A power master device voiceprint defect identification method and system

Abstract

The present application relates to the technical field of power equipment monitoring, and provides a power main equipment voiceprint defect identification method and system, which comprises: when the data entropy value in the feature vector library is lower than a second threshold value, defect identification is realized based on voiceprint data through joint implementation of a deep learning model and a mechanism model, a first defect identification result is obtained, and the deep learning model is optimized based on the first defect identification result; meanwhile, feature extraction is performed on the voiceprint data, vector retrieval is performed in the feature vector library, a second defect identification result is obtained, and if the identification results are inconsistent, the feature vector is added to the feature vector library; when the data entropy value in the feature vector library is higher than the second threshold value, vector retrieval is performed in the feature vector library, a second defect identification result is obtained, and the second defect identification result is taken as a final defect identification result. Not only all potential operating states of the power main equipment are basically covered, but also rapid diagnosis can be achieved, faults can be found in time, and the stable operation of the power grid equipment is beneficial.

Description

Technical Field

[0001] This invention belongs to the field of power equipment monitoring technology, and in particular relates to a method and system for identifying acoustic defects in main power equipment. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Online monitoring of power equipment can improve the timeliness of acquiring critical status parameters, facilitate the early detection of potential equipment problems, enhance power supply reliability, enable intelligent management and optimization of power grid operation, reduce the operating costs of power companies, and shorten fault handling time.

[0004] Although acoustic signature defect recognition technology is also incorporated into the overall architecture of intelligent inspection, there is currently insufficient attention and development regarding this technology for identifying acoustic signature defects inside transformers, resulting in slow application of corresponding defect identification and analysis techniques. Therefore, there is an urgent need to research and apply key technologies for intelligent inspection, such as acoustic signature defect recognition of transformers.

[0005] Existing voiceprint defect recognition technology has the following drawbacks:

[0006] (1) Most existing studies focus on a certain type of defect in transformers, and mainly concentrate on the diagnosis and identification of mechanical or insulation defects, such as winding loosening and deformation, DC bias, or insulation discharge. The identification results of the constructed diagnostic models do not cover all potential operating states of the main equipment of the substation, which to some extent limits the field application of voiceprint diagnostic technology.

[0007] (2) Most of the transformer acoustic fingerprint sample data comes from test results. The number of samples is insufficient and the environmental noise of the main substation equipment and the interference and influence of other power transmission and transformation equipment are not considered. The limitations of acoustic fingerprint samples will increase the misjudgment rate of the diagnostic results to a certain extent.

[0008] (3) Due to the large amount of data in the substation scenario, the model training is time-consuming, the data on the defect updates of each substation cannot be fed back in time, the training library is difficult to cover various situations of the main equipment, and the model cannot be updated and iterated in time. Summary of the Invention

[0009] To address the technical problems mentioned above, this invention provides a method and system for identifying voiceprint defects in power main equipment. After determining that the equipment to be identified is abnormal, the defect identification is achieved by combining a deep learning model and a mechanism model. This method basically covers all potential operating states of substation main equipment. Moreover, as the number of features covered in the feature vector library gradually increases, the voiceprint model detection is gradually discarded, and similarity matching is performed directly from the vector library. This enables rapid diagnosis and timely fault detection, which is beneficial to the stable operation of power grid equipment.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] The first aspect of the present invention provides a method for identifying acoustic defects in main power equipment, comprising:

[0012] After the device to be identified is determined to be abnormal, the voiceprint data of the device to be identified is obtained;

[0013] When the entropy value of the data in the feature vector library is lower than the second threshold, defect identification is achieved by combining a deep learning model and a mechanism model based on the voiceprint data to obtain the first defect identification result. The first defect identification result is used as the final defect identification result, and the deep learning model is optimized based on the first defect identification result. At the same time, feature extraction is performed on the voiceprint data to obtain feature vectors. Vector retrieval is performed in the feature vector library to obtain the second defect identification result. If the first defect identification result and the second defect identification result are inconsistent, the feature vector is added to the feature vector library.

[0014] When the entropy value of the data in the feature vector library is higher than the second threshold, feature extraction is performed on the voiceprint data to obtain feature vectors. Vector retrieval is then performed in the feature vector library to obtain the second defect identification result, which is then used as the final defect identification result.

[0015] Further, the anomaly determination step of the device to be identified includes: acquiring the voiceprint features and conditional features of the device to be detected, and obtaining the input features of the device to be detected through feature combination; inputting the input features of the device to be detected into an inference network, the inference network including a generator encoder and a generator decoder and an anomaly score calculation network respectively connected to the generator encoder; the generator encoder processes the input features of the device to be detected to obtain latent vectors; the generator decoder processes the latent vectors to obtain reconstructed features; the anomaly score calculation network processes the latent vectors to obtain an anomaly score; the reconstruction loss is calculated based on the latent vectors and the reconstructed features; the weighted sum of the reconstruction loss and the anomaly score is used as the risk value of the device to be detected; when the risk value of the device to be detected exceeds the risk threshold, it is determined to be an anomaly.

[0016] Furthermore, when the entropy value of the data in the feature vector library is lower than the first threshold, the weight of the mechanistic model is greater than the weight of the deep learning model, and vice versa.

[0017] Furthermore, when the entropy value of the data in the feature vector library is higher than the second threshold, if the similarity between the feature vector and the data in the feature vector library is lower than the set value during vector retrieval, then defect identification is re-implemented through a combination of deep learning model and mechanism model, and the feature vector is added to the feature vector library.

[0018] Furthermore, the feature extraction step includes: performing signal processing on the audio signal image in the voiceprint data to obtain a time-frequency domain spectrogram, inputting the time-frequency domain spectrogram into the feature extraction network, and outputting a feature vector of fixed dimensions.

[0019] Furthermore, the mechanism model identifies defects based on spectrogram features, which include the transformer body acoustic waveform signal-to-noise ratio, sound pressure level, odd-even harmonic ratio, high-frequency energy ratio, transformer spectral entropy, and resonant demodulation power frequency correlation.

[0020] Furthermore, the transformer body acoustic signature signal-to-noise ratio is used to determine if there is interference. If interference is determined to exist based on the transformer body acoustic signature signal-to-noise ratio, the first defect identification result is output as interference signal, and the mechanism model is no longer used.

[0021] The sound pressure level is used to identify heavy load faults and / or sensor anomalies.

[0022] The odd-even harmonic ratio is used to determine DC bias or harmonic current faults.

[0023] The high-frequency energy ratio is used to determine loosening-type faults;

[0024] The transformer spectrum entropy is used to diagnose abnormal noises in the cooler.

[0025] The resonant demodulation power frequency correlation is used for partial discharge fault diagnosis.

[0026] A second aspect of the present invention provides a power main equipment acoustic signature defect identification system, comprising:

[0027] The data acquisition module is configured to acquire the voiceprint data of the device to be identified after the device is determined to be abnormal.

[0028] The first identification module is configured to: when the entropy value of the data in the feature vector library is lower than the second threshold, identify defects based on the voiceprint data by jointly using a deep learning model and a mechanism model to obtain a first defect identification result, use the first defect identification result as the final defect identification result, and optimize the deep learning model based on the first defect identification result; at the same time, extract features from the voiceprint data to obtain feature vectors, perform vector retrieval in the feature vector library to obtain a second defect identification result; if the first defect identification result and the second defect identification result are inconsistent, add the feature vector to the feature vector library.

[0029] The second identification module is configured to: extract features from the voiceprint data when the entropy value of the data in the feature vector library is higher than the second threshold, obtain feature vectors, perform vector retrieval in the feature vector library, obtain the second defect identification result, and use the second defect identification result as the final defect identification result.

[0030] Furthermore, when the entropy value of the data in the feature vector library is lower than the first threshold, the weight of the mechanistic model is greater than the weight of the deep learning model, and vice versa.

[0031] Furthermore, when the entropy value of the data in the feature vector library is higher than the second threshold, if the similarity between the feature vector and the data in the feature vector library is lower than the set value during vector retrieval, then defect identification is re-implemented through a combination of deep learning model and mechanism model, and the feature vector is added to the feature vector library.

[0032] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for identifying acoustic defects in power main equipment.

[0033] A fourth aspect of the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method for identifying voiceprint defects in main power equipment.

[0034] A fifth aspect of the present invention provides a computer program product, which is a computer program product containing a computer program, wherein when a processor executes the computer program, it implements the steps in the above-described method for identifying acoustic defects in main power equipment.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] This invention innovatively proposes an incremental learning method for substation inspection image recognition. By constructing a local acoustic print defect database, a feature vector library, a main equipment acoustic print defect model, and a vector retrieval module, a lightweight incremental learning acoustic print image recognition method is developed. Without the participation of old samples, it can identify and detect new equipment defect categories. When local defects are updated, only the local acoustic print defect image database needs to be updated, without retraining the model. It can achieve "addition and use immediately", solving the problems of slow model training and update and local feature library upgrades and updates, and realizing rapid and accurate diagnosis of main equipment defects and faults.

[0037] This invention innovatively proposes a multi-model defect identification method for power equipment that integrates a voiceprint mechanism model and a deep learning model. The mechanism model is trained by screening spectral features in the low-frequency band, and the optimized deep learning model is trained using real-time data. The two results are combined to achieve accurate identification of abnormal voiceprints, solving the problem of easy misjudgment in power equipment defect identification, realizing accurate identification of abnormal voiceprints in power equipment, and improving the robustness of the model to equipment differences and environmental noise.

[0038] This invention innovatively proposes a method for detecting abnormal voiceprints in power main equipment. It reconstructs the original features using a large number of normal voiceprint samples and a small number of abnormal voiceprint samples. The degree of abnormality of the main equipment voiceprint is judged by the sample reconstruction loss. This solves the problem of low accuracy and insufficient generalization of deep learning models caused by insufficient voiceprint fault samples. It realizes the detection of abnormal signals under the condition of few samples and effectively improves the accuracy of voiceprint anomaly detection. Attached Figure Description

[0039] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0040] Figure 1 This is a flowchart illustrating the joint implementation of defect identification using a degree learning model and a mechanism model in Embodiment 1 of the present invention.

[0041] Figure 2 This is a flowchart of incremental learning in Embodiment 1 of the present invention;

[0042] Figure 3 This is a structural diagram of the feature extraction network in Embodiment 1 of the present invention. Detailed Implementation

[0043] To enable those skilled in the art to better understand the technical solutions in this application, the following will further illustrate the technical solutions for achieving the inventive objectives of this application through several specific embodiments. It should be noted that the technical solutions claimed in this application include, but are not limited to, the following embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this application.

[0044] Example 1

[0045] This embodiment provides a method for identifying acoustic defects in main power equipment.

[0046] This embodiment provides a method for identifying acoustic signature defects in main power equipment, the objectives of which include:

[0047] (1) Incremental learning: As the feature vector library gradually covers more features, the voiceprint model detection is gradually abandoned, and similarity matching is directly performed from the vector library, thereby achieving rapid diagnosis and timely detection of faults, which is conducive to the stable operation of power grid equipment. Moreover, only the local voiceprint defect image database needs to be updated, and there is no need to retrain the model. It can achieve "add and use immediately", solving the problems of slow model training and updating and local feature library upgrade and update.

[0048] (2) More accurate model identification: Defect identification is achieved through the joint use of deep learning models and mechanistic models, basically covering all potential operating states of substation main equipment. Moreover, considering the limited number of cases learned by the early deep learning model, more weight can be appropriately given to the mechanistic model. As the dataset becomes richer and expanded, the weight of the mechanistic model can be gradually reduced or eliminated, ultimately improving the model identification accuracy.

[0049] (3) Higher accuracy in detecting abnormal soundprints: Based on the latent vector and the reconstruction loss obtained from the reconstruction features, the degree of abnormality of the transformer soundprint is judged, realizing the detection of abnormal signals under the condition of few samples. By combining the abnormal probability score and the reconstruction loss, the quantitative evaluation of the risk status of the transformer soundprint is realized, which effectively improves the accuracy of detecting abnormal soundprints. After identifying abnormal soundprint data based on a large number of positive samples, the abnormal data is further identified by the multi-model feature fusion defect identification technology, which can identify the specific defect category.

[0050] This embodiment provides a method for identifying acoustic fingerprint defects in main power equipment, comprising three parts. First, it employs anomaly detection technology for acoustic fingerprint data, which identifies abnormal acoustic fingerprint data based on a large number of positive samples. Then, it uses multi-model feature fusion defect identification technology to further identify the abnormal data, thereby identifying specific defect categories. Furthermore, for newly emerging equipment defect categories, a self-evolving detection network is proposed, which can achieve the identification and detection of new equipment defect categories without the participation of old samples.

[0051] Step 1: Obtain the voiceprint features (voiceprint data) and conditional features of the device under test, and determine whether the device under test is abnormal.

[0052] As one implementation method, the voiceprint features and conditional features of the device to be detected are input into a pre-trained convolutional neural network to obtain the result of whether the device to be detected is abnormal.

[0053] As another implementation, the voiceprint features and conditional features of the device under test are obtained, and the input features of the device under test are obtained by feature combination. The input features of the device under test are input into an inference network, which includes a generator encoder and a generator decoder and an anomaly score calculation network connected to the generator encoder. The generator encoder processes the input features of the transformer under test to obtain latent vectors, the generator decoder processes the latent vectors to obtain reconstructed features, the anomaly score calculation network processes the latent vectors to obtain anomaly scores, and the reconstruction loss is calculated based on the latent vectors and reconstructed features. The weighted sum of the reconstruction loss and the anomaly score is used as the risk value of the device under test. When the risk value of the device under test exceeds the risk threshold, it is determined to be an anomaly.

[0054] Specifically, the voiceprint features s and conditional features c of the device to be detected are combined into input features x = (s, c), which are then processed by the generator encoder G.E The latent vector z is obtained and then processed by the generator decoder G. D Obtain the reconstructed features of x After passing through an additional encoder E, the reconstructed features of the latent vectors are obtained. Furthermore, combined features x and reconstruction During the training phase, the input is fed into the discriminator D. The discriminator learns continuously and strives to be as "unfoolproof" as possible. The discriminator and the generator (encoder G)... E and decoder G D (Combined) Through continuous adversarial training, the optimal solution is sought.

[0055] The first training phase process is as follows:

[0056] (1) Feature Combination: The acoustic signature feature s of the device under test is derived from the acoustic signature data of normal devices under test, such as MFCC, time-frequency diagram, Mel spectrum, etc. The conditional feature c includes the voltage level (110kV, 220kV, 500kV, 1000kV), cooling method (oil-immersed air cooling, forced oil circulation air cooling, forced guided oil circulation air cooling), manufacturer and affiliated site of the device under test. All conditional features are combined and one-hot encoded. Then the acoustic signature feature and conditional feature are combined into the input feature x = (s, c).

[0057] (2) Generate encoder G E The input feature x is processed by the generative encoder to obtain the latent vector z = G. E (x,c), generate encoder G E This includes convolutional layers, layer normalization, activation functions, and Dropout, which compress the dimensionality of the input feature x to represent its latent features. Layer normalization is used here to address the issue of batch normalization being susceptible to variations in the distribution of data within a batch. Dropout is only used during training; during testing or model evaluation, the probability should be set to 0 to disable Dropout.

[0058] (3) Generate decoder G D Implicit vector z = G E (x,c) is generated by decoder G D Then, after scaling up, the reconstructed feature vector is obtained. Generate decoder G D This includes transposed convolutional layers, normalized layers, and activation function layers. When When the difference between x and is small, it means that the input feature x can be approximately reconstructed. Therefore, transformer acoustic signature anomaly detection is achieved by calculating the error between the two. The reconstruction loss is defined using the Smooth L1 loss function as follows:

[0059]

[0060] in The i-th element represents the feature.

[0061] (4) Encoder E: In order to better learn the latent feature distribution of normal voiceprint data, the generated features After being encoded again by encoder E, the reconstructed latent vector is obtained. When this vector matches the latent vector z, it indicates a reconstructed feature. The latent distribution features are approximately similar to the latent distribution features of the input feature x. The encoding loss is still defined using the Smooth L1 loss function as follows:

[0062]

[0063] in This represents the i-th element of the feature. This method significantly reduces the time and space complexity of computation, improving the speed of fault detection. Encoders E and G E The structure is consistent.

[0064] (5) Discriminator network D.

[0065] The discriminator D functions similarly to traditional generative adversarial networks. It takes the input features and conditional features (x,c) as inputs and reconstructs the features. The input discriminator D is passed through a transposed convolutional layer, a batch normalization layer, and a layer with a leaky linear rectified activation function to scale it down multiple times. The compressed vector is then flattened and input into a fully connected layer to achieve binary classification (determining whether the input is original feature data or generated feature data). The adversarial loss is defined as follows:

[0066]

[0067] Overall, the model's overall loss during training is as follows:

[0068] L = L adv +αL rec +βL enc

[0069] During the training phase, the model was trained using acoustic signature monitoring data from different voltage levels, different stations, and different cooling methods, with weights α = 1 and β = 0.1.

[0070] The risk threshold is calculated under the condition of minimizing loss:

[0071] R = L rec +λL enc

[0072] Where λ is an adjustable parameter that is selected based on the batch data.

[0073] During the inference phase, the trained generative encoder G is used. E Generate decoder G D The encoder E is used to calculate the reconstruction loss and encoding loss, and to calculate the risk R′ value of the data to be detected. When the risk R′ value of the data to be detected exceeds the risk threshold R obtained during training, it can be identified as abnormal data.

[0074] Step 2: For devices identified as abnormal, when the data entropy value in the feature vector library is lower than the second threshold, multi-model feature fusion defect identification is performed through the voiceprint detection model (deep learning model + mechanism model).

[0075] In this embodiment, a defect identification method driven by the sound generation mechanism and data fusion of main power equipment is proposed, such as... Figure 1 As shown, the mechanism model is first identified and diagnosed through feature extraction and other steps according to the diagnostic rules. The feature acquisition in the spectrum adopts a low-frequency band selection strategy to improve the accuracy of data features and reduce the redundancy of data. Then, the real-time collected data is sent to the deep learning model. The two results are combined to achieve accurate identification of abnormal sound patterns, improve the robustness of the model to equipment differences and environmental noise, and meet the requirements of complex substation environments and fault conditions.

[0076] Step 201: Based on the spectrogram in the voiceprint data, the mechanism model is identified and diagnosed through feature extraction and other steps according to the diagnostic rules.

[0077] (1) Feature extraction.

[0078] Referring to the feature calculation formula in Table 1, feature extraction was completed. Extraction was performed only at 50, 100, 150, and 200 Hz in the spectrum, ensuring effective features while achieving dimensionality reduction and minimizing data redundancy. Specific features extracted from the spectrum include the transformer's acoustic signature signal-to-noise ratio, sound pressure level, odd-even harmonic ratio, high-frequency energy proportion, and transformer spectral entropy.

[0079] Where RMS is the effective value of the voiceprint signal (frequency band above 20Hz), and base is the sound pressure level reference, specifically 20uPa. X i*50 This represents the audioprint signal spectrum, with the corresponding index being the i-th 50Hz frequency. Wherein, This represents the percentage of energy at the corresponding frequency.

[0080] Table 1. Calculation Formulas for Main Features

[0081]

[0082] (2) Diagnostic rules.

[0083] For specific diagnostic rules, please refer to the characteristic manifestations of different faults in Table 2.

[0084] 1) External interference.

[0085] Among them, the transformer body acoustic waveform signal-to-noise ratio is mainly used to eliminate the influence of short-term strong interference. When the transformer body acoustic waveform signal-to-noise ratio drops sharply, it often indicates that the sensor is damaged or the environment is excessively noisy for a short time (thunder, car horns, firecracker noise). At this time, the signal-to-noise ratio is low (that is, the transformer body acoustic waveform signal-to-noise ratio is lower than the threshold), the reliability of the diagnostic results is low, and it can be determined that there is an interference signal. If the interference signal continues to exist, it can be considered that there is a probability of sensor damage.

[0086] 2) Heavy load.

[0087] When the sound pressure level is too high, since the operating noise of the transformer is positively correlated with the load (more precisely, it changes linearly with the square of the load current), when the sound pressure level increases significantly (i.e., the sound pressure level exceeds the set value), it often indicates an overload or heavy load fault. If the transformer operates in an overload state for a long time, it is very easy to cause irreversible damage to the transformer.

[0088] 3) DC bias.

[0089] When the odd-even harmonic ratio of the transformer's acoustic signature increases significantly (i.e., the odd-even harmonic ratio exceeds the set value), it can generally be directly determined that there is a DC bias defect. At the same time, it may be accompanied by an increase in noise. In this case, the DC bias fault can be directly output.

[0090] 4) Loose.

[0091] When the energy in the high-frequency band of a transformer (800Hz-2000Hz) increases significantly (the proportion of high-frequency energy exceeds a certain value), it often indicates resonance or loosening faults. At this time, the fault may be outside the transformer.

[0092] 5) Other abnormalities.

[0093] When one or more spectrum features of transformer acoustic signature data show a significant increase or decrease compared to historical data, or when the acoustic signature features of a transformer differ significantly from those of other transformers of the same model and operating conditions, it often indicates the presence of some anomalies. These anomalies may not be fully covered by existing faults, but they should still be treated as an anomaly alarm.

[0094] Table 2. Main Features, Meanings, and Uses

[0095]

[0096] (3) Alarm strategy.

[0097] To reduce the possibility of false alarms in the model under special interference conditions, such as when there is equipment construction nearby and the noise is at the same frequency as the transformer, false alarms may occur. Therefore, this mechanism model will only output alarm information after the soundprint meets the alarm conditions and continuously alarms for the same fault several times.

[0098] Step 202: Input the real-time collected voiceprint waveform data into the deep learning model.

[0099] STFT calculation is performed on the real-time acquired acoustic waveform data, and the data is reduced in dimension to the same data format as the deep learning model training stage. The improved deep learning model is then called to perform fault prediction, and the type and probability of the fault predicted by the deep learning model with the highest probability are obtained.

[0100] Step 203: Use a fusion-driven approach to achieve accurate identification of abnormal voiceprints from the two results.

[0101] Based on the diagnostic results of the mechanism diagnostic model and the deep learning model, a fusion-driven approach is adopted, and certain weights are assigned to each model to form the final output result, realizing the final output of transformer voiceprint recognition. Considering that the learning cases of the deep learning model are limited in the early stage, more weight can be given to the mechanism model. As the dataset becomes richer and expanded, the weight of the mechanism model is gradually reduced or eliminated.

[0102] The specific fusion-driven process can be divided into the following five steps:

[0103] 1) The mechanism model is used to identify abnormal data of non-transformer faults. When the noise of the acoustic data is large and the signal-to-noise ratio of the acoustic data of the transformer body is low, it may be due to sensor failure, abnormal sensor installation, low data acquisition quality, strong interference, etc. In this case, it is not easy to enter the fault identification logic, and the acoustic signal can be directly output as an interference signal.

[0104] 2) When the voiceprint data is initially judged by the mechanistic model, it indicates that the voiceprint is likely a genuine transformer voiceprint. At this point, the collected voiceprint samples are classified using both the mechanistic model and the deep learning model. The results are output: the discrimination result and confidence probability of the mechanistic model, and the maximum probability corresponding to the class and its probability of the deep learning model's recognition result, each corresponding to a label. 机理 P 机理 label 深度学习 P 深度学习 ;

[0105] 3) Set reliability weights w for the results of the mechanistic model and the deep learning model respectively. 机理 w 深度学习The reliability can be subdivided into overall reliability or category reliability. Taking overall reliability as an example, it can be determined based on the performance of the mechanism model and deep learning model on the test samples and combined with certain expert experience.

[0106] 4) Calculate the final output result and compare the combined result P(P=max(P) of the model result reliability weights and the confidence probabilities of each model output category. 机理 *w 机理 ,P 深度学习 *w 深度学习 Determine the output category with the highest probability, and then output the final result.

[0107] 5) Update the reliability weights of the mechanistic model and deep learning model results. When there is new progress in the research of a certain type of failure mechanism, the reliability weight of the corresponding type of failure in the mechanistic model can be increased. When the number of accumulated samples reaches a certain requirement and the accuracy of the deep learning model is high, such as reaching more than 95%, the reliability weight of the deep learning model can be appropriately increased.

[0108] Given that the early datasets contained insufficient transformer fault samples, mostly data collected from the same transformer at different angles, the trained machine learning models were difficult to apply widely to the identification of other transformer defects. Therefore, models based on smaller weights, such as w, could be used instead. 深度学习 =0.3, while the mechanistic model, due to its adjustable parameters, is easier to set appropriate parameters based on horizontal and vertical comparison results, and its accuracy is relatively high. It is recommended to base it on a higher weight, such as w 机理 =0.7.

[0109] In this embodiment, when the data entropy value in the feature vector library is lower than the first threshold, w 机理 Greater than w 深度学习 Conversely, it is less than. The first threshold is less than the second threshold.

[0110] Step 204: Add the latest collected main device voiceprint image (i.e., voiceprint data) and the comprehensive results to the dataset (local voiceprint defect database) and retrain the deep learning model.

[0111] Step 3: When the entropy value of the data in the feature vector library is lower than the second threshold, both Step 2 and the incremental learning module are enabled; when the entropy value of the data in the feature vector library is higher than the second threshold, only the incremental learning module is enabled.

[0112] Specifically, when the entropy value of the data in the feature vector library is lower than the second threshold, defect identification is achieved based on the voiceprint data through a combination of a deep learning model and a mechanism model to obtain a first defect identification result. The first defect identification result is used as the final defect identification result, and the deep learning model is optimized based on the defect identification result. At the same time, feature extraction is performed on the voiceprint data to obtain feature vectors. Vector retrieval is performed in the feature vector library to obtain a second defect identification result. If the first defect identification result and the second defect identification result are inconsistent, the feature vector is added to the feature vector library. When the entropy value of the data in the feature vector library is higher than the second threshold, feature extraction is performed on the voiceprint data to obtain feature vectors. Vector retrieval is performed in the feature vector library to obtain a second defect identification result, which is used as the final defect identification result.

[0113] It should be noted that when the entropy value of the data in the feature vector library is higher than the second threshold, if the similarity between the feature vector and the data in the feature vector library is lower than the set value during vector retrieval, then return to step 2, re-implement defect identification through the joint deep learning model and mechanism model, and add the feature vector to the feature vector library.

[0114] Currently, there are pain points in the field application of substation defect identification: due to the massive amount of data, deep learning model training is time-consuming, data updates on defects from various substations cannot be fed back in a timely manner, the training library is insufficient to cover various situations of the main equipment, and the model cannot be updated and iterated in a timely manner. There is an urgent need for targeted and differentiated solutions to improve the detection capabilities of each substation.

[0115] The incremental learning module consists of four modules: building a local voiceprint defect database, constructing a feature vector library, main device voiceprint defect model detection (i.e., the voiceprint detection model in step 2), and vector retrieval. It is a practical, lightweight, incremental learning general-purpose voiceprint image recognition system. When local defects are updated, the incremental learning module only needs to update the local voiceprint defect image database, without retraining the model, enabling "addition and use immediately" and solving the problems of slow model training and updates, and local feature library upgrades.

[0116] like Figure 2 As shown, the incremental learning module consists of four steps, which are described below:

[0117] (1) Build a local voiceprint defect database.

[0118] Acoustic print images of each main device in the local substation are collected at different times. If the main device fails to generate acoustic print defects for a continuous period of time, different forms of noise are artificially added to the normal acoustic print samples to simulate the generation of acoustic print defects, thereby generating a local acoustic print defect database, which is then stored on the local hard drive.

[0119] (2) Construct a feature vector library.

[0120] Feature extraction transforms the input image (the latest captured voiceprint image from the main device) into a fixed-dimensional feature vector. Image pairs with high similarity also have high feature similarity, while image pairs with low similarity have low feature similarity.

[0121] First, signal processing is performed on the audio signal images of the voiceprint defect database, such as framing, windowing, translation, and spectrum calculation, to obtain a time-frequency domain spectrum. The time-frequency domain spectrum is then input into the feature extraction network module, and the output is a fixed-dimensional feature vector, which is added to the vector library for subsequent vector retrieval.

[0122] To allow for flexible customization for image recognition tasks, the entire network is divided into Backbone, Neck, Head, and Loss sections, with the overall structure as follows: Figure 3 As shown. The functions of each module:

[0123] Backbone: The backbone network used to extract preliminary features from the input image, typically specified by the Backbone and BackboneStopLayer fields in the configuration file. The Backbone portion uses PP-LCNetV2_base, which, based on PPLCNet_V1, incorporates several optimizations including Rep strategy, PW convolution, Shortcut, improved activation functions, and improved SE module, resulting in final recognition accuracy similar to PPLCNet_x2_5, while reducing inference latency by 40%.

[0124] Neck: Used for feature enhancement and feature dimension transformation. It can be a simple FC Layer used for feature dimension transformation, or a more complex FPN structure used for feature enhancement. It is generally specified by the Neck field in the configuration file.

[0125] Head: Used to convert the output features of the Neck into logits, allowing the model to be trained as a classification task during the training phase.

[0126] Loss: Specifies the loss function used. The loss function uses Cross entropy loss and Triplet Angular Margin Loss, guiding network optimization during training with classification loss and angle-based triplet loss. An improvement on the original Triplet Loss (hard triplet loss) is made by changing the optimization objective from L2 Euclidean space to cosine space and adding a hard distance constraint between anchors and positive / negative targets, making the training and testing objectives closer and improving the model's generalization ability.

[0127] (3) Voiceprint model detection.

[0128] The audio signal image of the main equipment of the local substation is processed to obtain the time-frequency domain spectrum. The time-frequency spectrum is not only input into the mechanism diagnosis model and the deep learning model to output diagnosis result 1 (first defect identification result), but also input into the feature extraction network to obtain feature vectors, which are input into step (4) for vector retrieval to obtain diagnosis result 2 (second defect identification result).

[0129] (4) Vector retrieval.

[0130] For a given query vector, the similarity or distance of feature vectors is calculated in the established vector library, and the similarity ranking of candidate vectors is returned. Its main goal is to calculate the similarity or distance of feature vectors with all other query vectors in the established vector library for a given query vector, and obtain a similarity ranking. Each master device's voiceprint vector library supports both normal and abnormal voiceprint vector libraries.

[0131] In image recognition systems, open source is used here. Faiss To support this, Faiss's main principle is to build an index for the base vectors and then use the index to perform TopK similarity vector retrieval on the search vectors. Faiss has the following main advantages:

[0132] Good compatibility: Supports Windows, Linux, and MacOS systems;

[0133] Easy to install: Supports Python interface, can be installed directly using pip;

[0134] Rich algorithms: Supports multiple retrieval algorithms to meet the needs of different scenarios;

[0135] It supports both CPU and GPU, which can accelerate the search process.

[0136] (5) Combined diagnosis.

[0137] Combining the results of the first and second defect identifications, if either one of them diagnoses an abnormal voiceprint, the main equipment is judged to be malfunctioning. After verification, the voiceprint images of devices that were incorrectly identified or missed (where the error between the first and second defect identification results exceeds a set value) are processed through feature extraction and updated in the vector feature library.

[0138] As the vector library covers more and more features (the entropy value of the data in the feature vector library is higher than the second threshold), voiceprint model detection can be gradually discarded, and similarity matching can be performed directly from the vector library, thereby achieving rapid diagnosis, timely detection of faults, and contributing to the stable operation of power grid equipment.

[0139] As the feature vector library gradually expands, this invention gradually omits voiceprint model detection and directly performs similarity matching from the vector library, thereby achieving rapid diagnosis and timely fault detection, which is beneficial to the stable operation of power grid equipment. Moreover, it only requires updating the local voiceprint defect image database and does not require retraining the model, enabling "addition and use immediately" and solving the problems of slow model training and updates and local feature library upgrades.

[0140] After identifying the device to be identified as abnormal, this invention uses a combination of deep learning model and mechanism model to achieve defect identification, basically covering all potential operating states of substation main equipment, and expanding the field application of voiceprint diagnosis technology. In addition, considering the limited cases of early deep learning model learning, more weight is given to the mechanism model. As the dataset becomes richer and expanded, the weight of the mechanism model is gradually reduced or eliminated, ultimately improving the joint identification accuracy of deep learning model and mechanism model.

[0141] This invention uses latent vectors and reconstruction loss obtained from reconstruction features to determine the degree of abnormality of transformer acoustic signatures, achieving abnormal signal detection under conditions of few samples. By combining the abnormal probability score and reconstruction loss, it enables quantitative evaluation of the risk status of transformer acoustic signatures, effectively improving the accuracy of acoustic signature anomaly detection.

[0142] Example 2

[0143] This embodiment provides a power main equipment acoustic signature defect identification system, which specifically includes:

[0144] The data acquisition module is configured to acquire the voiceprint data of the device to be identified after the device is determined to be abnormal.

[0145] The first identification module is configured to: when the entropy value of the data in the feature vector library is lower than the second threshold, identify defects based on the voiceprint data by jointly using a deep learning model and a mechanism model to obtain a first defect identification result, use the first defect identification result as the final defect identification result, and optimize the deep learning model based on the first defect identification result; at the same time, extract features from the voiceprint data to obtain feature vectors, perform vector retrieval in the feature vector library to obtain a second defect identification result; if the first defect identification result and the second defect identification result are inconsistent, add the feature vector to the feature vector library.

[0146] The second identification module is configured to: extract features from the voiceprint data when the entropy value of the data in the feature vector library is higher than the second threshold, obtain feature vectors, perform vector retrieval in the feature vector library, obtain the second defect identification result, and use the second defect identification result as the final defect identification result.

[0147] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0148] Example 3

[0149] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the power main equipment acoustic defect identification method described in Embodiment 1 above.

[0150] Example 4

[0151] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the power main equipment voiceprint defect identification method described in Embodiment 1 above.

[0152] Example 5

[0153] This embodiment provides a computer program product, which is a computer program product containing a computer program. When the processor executes the computer program, it implements the steps in the power main equipment voiceprint defect identification method described in Embodiment 1 above.

[0154] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0155] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0156] The above description is merely a preferred embodiment of this application and is not intended to hinder this application in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of this application shall fall within the protection scope of this application.

Claims

1. A method for identifying acoustic defects in main power equipment, characterized in that, include: After the device to be identified is determined to be abnormal, the voiceprint data of the device to be identified is obtained; When the entropy value of the data in the feature vector library is lower than the second threshold, defect identification is achieved by combining a deep learning model and a mechanism model based on the voiceprint data to obtain the first defect identification result. The first defect identification result is used as the final defect identification result, and the deep learning model is optimized based on the first defect identification result. At the same time, feature extraction is performed on the voiceprint data to obtain feature vectors. Vector retrieval is performed in the feature vector library to obtain the second defect identification result. If the first defect identification result and the second defect identification result are inconsistent, the feature vector is added to the feature vector library. When the entropy value of the data in the feature vector library is higher than the second threshold, if the similarity between the feature vector and the data in the feature vector library is lower than the set value during vector retrieval, then defect identification will be re-implemented through the joint deep learning model and mechanism model, and the feature vector will be added to the feature vector library. When the entropy value of the data in the feature vector library is higher than the second threshold, feature extraction is performed on the voiceprint data to obtain feature vectors. Vector retrieval is then performed in the feature vector library to obtain the second defect identification result, which is then used as the final defect identification result.

2. The method for identifying acoustic defects in main power equipment as described in claim 1, characterized in that, The anomaly determination step of the device to be identified includes: acquiring the voiceprint features and conditional features of the device to be detected; obtaining the input features of the device to be detected through feature combination; inputting the input features of the device to be detected into an inference network, the inference network including a generator encoder and a generator decoder and an anomaly score calculation network connected to the generator encoder respectively; the generator encoder processes the input features of the device to be detected to obtain latent vectors; the generator decoder processes the latent vectors to obtain reconstructed features; the anomaly score calculation network processes the latent vectors to obtain anomaly scores; the reconstruction loss is calculated based on the latent vectors and the reconstructed features; the weighted sum of the reconstruction loss and the anomaly score is used as the risk value of the device to be detected; when the risk value of the device to be detected exceeds the risk threshold, it is determined to be an anomaly.

3. The method for identifying acoustic defects in main power equipment as described in claim 1, characterized in that, When the entropy value of the data in the feature vector library is lower than the first threshold, the weight of the mechanism model is greater than the weight of the deep learning model, and vice versa.

4. The method for identifying acoustic defects in main power equipment as described in claim 1, characterized in that, The feature extraction steps include: performing signal processing on the audio signal image in the voiceprint data to obtain a time-frequency domain spectrogram, inputting the time-frequency domain spectrogram into the feature extraction network, and outputting a feature vector of fixed dimensions.

5. The method for identifying acoustic defects in main power equipment as described in claim 1, characterized in that, The mechanism model identifies defects based on spectrogram features, which include the transformer body acoustic waveform signal-to-noise ratio, sound pressure level, odd-even harmonic ratio, high-frequency energy ratio, transformer spectral entropy, and resonant demodulation power frequency correlation.

6. The method for identifying acoustic defects in main power equipment as described in claim 5, characterized in that, The transformer body acoustic signature signal-to-noise ratio is used to determine if there is interference. If interference signal is determined to exist based on the transformer body acoustic signature signal-to-noise ratio, the first defect identification result is output as interference signal, and the mechanism model is no longer used. The sound pressure level is used to identify heavy load faults and / or sensor anomalies. The odd-even harmonic ratio is used to determine DC bias or harmonic current faults. The high-frequency energy ratio is used to determine loosening-type faults; The transformer spectrum entropy is used to diagnose abnormal noises in the cooler. The resonant demodulation power frequency correlation is used for partial discharge fault diagnosis.

7. A power main equipment voiceprint defect identification system, characterized in that, include: The data acquisition module is configured to acquire the voiceprint data of the device to be identified after the device is determined to be abnormal. The first identification module is configured to: when the entropy value of the data in the feature vector library is lower than the second threshold, identify defects based on the voiceprint data by jointly using a deep learning model and a mechanism model to obtain a first defect identification result, use the first defect identification result as the final defect identification result, and optimize the deep learning model based on the first defect identification result; at the same time, extract features from the voiceprint data to obtain feature vectors, perform vector retrieval in the feature vector library to obtain a second defect identification result; if the first defect identification result and the second defect identification result are inconsistent, add the feature vector to the feature vector library. When the entropy value of the data in the feature vector library is higher than the second threshold, if the similarity between the feature vector and the data in the feature vector library is lower than the set value during vector retrieval, then defect identification will be re-implemented through the joint deep learning model and mechanism model, and the feature vector will be added to the feature vector library. The second identification module is configured to: extract features from the voiceprint data when the entropy value of the data in the feature vector library is higher than the second threshold, obtain feature vectors, perform vector retrieval in the feature vector library, obtain the second defect identification result, and use the second defect identification result as the final defect identification result.

8. The power main equipment voiceprint defect identification system as described in claim 7, characterized in that, When the entropy value of the data in the feature vector library is lower than the first threshold, the weight of the mechanism model is greater than the weight of the deep learning model, and vice versa.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the power main equipment acoustic defect identification method as described in any one of claims 1-6.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the power main equipment acoustic defect identification method as described in any one of claims 1-6.

11. A computer program product, which is a computer program product containing a computer program, characterized in that, When the processor executes the computer program, it implements the steps in the power main equipment acoustic defect identification method as described in any one of claims 1-6.

Images