discharge Ultrasonic Signal Based on GFCC

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22947725

Country of ref document: EP

Kind code of ref document: A1