CN112348052A - A method for localizing abnormal sound sources of power transmission and transformation equipment based on improved EfficientNet - Google Patents

A method for localizing abnormal sound sources of power transmission and transformation equipment based on improved EfficientNet Download PDF

Info

Publication number
CN112348052A
CN112348052A CN202011063845.7A CN202011063845A CN112348052A CN 112348052 A CN112348052 A CN 112348052A CN 202011063845 A CN202011063845 A CN 202011063845A CN 112348052 A CN112348052 A CN 112348052A
Authority
CN
China
Prior art keywords
power
abnormal
classification model
sound source
abnormal sound
Prior art date
Legal status (The legal status 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 status listed.)
Granted
Application number
CN202011063845.7A
Other languages
Chinese (zh)
Other versions
CN112348052B (en
Inventor
周雪枫
赵培峰
赵国伟
马小源
王强
赵锐
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Datong Power Supply Co of State Grid Shanxi Electric Power Co Ltd
Original Assignee
Datong Power Supply Co of State Grid Shanxi Electric Power Co Ltd
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 Datong Power Supply Co of State Grid Shanxi Electric Power Co Ltd filed Critical Datong Power Supply Co of State Grid Shanxi Electric Power Co Ltd
Priority to CN202011063845.7A priority Critical patent/CN112348052B/en
Publication of CN112348052A publication Critical patent/CN112348052A/en
Application granted granted Critical
Publication of CN112348052B publication Critical patent/CN112348052B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S5/00—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations
    • G01S5/18—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations using ultrasonic, sonic, or infrasonic waves
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
    • Y02E40/00—Technologies for an efficient electrical power generation, transmission or distribution
    • Y02E40/70—Smart grids as climate change mitigation technology in the energy generation sector
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)

Abstract

本申请公开了一种基于改进EfficientNet的输变电设备异常声源定位方法，包括：利用声音采集单元，采集电力设备运行过程中的样本功率谱图，对样本功率谱图中的异常功率谱图进行异常功率值标注；根据卷积神经网络，引入复合系数，构建初始分类模型，并利用样本功率谱图对初始分类模型进行训练，将训练后的初始分类模型记作功率谱图分类模型；采集待检设备的待检功率谱图，根据功率谱图分类模型对待检功率谱图进行异常功率识别，当识别出异常功率时，采用融合算法，计算待检设备的异常声源坐标。通过本申请中的技术方案，结合三通道异常功率特征值，根据麦克风与异常声源间的传播时延，实现异常声源坐标的计算，完成电力设备运行过程中的异常声源定位。

Figure 202011063845

The present application discloses a method for locating abnormal sound sources of power transmission and transformation equipment based on improved EfficientNet. Carry out abnormal power value labeling; according to the convolutional neural network, introduce compound coefficients, construct the initial classification model, and use the sample power spectrum to train the initial classification model, and record the trained initial classification model as the power spectrum classification model; collect For the power spectrum of the device to be tested, the abnormal power of the power spectrum to be tested is identified according to the power spectrum classification model. When the abnormal power is identified, a fusion algorithm is used to calculate the coordinates of the abnormal sound source of the device to be tested. Through the technical solution in this application, combined with the three-channel abnormal power characteristic value, according to the propagation delay between the microphone and the abnormal sound source, the calculation of the coordinates of the abnormal sound source is realized, and the abnormal sound source localization during the operation of the power equipment is completed.

Figure 202011063845

Description

Power transmission and transformation equipment abnormal sound source positioning method based on improved EfficientNet
Technical Field
The application relates to the technical field of power system equipment detection, in particular to a power transmission and transformation equipment abnormal sound source positioning method based on improved EfficientNet.
Background
For equipment abnormity detection, the earlier machine equipment fault and damage are discovered, which means huge saving and improvement of industrial process efficiency, and has important significance for safe operation of a power distribution system.
In an electric power system, a conventional inspection process is to inspect equipment devices one by one through professional instruments by skilled maintenance technicians to judge whether the equipment is abnormal, the operation environment is complex in the inspection and inspection process, the inspection time is long, the working efficiency is low, the labor intensity is high, and the high requirements are provided for the technical capability and physical quality of the technicians.
The abnormal sound source positioning is a common method for equipment abnormality detection, and can timely and quickly find out equipment abnormal points and improve the equipment maintenance efficiency.
In the prior art, a microphone sound source positioning method is usually adopted, and azimuth information such as a sound source direction angle, an elevation angle, a distance and the like is calculated by combining mathematical geometry knowledge according to time errors of sound source signals reaching microphones with different spatial positions through a time delay estimation algorithm, so that sound source positioning is realized.
However, the position parameters calculated by using the method have low precision and are easily influenced by external environment noise, and particularly under the environment with high reverberation and strong noise and low signal-to-noise ratio (SNR), the abnormal sound source positioning performance of the method is very poor, and the requirement of equipment abnormal detection cannot be met.
Disclosure of Invention
The purpose of this application lies in: and calculating the coordinate of the abnormal sound source according to the propagation delay between the microphone and the abnormal sound source by combining the three-channel abnormal power characteristic value, and completing the positioning of the abnormal sound source in the operation process of the electrical equipment.
The technical scheme of the first aspect of the application is as follows: the method for positioning the abnormal sound source of the power transmission and transformation equipment based on the improved EfficientNet comprises the following steps: step 1, a sound acquisition unit is used for acquiring a sample power spectrogram in the operation process of electric equipment and carrying out abnormal power value labeling on an abnormal power spectrogram in the sample power spectrogram, wherein the sample power spectrogram comprises a normal power spectrogram and an abnormal power spectrogram; step 2, introducing a composite coefficient according to a convolutional neural network and an EfficientNet network, constructing an initial classification model, training the initial classification model by using a sample power spectrogram, and recording the trained initial classification model as a power spectrogram classification model; and 3, acquiring a power spectrogram to be detected of the equipment to be detected, identifying abnormal power of the power spectrogram to be detected according to the power spectrogram classification model, and calculating the coordinates of an abnormal sound source of the equipment to be detected by adopting a fusion algorithm when the abnormal power is identified.
In any one of the above technical solutions, further, the initial classification model includes an initial convolutional layer, the initial convolutional layer is configured to perform convolution and standardized calculation on input data, and in step 2, the training of the initial classification model by using the sample power spectrogram specifically includes:
step 21, randomly dividing a sample power spectrogram into training data and testing data;
step 22, inputting training data into the initial convolutional layer to obtain a standardized input image;
step 23, performing multilayer convolution module calculation on the standardized input image, and performing data enhancement on the intermediate value calculated by each layer of the network according to a composite coefficient, wherein the calculation formula of the composite coefficient is as follows:
Figure BDA0002713142560000021
wherein w is the width of the current network, d is the depth of the current network, (h, p) is the resolution of the standardized input image, r is the resolution magnification, and α is the learning rate of the current network;
step 24, according to the calculation result calculated by the multilayer convolution module, performing first revision on the parameters in the initial classification model, wherein the parameters comprise a learning rate alpha;
and 25, performing second revision on the parameters in the first revised initial classification model according to the test data, and recording the second revised initial classification model as a power spectrogram classification model.
In any of the above technical solutions, further, the revision formula of the parameters is:
Figure BDA0002713142560000031
Figure BDA0002713142560000032
Memory(N)≤targetmemory
FLOPS(N)≤target_flops
in the formula (I), the compound is shown in the specification,
Figure BDA0002713142560000033
for the ith convolution module in the initial classification model, Li represents the number of times the ith convolution module block repeats,
Figure BDA0002713142560000034
is to perform a derivation calculation on the Li,
Figure BDA0002713142560000035
represents the ith convolution block, formed by convolution layer FiRepetition of
Figure BDA0002713142560000036
In the second configuration, the result is subjected to a derivative calculation, i ═ 1,2, …, m, < > is an OR operation, targetmemoryFor the target value of resource consumption, target _ flops are the calculated amountThe consumption target value, Memory (-) is a resource consumption function, FLOPS (-) is a calculated amount consumption function, N (-) is a calculated result calculated by a layer of convolution module, and Accuracy (-) is a constraint function which enables the Accuracy of the network to be the highest within the constraint condition.
In any one of the above technical solutions, further, step 3 specifically includes: step 31, acquiring a power spectrogram to be detected of equipment to be detected by using a sound acquisition unit, and performing abnormal power identification on the power spectrogram to be detected according to a power spectrogram classification model to generate an abnormal power characteristic value of the power spectrogram to be detected, wherein the sound acquisition unit at least comprises 3 sound sensors; step 32, calculating the distance between the abnormal sound source and the sound sensor according to the time between two adjacent maximum power values in the power spectrogram to be detected, which is acquired by any two sound sensors, and the abnormal power characteristic value; and step 33, calculating the abnormal sound source coordinate of the abnormal sound source according to the distance and the space coordinate of the sound sensor.
In any of the above technical solutions, further, the calculation formula of the distance is:
Figure BDA0002713142560000037
in the formula, WuThe abnormal power characteristic value of the power spectrogram to be detected, which is acquired by the u-th sound sensor, u is 1, …, h, …, j, h and j are the labels of the sound sensors, a is a preset precision coefficient, and r is a preset precision coefficientuIs the distance between the abnormal sound source and the u-th sound sensor, VsIs the speed of propagation of sound in air, ThThe time difference between two adjacent maximum power values corresponding to the h-th sound sensor is obtained.
The technical scheme of the second aspect of the application is as follows: the device comprises a sound collection unit, a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program stored in the memory, the method for positioning the abnormal sound source of the power transmission and transformation equipment based on the improved EfficientNet is realized so as to position the abnormal sound source of the equipment to be detected.
The beneficial effect of this application is:
according to the technical scheme, the ternary microphone array is arranged to serve as a sound collecting unit, a trained power spectrogram classification model is utilized, whether the sound power value in the power spectrogram to be detected belongs to normal sound or abnormal sound is judged in real time, three-channel abnormal power characteristic values are combined, calculation of an abnormal sound source coordinate is achieved according to propagation delay between a microphone and an abnormal sound source, abnormal sound source positioning in the operation process of electric equipment is completed, the accuracy of calculation of the abnormal sound source coordinate is improved, the positioning time is shortened, and the positioning efficiency is improved.
The method provided by the application can be used for accurately classifying the sound emitted by the power transmission and transformation equipment, dividing the sound stream segments with fixed frame number into normal sound or abnormal sound, accurately capturing the abnormal sound stream segments and predicting the abnormal power characteristic value corresponding to the abnormal frame, and realizing accurate classification and prediction.
According to the method, the traditional time delay estimation algorithm is not used for sound source positioning, an innovative positioning method of fusing abnormal sound source power values and time delay data is adopted, a plurality of azimuth information is obtained through a plurality of formulas and geometric calculation, so that the equation is solved to obtain the target sound source coordinate, the three-dimensional coordinate vector value of the target abnormal sound source can be directly obtained through the method, and the accuracy of the result is high.
The method fully utilizes the strong modeling and fitting capability of the neural network, accurately classifies the input sound flow into normal and abnormal sound flows, predicts and captures the abnormal power characteristic value of the abnormal sound source, and then integrates the time delay information to calculate the azimuth coordinate, thereby realizing the real-time detection and positioning of the abnormal sound source of the power transmission and transformation equipment.
Drawings
The advantages of the above and/or additional aspects of the present application will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 is a schematic flow chart diagram of a method for improving the EfficientNet-based positioning of an abnormal sound source of an electric transmission and transformation device according to an embodiment of the present application;
FIG. 2 is a schematic diagram of a normal power spectrum and an abnormal power spectrum according to one embodiment of the present application;
FIG. 3 is a schematic diagram of a conventional convolutional neural network model according to one embodiment of the present application;
FIG. 4 is a power spectrogram classification model with the introduction of complex coefficients according to one embodiment of the present application;
fig. 5 is a schematic diagram of an abnormal sound source target coordinate calculation according to an embodiment of the present application.
Detailed Description
In order that the above objects, features and advantages of the present application can be more clearly understood, the present application will be described in further detail with reference to the accompanying drawings and detailed description. It should be noted that the embodiments and features of the embodiments of the present application may be combined with each other without conflict.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application, however, the present application may be practiced in other ways than those described herein, and therefore the scope of the present application is not limited by the specific embodiments disclosed below.
As shown in fig. 1, the present embodiment provides a method for positioning an abnormal sound source of an electric transmission and transformation device based on improved EfficientNet, where the method includes:
step 1, a sound acquisition unit is used for acquiring a sample power spectrogram in the operation process of electric equipment and carrying out abnormal power value labeling on an abnormal power spectrogram in the sample power spectrogram, wherein the sample power spectrogram comprises a normal power spectrogram and an abnormal power spectrogram;
in this embodiment, a conventional ternary microphone array is used as a sound collection unit, and the ternary microphone array is deployed around the power equipment, so as to collect sound data in the operation process of the power equipment, and obtain a sample power spectrogram, which is used as a database of a power spectrogram classification model.
When data is collected, a three-element microphone array can be arranged by adopting a planar or three-dimensional topological structure, the microphones and a corresponding PC are connected through three leads, audio processing software such as Audacity is installed on the PC, sound collected by multiple channels is analyzed by the audio processing software, normal working sound and abnormal working sound of the power transmission and transformation equipment are recorded on site, a sample power spectrogram output by the Audacity software is intercepted in real time, 10000 pieces of normal and abnormal power spectrograms of the power equipment are collected together by taking 100 frames as the total width of the spectrograms, and abnormal power values of the power equipment are marked in the abnormal power spectrogram, as shown in FIG. 2.
In this embodiment, the sound category of the normal power spectrogram is labeled as "normal", the sound category of the abnormal power spectrogram is labeled as "abnormal", and the "power value" of the abnormal power spectrogram is labeled.
As shown in fig. 3, in this embodiment, a conventional convolutional neural network model may be used as an initial model, 80% of the collected sample power spectrogram is used as training data, and the remaining 20% is used as test data, and the initial model is trained and tested to obtain a final power spectrogram classification model, thereby implementing abnormal power identification on the device to be detected.
In the embodiment, in order to improve the accuracy of the abnormal power identification of the equipment to be detected and the utilization degree of data resources, a conventional convolutional neural network model is improved to form a lightweight image classification deep learning model.
Step 2, introducing a composite coefficient according to a convolutional neural network and an EfficientNet network, constructing an initial classification model, training the initial classification model by using a sample power spectrogram, and recording the trained initial classification model as a power spectrogram classification model, wherein the initial classification model comprises an initial convolutional layer which is used for carrying out convolution and standardized calculation on input data;
in this embodiment, a convolutional neural network and an EfficientNet network model are combined, a complex coefficient is introduced, and all dimensions such as depth, width, resolution and the like of input data are scaled uniformly to adjust the extended network depth, so that better performance can be obtained without increasing the operation time.
As shown in fig. 4, in the present embodiment, the stem structure is used as an initial convolution layer, and convolution and normalization calculation are performed on input data.
In this embodiment, the power spectrum classification model after the introduction of the complex coefficient may be regarded as a network model formed by combining a plurality of convolution modules, and each time one convolution module block outputs a result, the dimension of the output result is scaled by the adaptive complex coefficient, so as to achieve the purpose of reducing the search space.
Specifically, in step 2, training the initial classification model by using the sample power spectrogram specifically includes:
and step 21, randomly dividing the sample power spectrogram into training data and testing data, wherein 80% of the sample power spectrogram is used as the training data, and the rest 20% of the sample power spectrogram is used as the testing data.
And step 22, inputting training data into the initial convolutional layer to obtain a standardized input image, and then entering a corresponding activation function and convolution module block.
In this embodiment, the convolution module block structure is divided into a main network and a residual edge structure.
Specifically, after being processed by a stem structure, training data firstly enters a block backbone network, and a feature layer of the training data (input image) is subjected to dimension-increasing operation by using a 1 × 1 convolution, then is subjected to standardization operation and an activation function, and then returns to the image feature layer after dimension-increasing.
Again, a depth separable convolution is performed, where feature size can be selected to be 3 x 3 or 5 x 5, and then a normalization operation and activation function are performed, applying an attention mechanism.
The attention mechanism in this embodiment is also divided into two parts, first, a feature layer of the attention mechanism is applied, here, an operation of global pooling is performed on an input image feature layer, only a dimension of the number of channels remains, then, an operation of reshape is performed, two dimensions of height and width are added to a result after the image is globally pooled, and then convolution is performed, here, convolution of 1 × 1 is performed first to compress the number of channels, then, the number of channels is extended through convolution of 1 × 1, after two times of convolution processing of 1 × 1, the obtained feature layer and the input feature layer have the same dimension, and then, sigmod operation is performed to fix the value between 0 and 1. Finally, multiplying the output under the attention mechanism with the image feature layer needing attention in the attention mechanism, so that each layer of feature layer needing attention is applied with the attention mechanism.
After the attention mechanism is completed, dimension reduction is carried out by utilizing a convolution of 1 multiplied by 1, then standardization operation and an activation function are carried out, a result under a main network is output, and the output result is added with an output result of a residual edge structure, so that a complete calculation process of a convolution module block structure is realized.
Step 23, performing multilayer convolution module calculation on the standardized input image, and performing data enhancement on the intermediate value calculated by each layer of the network according to a composite coefficient, wherein the calculation formula of the composite coefficient is as follows:
Figure BDA0002713142560000081
wherein w is the width of the current network, d is the depth of the current network, (h, p) is the resolution of the standardized input image, r is the resolution multiplying power, the initial value is 1, then the resolution is changed by continuously updating and optimizing in the network through constraint and multiplying the resolution (h, p), and alpha is the learning rate of the current network;
and 24, performing first revision on parameters in the initial classification model according to the calculation result calculated by the multilayer convolution module, wherein the parameters at least comprise a learning rate alpha, the width w of the current network, the depth d of the current network, resolution (h, p) and resolution multiplying power r.
In the process of parameter revision, the revision formula of the parameters is as follows:
Figure BDA0002713142560000082
Figure BDA0002713142560000083
Memory(N)≤targetmemory
FLOPS(N)≤target_flops
namely, the meanings of the above constraints are:
memory (n) is less than or equal to target _ memory, and the consumption of memory resources is less than a target value;
flops (n) is less than or equal to targetflops, and the consumption of the calculated amount is less than the target value.
In the formula (I), the compound is shown in the specification,
Figure BDA0002713142560000091
accuracy (. circle.) is a constraint function that maximizes the Accuracy of the network within the constraints, for the ith convolution module in the initial classification model,
Figure BDA0002713142560000092
represents the ith convolution block, formed by convolution layer FiRepetition of
Figure BDA0002713142560000093
And then the derivative calculation is carried out on the result, Li is the repeated times of the ith convolution module block,
Figure BDA0002713142560000094
to perform the derivation calculation for Li, i is 1,2, …, m, targetmemoryThe target _ FLOPS are the resource consumption target values, the target _ FLOPS are the calculated amount consumption target values, the Memory (-) is the resource consumption function, the FLOPS (-) is the calculated amount consumption function, the N (-) is the calculation result calculated by one layer of convolution module, the XNOR operation is performed, namely the calculation results calculated by the plurality of convolution modules N are subjected to XNOR operation integrally.
And 25, performing second revision on the parameters in the first revised initial classification model according to the test data, and recording the second revised initial classification model as a power spectrogram classification model.
It should be noted that the construction process and the testing process of the initial classification model are similar to the training process, and are not described herein again.
And 3, acquiring a power spectrogram to be detected of the equipment to be detected, identifying abnormal power of the power spectrogram to be detected according to the power spectrogram classification model, and calculating the coordinates of an abnormal sound source of the equipment to be detected by adopting a fusion algorithm when the abnormal power is identified.
In the embodiment, the three-element microphone array is used as a sound collecting unit, so that the three-dimensional coordinate positioning of the abnormal sound source is realized.
And inputting the power spectrogram to be detected into a previously trained model (power spectrogram classification model), and outputting a sound type (normal or abnormal) obtained after the three microphones (sound sensors) receive the sound and a predicted power value in real time.
Judging through sound types, if any classification result of the outputs of the three sound sensor channels is normal, judging the sound to be in a normal state, and not processing; otherwise, the obtained three-channel power value is used as an abnormal power characteristic value for subsequent processing, and after the abnormal power characteristic value is obtained, the time delay data of the sound sensor is fused, and the coordinate calculation of the abnormal sound source is performed on the time delay data, as shown in fig. 5.
The step 3 specifically includes:
and step 31, acquiring a power spectrogram to be detected of the equipment to be detected by using a sound acquisition unit, identifying abnormal power of the power spectrogram to be detected according to a power spectrogram classification model, and when it is judged that power values obtained in the power spectrogram to be detected corresponding to the three sound sensor channels exceed a preset safety threshold, taking the obtained power values as abnormal power characteristic values of the power spectrogram to be detected.
The predicted power value obtained in this embodiment is a predicted value obtained by performing image recognition on the power spectrogram to be detected through a power spectrogram classification model, and the prediction process is not limited in this embodiment.
Step 32, calculating the distance between the abnormal sound source and the sound sensor according to the time between two adjacent maximum power values in the power spectrogram to be detected acquired by any two sound sensors and the abnormal power characteristic value, wherein the calculation formula of the distance is as follows:
Figure BDA0002713142560000101
in the formula, WuThe abnormal power characteristic value of the power spectrogram to be detected, which is acquired by the u-th sound sensor, u is 1, …, h, …, j, h and j are the labels of the sound sensors, a is a preset precision coefficient, and r is a preset precision coefficientuIs the distance between the abnormal sound source and the u-th sound sensor, VsIs the speed of propagation of sound in air, ThThe time difference between two adjacent maximum power values corresponding to the h-th sound sensor is obtained.
Specifically, the calculation formula of the distance from the abnormal sound source to any microphone (sound sensor) in the ternary microphone array can be obtained according to the corresponding relation between the distance from the abnormal sound source and the power, and is as follows:
Figure BDA0002713142560000102
wherein b is a proportionality coefficient, i.e. distance ruAnd power
Figure BDA0002713142560000103
Proportional, therefore, for a ternary microphone array, the ratio of the distances of the abnormal sound source to the three microphones is:
Figure BDA0002713142560000104
further, since each microphone (sound sensor) in the ternary microphone array is located at a different position on the electric device, the propagation duration of the abnormal sound source to reach each microphone is also different.
Setting the microphones h and j, the distance difference from the abnormal sound source to the microphones h and j is determined by the propagation delay of the abnormal sound source, i.e. rh-rj＝Vs*|Th-Tj|。
In calculating the propagation delay, the time for the abnormal sound source to reach the corresponding sound sensor needs to be determined respectively. For each sound sensor, on the same frequency, the time difference between two adjacent maximum power values is the time for the abnormal sound source to reach the sound sensor, so that the propagation delay of the abnormal sound source can be determined through the corresponding power spectrogram of the sound sensor.
And step 33, calculating the abnormal sound source coordinate of the abnormal sound source according to the distance and the space coordinate of the sound sensor.
Setting the spatial coordinates of three microphones in the ternary microphone array as follows in sequence: (0, -y)1,0)、(0,y1,0)、(x10,0), the abnormal sound source coordinates of the abnormal sound source are: (x, y, z), then according to the position relationship between the three-dimensional vectors, the vector relationship can be obtained as follows:
Figure BDA0002713142560000111
therefore, the abnormal sound source coordinates (x, y, z) of the abnormal sound source can be calculated.
On the basis of the foregoing embodiments, this embodiment further illustrates an apparatus for positioning an abnormal sound source of an electric transmission and transformation device, where the apparatus includes a sound collection unit, a processor, and a memory, where the memory stores a computer program, and when the processor executes the computer program stored in the memory, the method for positioning an abnormal sound source of an electric transmission and transformation device based on improved EfficientNet in the foregoing embodiments is implemented to perform abnormal sound source positioning on a device to be detected.
Through testing of the abnormal sound source positioning method and device, the strong modeling and fitting capability of the neural network can be fully utilized, the input sound flow is accurately classified into normal and abnormal, the abnormal power characteristic value of the abnormal sound source is predicted and captured, and the time delay information is fused to calculate the azimuth coordinate, so that the real-time detection and positioning of the abnormal sound source of the power transmission and transformation equipment are realized.
The experimental results show that: the test model and the sound processing software are deployed in the front-end embedded equipment, abnormal sounds emitted by sound source devices such as transformers can be captured quickly and accurately, tracking and positioning are carried out, intelligent real-time detection and positioning of abnormal sound sources of the power transmission and transformation equipment are achieved, and the test model and the sound processing software have good performance and popularization and application prospects in positioning of abnormal sound sources in indoor and outdoor environments.
The technical scheme of the application is described in detail in the above with reference to the accompanying drawings, and the application provides a method for positioning an abnormal sound source of power transmission and transformation equipment based on improved EfficientNet, which comprises the following steps: acquiring a sample power spectrogram in the operation process of the electrical equipment by using a sound acquisition unit, and carrying out abnormal power value labeling on an abnormal power spectrogram in the sample power spectrogram; constructing an initial classification model according to the convolutional neural network, training the initial classification model by using a sample power spectrogram, and recording the trained initial classification model as a power spectrogram classification model; acquiring a power spectrogram to be detected of equipment to be detected, identifying abnormal power of the power spectrogram to be detected according to the power spectrogram classification model, and calculating the coordinates of an abnormal sound source of the equipment to be detected by adopting a fusion algorithm when identifying the abnormal power. According to the technical scheme, the three-channel abnormal power characteristic value is combined, the coordinate of the abnormal sound source is calculated according to the propagation delay between the microphone and the abnormal sound source, and the abnormal sound source positioning in the operation process of the electric equipment is completed.
The steps in the present application may be sequentially adjusted, combined, and subtracted according to actual requirements.
The units in the device can be merged, divided and deleted according to actual requirements.
Although the present application has been disclosed in detail with reference to the accompanying drawings, it is to be understood that such description is merely illustrative and not restrictive of the application of the present application. The scope of the present application is defined by the appended claims and may include various modifications, adaptations, and equivalents of the invention without departing from the scope and spirit of the application.

Claims (6)

1.一种基于改进EfficientNet的输变电设备异常声源定位方法，其特征在于，该方法包括：1. a power transmission and transformation equipment abnormal sound source localization method based on improving EfficientNet, is characterized in that, this method comprises: 步骤1，利用声音采集单元，采集电力设备运行过程中的样本功率谱图，对所述样本功率谱图中的异常功率谱图进行异常功率值标注，其中，所述样本功率谱图包括正常功率谱图以及所述异常功率谱图；Step 1: Use a sound acquisition unit to collect a sample power spectrogram during the operation of the power equipment, and mark the abnormal power value of the abnormal power spectrogram in the sample power spectrogram, wherein the sample power spectrogram includes the normal power a spectrogram and the abnormal power spectrogram; 步骤2，根据所述卷积神经网络和EfficientNet网络，引入复合系数，构建初始分类模型，并利用所述样本功率谱图对所述初始分类模型进行训练，将训练后的初始分类模型记作功率谱图分类模型；Step 2, according to described convolutional neural network and EfficientNet network, introduce compound coefficient, build initial classification model, and utilize described sample power spectrum to carry out training to described initial classification model, the initial classification model after training is recorded as power Spectral classification model; 步骤3，采集待检设备的待检功率谱图，根据所述功率谱图分类模型对所述待检功率谱图进行异常功率识别，当识别出异常功率时，采用融合算法，计算所述待检设备的异常声源坐标。Step 3: Collect the power spectrum to be tested of the device to be tested, and perform abnormal power identification on the power spectrum to be tested according to the power spectrum classification model. When abnormal power is identified, a fusion algorithm is used to calculate the power spectrum to be tested. Check the abnormal sound source coordinates of the equipment. 2.如权利要求1所述的基于改进EfficientNet的输变电设备异常声源定位方法，其特征在于，所述初始分类模型包括初始卷积层，所述初始卷积层用于对输入数据进行卷积及标准化计算，其特征在于，2. the abnormal sound source localization method of power transmission and transformation equipment based on improved EfficientNet as claimed in claim 1, is characterized in that, described initial classification model comprises initial convolutional layer, and described initial convolutional layer is used for input data. Convolution and normalization calculation, characterized in that, 所述步骤2中，利用所述样本功率谱图对所述初始分类模型进行训练，具体包括：In the step 2, the initial classification model is trained by using the sample power spectrum, which specifically includes: 步骤21，将所述样本功率谱图随机分为训练数据和测试数据；Step 21, randomly dividing the sample power spectrogram into training data and test data; 步骤22，所述训练数据输入至所述初始卷积层，得到标准化输入图像；Step 22, the training data is input into the initial convolution layer to obtain a standardized input image; 步骤23，对所述标准化输入图像进行多层卷积模块计算，并根据所述复合系数对每一层网络计算出的中间值进行数据增强，其中，所述复合系数的计算公式为：Step 23, performing multi-layer convolution module calculation on the standardized input image, and performing data enhancement on the intermediate value calculated by each layer of the network according to the composite coefficient, wherein the calculation formula of the composite coefficient is:
Figure FDA0002713142550000011
Figure FDA0002713142550000011
 式中，w为当前网络的宽度，d为当前网络的深度，(h,p)为标准化输入图像的分辨率，r是分辨率倍率，α是当前网络的学习率；In the formula, w is the width of the current network, d is the depth of the current network, (h, p) is the resolution of the normalized input image, r is the resolution magnification, and α is the learning rate of the current network; 步骤24，根据多层卷积模块计算的计算结果，对所述初始分类模型中的参数进行第一修订，其中，所述参数包括所述学习率α；Step 24, according to the calculation result calculated by the multi-layer convolution module, perform a first revision to the parameters in the initial classification model, wherein the parameters include the learning rate α; 步骤25，根据所述测试数据，对所述第一修订后的初始分类模型中的参数进行第二修订，将所述第二修订后的初始分类模型记作所述功率谱图分类模型。Step 25: Perform a second revision on the parameters in the first revised initial classification model according to the test data, and record the second revised initial classification model as the power spectrum classification model. 3.如权利要求2所述的基于改进EfficientNet的输变电设备异常声源定位方法，其特征在于，所述参数的修订公式为：3. the abnormal sound source localization method of power transmission and transformation equipment based on improved EfficientNet as claimed in claim 2, is characterized in that, the revision formula of described parameter is:
Figure FDA0002713142550000021
Figure FDA0002713142550000021
Figure FDA0002713142550000022
Figure FDA0002713142550000022
 Memory(N)≤targetmemory Memory(N)≤target memory FLOPS(N)≤target_flopsFLOPS(N)≤target_flops 式中，
Figure FDA0002713142550000023
为所述初始分类模型中第i个卷积模块，Li表示第i个卷积模块block重复的次数，
Figure FDA0002713142550000024
是对Li进行求导计算，
Figure FDA0002713142550000025
表示第i个卷积模块block，由卷积层Fi重复
Figure FDA0002713142550000026
次构成，再将结果进行求导计算，i＝1,2,…,m，⊙为同或运算，targetmemory为资源消耗目标值，target_flops为计算量消耗目标值，Memory(·)为资源消耗函数，FLOPS(·)为计算量消耗函数，N(·)为一层卷积模块计算的计算结果，Accuracy(·)是在约束条件内使网络的精度最高的约束函数。In the formula,
Figure FDA0002713142550000023
is the ith convolution module in the initial classification model, Li represents the number of repetitions of the ith convolution module block,
Figure FDA0002713142550000024
is the derivative calculation of Li,
Figure FDA0002713142550000025
Represents the i-th convolutional module block, repeated by the convolutional layer F i
Figure FDA0002713142550000026
Then, the result is calculated by derivation, i=1,2,...,m, ⊙ is the same OR operation, target memory is the target value of resource consumption, target_flops is the target value of calculation consumption, and Memory( ) is the resource consumption function, FLOPS(·) is the calculation consumption function, N(·) is the calculation result calculated by the one-layer convolution module, and Accuracy(·) is the constraint function that maximizes the accuracy of the network within the constraints. 4.如权利要求1所述的基于改进EfficientNet的输变电设备异常声源定位方法，其特征在于，所述步骤3，具体包括：4. the abnormal sound source localization method of power transmission and transformation equipment based on improved EfficientNet as claimed in claim 1, is characterized in that, described step 3, specifically comprises: 步骤31，利用所述声音采集单元采集待检设备的待检功率谱图，根据所述功率谱图分类模型对所述待检功率谱图进行异常功率识别，生成所述待检功率谱图的异常功率特征值，其中，所述声音采集单元至少包括3个声音传感器；Step 31: Use the sound acquisition unit to collect the power spectrum to be tested of the device to be tested, perform abnormal power identification on the power spectrum to be tested according to the power spectrum classification model, and generate a power spectrum of the power spectrum to be tested. Abnormal power characteristic value, wherein, the sound acquisition unit includes at least 3 sound sensors; 步骤32，根据任意两个所述声音传感器采集到的所述待检功率谱图中相邻两个最大功率值间的时间、以及所述异常功率特征值，计算异常声源与所述声音传感器之间的间距；Step 32: Calculate the abnormal sound source and the sound sensor according to the time between two adjacent maximum power values in the power spectrum to be checked collected by any two of the sound sensors and the abnormal power characteristic value. spacing between; 步骤33，根据所述间距、以及所述声音传感器的空间坐标，计算所述异常声源的异常声源坐标。Step 33: Calculate the coordinates of the abnormal sound source of the abnormal sound source according to the distance and the spatial coordinates of the sound sensor. 5.如权利要求4所述的基于改进EfficientNet的输变电设备异常声源定位方法，其特征在于，所述间距的计算公式为：5. the abnormal sound source localization method of power transmission and transformation equipment based on improved EfficientNet as claimed in claim 4, is characterized in that, the calculating formula of described spacing is:
Figure FDA0002713142550000031
Figure FDA0002713142550000031
 式中，Wu为第u个声音传感器采集到的待检功率谱图的异常功率特征值，u＝1,…,h,…,j，h、j为声音传感器的标号，a为预设精度系数，ru为所述异常声源与第u个声音传感器之间的间距，Vs为空气中声音的传播速度，Th为第h个声音传感器对应的相邻两个最大功率值间的时间差。In the formula, Wu is the abnormal power characteristic value of the power spectrum to be detected collected by the uth sound sensor, u=1,...,h,...,j, h, j are the labels of the sound sensor, a is the preset Coefficient of accuracy, r u is the distance between the abnormal sound source and the u th sound sensor, V s is the propagation speed of sound in the air, Th is the distance between the two adjacent maximum power values corresponding to the h th sound sensor time difference. 6.一种输变电设备异常声源定位的装置，其特征在于，该装置包括声音采集单元、处理器以及存储器，所述存储器中存储有计算机程序，所述处理器执行所述存储器中存储有计算机程序时实现如权利要求1至6中任一项所述的基于改进EfficientNet的输变电设备异常声源定位方法，以对待检设备进行异常声源定位。6. A device for locating abnormal sound sources of power transmission and transformation equipment, characterized in that the device comprises a sound collection unit, a processor and a memory, wherein a computer program is stored in the memory, and the processor executes the memory stored in the memory. When there is a computer program, the method for locating abnormal sound sources of power transmission and transformation equipment based on improved EfficientNet according to any one of claims 1 to 6 is realized, so as to locate abnormal sound sources of the equipment to be inspected.
CN202011063845.7A 2020-09-30 2020-09-30 Power transmission and transformation equipment abnormal sound source positioning method based on improvement EFFICIENTNET Active CN112348052B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011063845.7A CN112348052B (en) 2020-09-30 2020-09-30 Power transmission and transformation equipment abnormal sound source positioning method based on improvement EFFICIENTNET

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011063845.7A CN112348052B (en) 2020-09-30 2020-09-30 Power transmission and transformation equipment abnormal sound source positioning method based on improvement EFFICIENTNET

Publications (2)

Publication Number Publication Date
CN112348052A true CN112348052A (en) 2021-02-09
CN112348052B CN112348052B (en) 2024-09-06

Family

ID=74360384

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011063845.7A Active CN112348052B (en) 2020-09-30 2020-09-30 Power transmission and transformation equipment abnormal sound source positioning method based on improvement EFFICIENTNET

Country Status (1)

Country Link
CN (1) CN112348052B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113379005A (en) * 2021-08-12 2021-09-10 新风光电子科技股份有限公司 Intelligent energy management system and method for power grid power equipment
CN114325214A (en) * 2021-11-18 2022-04-12 国网辽宁省电力有限公司电力科学研究院 An on-line monitoring method of electric power based on microphone array sound source localization technology
US20230104123A1 (en) * 2021-10-06 2023-04-06 Samsung Electronics Co., Ltd. Method and apparatus with abnormal channel of microphone array detection and compensation signal generation
CN116229036A (en) * 2023-05-06 2023-06-06 山东国研自动化有限公司 Data acquisition system, method, computer device and storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002323371A (en) * 2001-04-27 2002-11-08 Toshiba Corp Sound diagnostic device and sound diagnostic method
CN106021789A (en) * 2016-06-01 2016-10-12 北京交通大学 Fuzzy-intelligence-based rail car suspension system fault classification method and system
CN110705456A (en) * 2019-09-29 2020-01-17 华南理工大学 Micro motor abnormity detection method based on transfer learning

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002323371A (en) * 2001-04-27 2002-11-08 Toshiba Corp Sound diagnostic device and sound diagnostic method
CN106021789A (en) * 2016-06-01 2016-10-12 北京交通大学 Fuzzy-intelligence-based rail car suspension system fault classification method and system
CN110705456A (en) * 2019-09-29 2020-01-17 华南理工大学 Micro motor abnormity detection method based on transfer learning

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113379005A (en) * 2021-08-12 2021-09-10 新风光电子科技股份有限公司 Intelligent energy management system and method for power grid power equipment
US20230104123A1 (en) * 2021-10-06 2023-04-06 Samsung Electronics Co., Ltd. Method and apparatus with abnormal channel of microphone array detection and compensation signal generation
CN114325214A (en) * 2021-11-18 2022-04-12 国网辽宁省电力有限公司电力科学研究院 An on-line monitoring method of electric power based on microphone array sound source localization technology
CN116229036A (en) * 2023-05-06 2023-06-06 山东国研自动化有限公司 Data acquisition system, method, computer device and storage medium

Also Published As

Publication number Publication date
CN112348052B (en) 2024-09-06

Similar Documents

Publication Publication Date Title
CN112348052B (en) Power transmission and transformation equipment abnormal sound source positioning method based on improvement EFFICIENTNET
CN112183647B (en) Method for detecting and positioning sound faults of substation equipment based on deep learning
CN104407328B (en) Closing space sound localization method based on space impulse response coupling and system
KR102509699B1 (en) Sound-based Multi-device Operation Monitoring Method and System for the Same
CN115758200A (en) Vibration signal fault identification method and system based on similarity measurement
CN114487129B (en) Flexible material damage identification method based on acoustic emission technology
CN117289067B (en) Transformer running state on-line monitoring system
CN114297912B (en) Cutter abrasion prediction method based on deep learning
CN113671363A (en) High-voltage circuit breaker state identification system and method
CN113156278A (en) CNN network-based GIS partial discharge positioning method
CN111859241B (en) An Unsupervised Sound Source Orientation Method Based on Acoustic Transfer Function Learning
CN117975372A (en) A construction site safety detection system and method based on the combination of YOLOv8 and Transformer encoder
CN112711873A (en) Method, system and medium for carrying out damage positioning based on lamb wave and neural network
CN111076960A (en) Voiceprint quality detection method based on artificial intelligence algorithm
CN118552530A (en) A defect detection method and system for automobile parts based on visual inspection
CN115101149B (en) Method for predicting total energy of microstructure of material
CN117332352A (en) A lightning arrester signal defect identification method based on BAM-AlexNet
CN116259313A (en) A sound event localization and detection method based on temporal convolutional network
Zhang et al. A task-specific meta-learning framework for few-shot sound event detection
Zhang et al. Diagnosis model of noise-type defects for dry-type transformer based on time–frequency-space tensors and improved prototypical network under small sample conditions
CN112183745A (en) High-voltage cable partial discharge mode identification method based on particle swarm algorithm and DBN
CN111721399A (en) A kind of hydraulic building structure vibration test system and test method
US20230152187A1 (en) Fault signal locating and identifying method of industrial equipment based on microphone array
CN116989593B (en) Monitoring and early warning method for operation faults of electric furnace
CN119091924B (en) A Noise Area Detection and Identification Method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant