Method and system for recognizing voiceprint defect of power master equipment

Abstract

The invention relates to the technical field of power equipment monitoring, and provides a power master equipment voiceprint defect recognition method and system, and the method comprises the steps: achieving the defect recognition based on voiceprint data through the combination of a deep learning model and a mechanism model when a data entropy value in a feature vector library is lower than a second threshold value, and obtaining a first defect recognition result, the deep learning model is optimized based on the first defect recognition result; carrying out feature extraction on the voiceprint data, carrying out vector retrieval in a feature vector library to obtain a second defect recognition result, and if the recognition results are inconsistent, adding feature vectors into the feature vector library; and when the data entropy value in the feature vector library is higher than a second threshold value, performing vector retrieval in the feature vector library to obtain a second defect identification result, and taking the second defect identification result as a final defect identification result. All potential operation states of the power transformation main equipment are basically covered, rapid diagnosis can be achieved, faults can be found in time, and stable operation of power grid equipment is facilitated.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment monitoring, and in particular relates to a method and system for identifying voiceprint defects of power main equipment. Background Art

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

[0003] Carrying out online monitoring of power equipment can improve the timeliness of obtaining key status quantities, help discover equipment hidden dangers in advance, enhance power supply reliability, realize intelligent management and optimization of power grid operation, reduce the operating costs of power companies and shorten fault handling time.

[0004] Although voiceprint defect recognition technology is also included in the overall framework of intelligent inspection, the current attention and development of voiceprint defect recognition technology for transformer internal operation is insufficient, and the corresponding defect recognition and analysis technology is slowly being applied. Therefore, it is urgent to study and apply key intelligent inspection technologies such as transformer voiceprint defect recognition.

[0005] The existing voiceprint defect recognition technology has the following disadvantages:

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

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

[0008] (3) Due to the huge amount of substation scene data, model training is time-consuming, the data of 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] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a method and system for identifying voiceprint defects of power main equipment. After the equipment to be identified is determined to be abnormal, defect identification is realized through the joint use of deep learning model and mechanism model, which basically covers all potential operating states of the substation main equipment. Moreover, as the number of features covered in the feature vector library gradually increases, the voiceprint model detection is gradually abandoned, and similarity matching is directly performed from the vector library, thereby achieving rapid diagnosis and timely fault detection, which is conducive to the stable operation of power grid equipment.

[0010] In order to achieve the above object, the present invention adopts the following technical solution:

[0011] A first aspect of the present invention provides a method for identifying voiceprint defects of a power main device, 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 data entropy value in the feature vector library is lower than the second threshold, defect recognition is implemented based on the voiceprint data through the deep learning model and the mechanism model to obtain a first defect recognition result, the first defect recognition result is used as the final defect recognition result, and the deep learning model is optimized based on the first defect recognition result; at the same time, feature extraction is performed on the voiceprint data to obtain a feature vector, and a vector search is performed in the feature vector library to obtain a second defect recognition result. If the first defect recognition result and the second defect recognition result are inconsistent, the feature vector is added to the feature vector library;

[0014] When the data entropy value in the feature vector library is higher than the second threshold, feature extraction is performed on the voiceprint data to obtain a feature vector, and 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.

[0015] Furthermore, the abnormality determination step of the device to be identified includes: obtaining 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 includes a generation encoder and a generation decoder and an abnormality score calculation network respectively connected to the generation encoder, the generation encoder processes the input features of the transformer to be detected to obtain a latent vector, the generation decoder processes the latent vector to obtain a reconstructed feature, the abnormality score calculation network processes the latent vector to obtain an abnormality score, the reconstruction loss is calculated based on the latent vector and the reconstructed feature, and the weighted sum of the reconstruction loss and the abnormality 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 abnormal.

[0016] Furthermore, when the data entropy value 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, otherwise it is less than.

[0017] Furthermore, when the data entropy value in the feature vector library is higher than the second threshold, if in the vector retrieval, the similarity between the feature vector and the data in the feature vector library is lower than the set value, the defect identification is re-implemented through the deep learning model and the 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 a feature extraction network, and outputting a feature vector of a fixed dimension.

[0019] Furthermore, the mechanism model identifies defects based on the features of the spectrum graph, and the features of the spectrum graph include the signal-to-noise ratio of the transformer body soundprint, sound pressure level, odd-even harmonic ratio, high-frequency energy proportion, transformer spectral entropy and resonance demodulation power frequency correlation.

[0020] Furthermore, the transformer body voiceprint signal-to-noise ratio is used to determine whether interference is incorrect. If an interference signal is determined to exist based on the transformer body voiceprint signal-to-noise ratio, the first defect recognition result is output as an interference signal, and the mechanism model is no longer combined;

[0021] The sound pressure level is used to identify heavy load faults and / or sensor abnormality judgments;

[0022] The odd-even harmonic ratio is used for DC bias magnetic or harmonic current fault judgment;

[0023] The high-frequency energy ratio is used to judge loose faults;

[0024] The transformer spectral entropy is used to judge abnormal noise faults of coolers;

[0025] The resonance demodulation power frequency correlation is used for partial discharge fault judgment.

[0026] A second aspect of the present invention provides a system for identifying voiceprint defects of a power main device, comprising:

[0027] A data acquisition module, which is configured to: acquire voiceprint data of the device to be identified after the device to be identified is determined to be abnormal;

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

[0029] The second recognition module is configured as follows: when the data entropy value in the feature vector library is higher than the second threshold value, feature extraction is performed on the voiceprint data to obtain a feature vector, vector retrieval is performed in the feature vector library to obtain a second defect recognition result, and the second defect recognition result is used as the final defect recognition result.

[0030] Furthermore, when the data entropy value 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, otherwise it is less than.

[0031] Furthermore, when the data entropy value in the feature vector library is higher than the second threshold, if in the vector retrieval, the similarity between the feature vector and the data in the feature vector library is lower than the set value, the defect identification is re-implemented through the deep learning model and the 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 in the method for identifying voiceprint defects of a power master device as described above.

[0033] The 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 when the processor executes the program, the steps in the method for identifying voiceprint defects of a power master device as described above are implemented.

[0034] A fifth aspect of the present invention provides a computer program product, which is a computer program product containing a computer program. When a processor executes the computer program, the steps in the method for identifying voiceprint defects of a power master device as described above are implemented.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention innovatively proposes an incremental learning method for substation inspection image recognition. By constructing a local voiceprint defect database, a feature vector library, a main equipment voiceprint defect model and a vector retrieval module, a lightweight incremental learning voiceprint image recognition method is developed. Without the participation of old samples, the recognition and detection of new equipment defect categories can be realized. When local defects are updated, only the local voiceprint defect image database needs to be updated without retraining the model. It can achieve "instant add-on", which solves the problems of slow model training and update and local feature library upgrade and update, and realizes the rapid and accurate diagnosis of main equipment defect faults.

[0037] The present invention innovatively proposes a multi-model defect recognition method for power equipment that integrates the voiceprint mechanism model and the deep learning model. The mechanism model is trained by screening the spectral features in the low-frequency band, and the deep learning model is optimized by training with real-time data. The two results are fused and driven to achieve accurate recognition of abnormal voiceprints, which solves the problem of misjudgment in defect recognition of power equipment, realizes accurate recognition of abnormal voiceprints of power equipment, and improves the robustness of the model in the face of equipment differences and environmental noise.

[0038] The present invention innovatively proposes a method for detecting voiceprint anomalies of power main equipment, which uses a large number of normal voiceprint samples and a small number of abnormal voiceprint samples to reconstruct the original features, and judges the degree of abnormality of the voiceprint of the main equipment through the sample reconstruction loss. It solves the problems of low accuracy and insufficient generalization of the deep learning model caused by insufficient voiceprint fault samples, realizes the detection of abnormal signals under the condition of few samples, and effectively improves the accuracy of voiceprint anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0040] Figure 1 It is a flow chart of the method of realizing defect identification by combining the degree learning model and the mechanism model in the first embodiment of the present invention;

[0041] Figure 2 is a flowchart of incremental learning according to the first embodiment of the present invention;

[0042] Figure 3 It is a structural diagram of the feature extraction network of the first embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order 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 invention purpose of this application through several specific embodiments. It should be noted that the technical solutions claimed for protection 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 making creative efforts shall fall within the scope of protection of this application.

[0044] Embodiment 1

[0045] This embodiment provides a method for identifying voiceprint defects of a power master device.

[0046] This embodiment provides a method for identifying voiceprint defects of a power main device, the objectives of which include:

[0047] (1) Incremental learning: As the number of features included in the feature vector library increases, voiceprint model detection is gradually abandoned, and similarity matching is performed directly from the vector library, thereby achieving rapid diagnosis and timely detection of faults, which is conducive to the stable operation of power grid equipment. In addition, only the local voiceprint defect image database needs to be updated without retraining the model, which can achieve "instant use" and solve the problem of slow model training and updating and local feature library upgrade and update.

[0048] (2) More accurate model recognition: Defect recognition is achieved through the combination of deep learning models and mechanism models, which basically covers all potential operating states of substation main equipment. In addition, considering that the early deep learning model has limited learning cases, the mechanism model can be given more weight appropriately. As the data set is enriched and expanded, the weight of the mechanism model can be gradually reduced or cancelled, ultimately improving the model recognition accuracy.

[0049] (3) Higher accuracy of voiceprint anomaly detection: The reconstruction loss obtained based on latent vectors and reconstruction features is used to judge the abnormality of the transformer voiceprint, realizing the detection of abnormal signals under the condition of few samples. The abnormal probability score and reconstruction loss are combined to realize the quantitative evaluation of the risk status of the transformer voiceprint, which effectively improves the accuracy of voiceprint anomaly detection. After identifying the voiceprint abnormal data based on massive positive samples, the abnormal data is further identified through multi-model feature fusion defect recognition technology, and the specific defect category can be identified.

[0050] The present embodiment provides a method for identifying voiceprint defects of power main equipment, which includes three parts. The first is a voiceprint data anomaly detection technology, which identifies abnormal voiceprint data based on massive positive samples; then, the abnormal data is further identified through multi-model feature fusion defect recognition technology, and the specific defect category can be identified; in addition, for newly emerging equipment defect categories, a self-evolving detection network is proposed, which can realize 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 to be detected, and determine whether the device to be detected is abnormal.

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

[0053] As another implementation method, the voiceprint features and conditional features of the device to be detected are obtained, and the input features of the device to be detected are obtained by feature combination; the input features of the device to be detected are input into an inference network, and the inference network includes a generative encoder and a generative decoder and an anomaly score calculation network respectively connected to the generative encoder, the generative encoder processes the input features of the transformer to be detected to obtain a latent vector, the generative decoder processes the latent vector to obtain a reconstructed feature, the anomaly score calculation network processes the latent vector to obtain an anomaly score, and a reconstruction loss is calculated based on the latent vector and the reconstructed feature, and a 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 abnormal.

[0054] Specifically, the voiceprint feature s and conditional feature c of the device to be detected are combined into the input feature x = (s, c), and then the encoder G is generated.E Get the hidden vector z, and generate the decoder G D Get the reconstructed features of x After an additional encoder E, the reconstructed features of the latent vector are obtained In addition, combining the features x and reconstructing In the training phase, the input is sent to the discriminator D, which continuously learns to be as "undeceivable" as possible. The discriminator and the generator (encoder G E and decoder G D Through continuous adversarial training, we seek the optimal solution.

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

[0056] (1) Feature combination: The voiceprint feature s of the device to be tested is derived from the normal voiceprint data of the device to be tested, such as MFCC, time-frequency diagram, mel-spectrogram, etc. The conditional feature c includes the voltage level (110kV, 220kV, 500kV, 1000kV) of the device to be tested, the cooling method (oil-immersed air cooling, forced oil circulation air cooling, forced guided oil circulation air cooling), the manufacturer and the station to which it belongs. All conditional features are combined and one-hot encoded. Then the voiceprint feature and the conditional feature are combined into the input feature x = (s, c).

[0057] (2) Generate encoder G E :The input feature x is calculated by the generative encoder to obtain the latent vector z = G E (x,c), generating encoder G E Including convolutional layers, layer normalization, activation functions, and Dropout, etc., to achieve dimensional compression of the input feature x and represent the potential features of the input feature. Layer Normalization is used here to optimize the problem that batch normalization is easily affected by changes in the distribution of data within the batch. Dropout is only used in the training phase. When testing or evaluating the model, the probability needs to be set to 0 to turn off Dropout.

[0058] (3) Generate decoder G D : Hidden vector z = G E (x,c) passes through the generative decoder G D Afterwards, after the scale is enlarged upward, the reconstructed feature vector is obtained. Generate decoder G D It includes transposed convolution layer, normalization layer and activation function layer. If the difference between x and x is very small, it means that the input feature x can be approximately reconstructed, and then the transformer voiceprint 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 Represents the i-th element of the feature.

[0061] (4) Encoder E: In order to better learn the potential 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 is consistent with the latent vector z, it indicates the reconstructed feature The potential distribution characteristics of are similar to those of the input feature x, and the encoding loss is still defined using the Smooth L1 loss function as follows:

[0062]

[0063] in Represents the i-th element of the feature. This method greatly reduces the time complexity and space complexity of the calculation and improves the speed of fault detection. Encoders E and G E The structure is consistent.

[0064] (5) Discriminator network D.

[0065] The role of the discriminator D is consistent with the traditional adversarial generation network. Input features and conditional features (x, c) are input and reconstructed features The input discriminator D is scaled down multiple times through the transposed convolution layer, batch normalization layer, and leaky linear rectifier activation function layer. The compressed vector is flattened and input into the fully connected layer to achieve the purpose of binary classification (determining whether the input is original feature data or generated feature data). The adversarial loss is defined as follows:

[0066]

[0067] In general, the overall loss of the model during training is as follows:

[0068] L=L adv +αL rec +βL enc

[0069] During the training phase, the model is trained using voiceprint monitoring data of different voltage levels, from different sites and with different cooling methods, with weights α=1, β=0.1.

[0070] In the case of minimal loss, the risk threshold is calculated:

[0071] R=L rec +λL enc

[0072] Where λ is an adjustable parameter, which is selected according to the batch data.

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

[0074] Step 2: For devices to be identified that are judged to be 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 the main power equipment is proposed. Figure 1 As shown in the figure, the recognition and diagnosis of the mechanism model is first realized through feature extraction and other steps according to the diagnosis rules. The feature acquisition in the spectrum diagram adopts a low-frequency band selection strategy to improve the accuracy of data features and reduce the redundancy of data volume; then the real-time collected data is transmitted to the deep learning model; the fusion drive idea is used for the two results to realize the accurate recognition of abnormal voiceprints, improve the robustness of the model in the face of equipment differences and environmental noise, and meet the complex environment and fault conditions of the substation.

[0076] Step 201: Based on the frequency spectrum in the voiceprint data, recognition and diagnosis of the mechanism model are realized through feature extraction and other steps according to diagnostic rules.

[0077] (1) Feature extraction

[0078] Refer to the feature calculation formula in Table 1 to complete the feature extraction work. Only 50, 100, 150, and 200 Hz in the spectrum graph are extracted, which not only ensures effective features, but also achieves dimensionality reduction and reduces data redundancy. The spectrum graph features specifically include transformer body soundprint signal-to-noise ratio, sound pressure level, odd-even harmonic ratio, high-frequency energy ratio, transformer spectral entropy and other feature extractions.

[0079] Among them, RMS is the effective value of the voiceprint signal (frequency band above 20Hz), and base is the reference standard of the sound pressure level, specifically 20uPa. i*50 is the spectrum of the voiceprint signal, and the corresponding index is the i-th 50Hz frequency. is the energy proportion at the corresponding frequency.

[0080] Table 1. Main characteristic calculation formula

[0081]

[0082] (2) Diagnostic rules.

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

[0084] 1) External interference.

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

[0086] 2) Overload.

[0087] When the sound pressure level is too large, due to the positive correlation between the transformer operating noise and the load (to be precise, it changes linearly with the square of the load current), when the sound pressure level increases significantly (that is, the sound pressure level exceeds the set value), it often indicates an overload or heavy load fault. If it works in an overload state for a long time, it is very easy to cause irreversible damage to the transformer.

[0088] 3) DC bias magnetization.

[0089] When the odd-even harmonic ratio of the transformer sound pattern increases significantly (that is, the odd-even harmonic ratio exceeds the set value), it can generally be directly determined that a DC bias magnetic defect exists, which may be accompanied by an increase in noise. At this time, a DC bias magnetic fault can be directly output.

[0090] 4) Loose.

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

[0092] 5) Other abnormalities.

[0093] When one or more frequency spectrum features of transformer voiceprint data increase or decrease significantly compared with the history, or when the voiceprint features of a transformer are significantly different from those of other transformers of the same model and working conditions, it often indicates that there are some abnormalities, and the existing faults may not be fully covered, but it should still be treated as an abnormal alarm.

[0094] Table 2. Main features, meanings and uses

[0095]

[0096] (3) Alarm strategy

[0097] In order to reduce the possibility of false alarms under special interference, such as when there is equipment construction nearby and the noise has the same frequency as the transformer, false alarms may occur. Therefore, this mechanism diagnosis model will only output alarm information when the soundprint meets the alarm conditions and continues to alarm the same fault several times.

[0098] Step 202: Transmit the voiceprint waveform data collected in real time to the deep learning model.

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

[0100] Step 203: A fusion drive approach is used for the two results to achieve accurate recognition of abnormal voiceprints.

[0101] Based on the diagnosis results of the mechanism diagnosis model and the deep learning model, a fusion-driven approach is adopted to set certain weights for them respectively to constitute the final output results, thereby realizing the final output of the transformer voiceprint recognition results. Taking into account the limited learning cases of the early deep learning model, more weight can be given to the mechanism model. As the data set is enriched and expanded, the weight of the mechanism model can be gradually reduced or cancelled.

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

[0103] 1) The mechanism model is used to realize the discrimination of abnormal data not related to transformer faults. When the voiceprint data noise is large and the signal-to-noise ratio of the transformer voiceprint is low, it may be due to sensor failure, abnormal sensor installation, low data acquisition quality, strong interference, etc. At this time, it is not easy to enter the fault discrimination logic, and the voiceprint signal can be directly output as an interference signal;

[0104] 2) When the voiceprint data passes the initial judgment of the mechanism model, it means that the voiceprint is likely to be the real transformer voiceprint data. At this time, the collected voiceprint samples are predicted by the mechanism model and the deep learning model respectively, and the discrimination result and confidence probability of the mechanism model and the maximum probability corresponding to the category and its probability of the deep learning model recognition result are output, corresponding to the label 机理 , P 机理 、label 深度学习 , P 深度学习 ;

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

[0106] 4) Calculate the final output result, compare the reliability weight of the model result and the comprehensive result P (P = max (P 机理 *w 机理 ,P 深度学习 *w 深度学习 ), determine the output category with the highest probability, and finally output the result;

[0107] 5) Update the reliability weights of the mechanism model and deep learning model results. When there is new progress in the research on a certain type of fault mechanism, the reliability weight of the corresponding category of faults in the mechanism 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 above 95%, the reliability weight of the deep learning model can be appropriately increased.

[0108] Given that the transformer fault samples contained in the early data sets are not rich enough and are mostly data collected from different angles of the same transformer, the trained machine learning model is difficult to be widely applied to other transformer defect identification. 深度学习 =0.3, while the mechanism model is easy to set appropriate parameters according to the horizontal and vertical comparison results because of its adjustable parameters. It has relatively high accuracy and is recommended to be based on higher weights, 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 , otherwise it is less than. The first threshold is less than the second threshold.

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

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

[0112] Specifically, when the data entropy value in the feature vector library is lower than the second threshold, defect identification is realized jointly through the deep learning model and the mechanism model based on the voiceprint data 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 a feature vector, and a vector search 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 data entropy value in the feature vector library is higher than the second threshold, feature extraction is performed on the voiceprint data to obtain a feature vector, and a vector search is performed in the feature vector library to obtain a second defect identification result, and the second defect identification result is used as the final defect identification result.

[0113] It should be noted that when the data entropy value 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 deep learning model and mechanism model, and add the feature vector to the feature vector library.

[0114] At present, there are pain points in the on-site application of defect identification of substation equipment: due to the huge amount of data, the deep learning model training is time-consuming, the updated data of defects in 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. Targeted and differentiated solutions are urgently needed to improve the detection level of each station.

[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 universal 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, which can achieve "instant use" and solve the problem of slow model training and update and local feature library upgrade and update.

[0116] like Figure 2 As shown, the incremental learning module is divided into four steps, which are described as follows:

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

[0118] The voiceprint images of each main device in the local substation are collected at different times. If the main device fails to produce voiceprint defects for a continuous period of time, different forms of noise are artificially added to the normal voiceprint samples to simulate voiceprint defects, generate a local voiceprint defect database, and save it on the local hard disk.

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

[0120] Feature extraction is to convert the input image (the latest collected main device voiceprint image) into a fixed-dimensional feature vector. The image pairs with high similarity have high feature similarity, and the image pairs with low similarity have low feature similarity.

[0121] First, the audio signal image of the voiceprint defect database is processed, such as framing, windowing, translation, spectrum calculation and other operations to obtain the time-frequency domain spectrogram, which is then input into the feature extraction network module and output as a feature vector of fixed dimension, which is added to the vector library for subsequent vector retrieval.

[0122] In order to flexibly customize the image recognition task, the entire network is divided into Backbone, Neck, Head and Loss parts. The overall structure is as follows Figure 3 As shown. The functions of each module are:

[0123] Backbone: The backbone network used to extract preliminary features of the input image, which is generally specified by the Backbone and BackboneStopLayer fields in the configuration file. The Backbone part uses PP-LCNetV2_base, which adds multiple optimization points including Rep strategy, PW convolution, Shortcut, activation function improvement, SE module improvement, etc. on the basis of PPLCNet_V1, making the final recognition accuracy close to PPLCNet_x2_5, and reducing the inference delay by 40%.

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

[0125] Head: It is used to convert the output feature of Neck into logits, so that the model can be trained as a classification task during the training phase.

[0126] Loss: Specifies the Loss function used. Cross entropy loss and TripletAngularMarginLoss are used for the Loss part. Classification loss and angle-based triplet loss are used to guide the network optimization during training. Based on the original TripletLoss (difficult triplet loss), the optimization target is changed from L2 Euclidean space to cosine space, and a hard distance constraint between anchor and positive / negative is added to make the training and testing targets closer and improve the generalization ability of the model.

[0127] (3) Voiceprint model detection.

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

[0129] (4) Vector retrieval.

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

[0131] In the image recognition system, open source Faiss To support this part, the main principle of Faiss is to build an index of base vectors vector data, and then use the index to implement TopK similar vector retrieval for search vectors. Faiss has the following main advantages:

[0132] Good adaptability: supports Windos, Linux, and MacOS systems;

[0133] Easy installation: supports Python interface and 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 speed up the retrieval process.

[0136] (5) Combined diagnosis

[0137] Combining the first defect recognition result and the second defect recognition result, as long as one of the diagnosis of voiceprint abnormality is met, the main device is judged to be abnormal. After verification, the device voiceprint image with identification error or missed detection (the error between the first defect recognition result and the second defect recognition result exceeds the set value) is extracted and the vector feature library is updated.

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

[0139] As the number of features covered in the feature vector library gradually increases, the present invention gradually abandons 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, only the local voiceprint defect image database needs to be updated without retraining the model, and can achieve "add-on-the-fly" and solve the problems of slow model training and updating and local feature library upgrade and update.

[0140] After the device to be identified is judged as abnormal, the present invention realizes defect identification through the joint use of deep learning model and mechanism model, which basically covers all potential operating states of the main substation equipment and expands the on-site application of voiceprint diagnosis technology; and considering that the early deep learning model learning cases are limited, the mechanism model is appropriately given more weight. As the data set is enriched and expanded, the weight of the mechanism model is gradually reduced or cancelled, and finally the joint recognition accuracy of the deep learning model and the mechanism model is improved.

[0141] The present invention judges the abnormality of the transformer voiceprint based on the reconstruction loss obtained by latent vector and reconstruction features, realizes the detection of abnormal signals under few sample conditions, and combines the abnormal probability score and reconstruction loss to realize the quantitative evaluation of the risk status of the transformer voiceprint, effectively improving the accuracy of voiceprint anomaly detection.

[0142] Embodiment 2

[0143] This embodiment provides a system for identifying voiceprint defects of a power main device, which specifically includes:

[0144] A data acquisition module, which is configured to: acquire voiceprint data of the device to be identified after the device to be identified is determined to be abnormal;

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

[0146] The second recognition module is configured as follows: when the data entropy value in the feature vector library is higher than the second threshold value, feature extraction is performed on the voiceprint data to obtain a feature vector, vector retrieval is performed in the feature vector library to obtain a second defect recognition result, and the second defect recognition result is used as the final defect recognition result.

[0147] It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0148] Embodiment 3

[0149] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps in the method for identifying voiceprint defects of a power master device as described in the first embodiment above are implemented.

[0150] Embodiment 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, the steps in the method for identifying voiceprint defects of a power master device as described in the first embodiment above are implemented.

[0152] Embodiment 5

[0153] This embodiment provides a computer program product, which is a computer program product containing a computer program. When a processor executes the computer program, the steps in the method for identifying voiceprint defects of a power master device as described in the above-mentioned embodiment 1 are implemented.

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

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

[0156] The above is only a preferred embodiment of the present application and does not constitute any form of obstruction to the present application. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present application shall fall within the protection scope of the present application.

Claims

1. A method for identifying voiceprint defects of power main 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 data entropy value in the feature vector library is lower than the second threshold, defect recognition is implemented based on the voiceprint data through the deep learning model and the mechanism model to obtain a first defect recognition result, the first defect recognition result is used as the final defect recognition result, and the deep learning model is optimized based on the first defect recognition result; at the same time, feature extraction is performed on the voiceprint data to obtain a feature vector, and a vector search is performed in the feature vector library to obtain a second defect recognition result. If the first defect recognition result and the second defect recognition result 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, feature extraction is performed on the voiceprint data to obtain a feature vector, and 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.

2. A method for identifying voiceprint defects of a power main device according to claim 1, characterized in that: The abnormality determination step of the device to be identified includes: obtaining the voiceprint feature and conditional feature of the device to be detected, and obtaining the input feature of the device to be detected by feature combination; inputting the input feature of the device to be detected into an inference network, wherein the inference network includes a generative encoder, a generative decoder and an abnormality score calculation network respectively connected to the generative encoder, the generative encoder processes the input feature of the transformer to be detected to obtain a latent vector, the generative decoder processes the latent vector to obtain a reconstructed feature, the abnormality score calculation network processes the latent vector to obtain an abnormality score, and a reconstruction loss is calculated based on the latent vector and the reconstructed feature, and a weighted sum of the reconstruction loss and the abnormality score is used as the risk value of the device to be detected. When the risk value of the device to be detected exceeds a risk threshold, it is determined to be abnormal.

3. A method for identifying voiceprint defects of a power main device according to claim 1, characterized in that: When the data entropy value 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, otherwise it is less than.

4. A method for identifying voiceprint defects of a power main device according to claim 1, characterized in that: When the data entropy value in the feature vector library is higher than the second threshold, if in the vector retrieval, the similarity between the feature vector and the data in the feature vector library is lower than the set value, the defect identification is re-implemented through the deep learning model and the mechanism model, and the feature vector is added to the feature vector library.

5. A method for identifying voiceprint defects of a power main device according to claim 1, characterized in that: 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 a feature extraction network, and outputting a feature vector of fixed dimension.

6. A method for identifying voiceprint defects of a power main device according to claim 1, characterized in that: The mechanism model identifies defects based on the features of the spectrum graph, and the features of the spectrum graph include the signal-to-noise ratio of the transformer body soundprint, sound pressure level, odd-even harmonic ratio, high-frequency energy proportion, transformer spectral entropy and resonance demodulation power frequency correlation.

7. A method for identifying voiceprint defects of a power main device according to claim 1, characterized in that: The transformer body voiceprint signal-to-noise ratio is used to determine whether interference is incorrect. If an interference signal is determined to exist based on the transformer body voiceprint signal-to-noise ratio, the first defect recognition result is output as an interference signal, and the mechanism model is no longer combined; The sound pressure level is used to identify heavy load faults and / or sensor abnormality judgments; The odd-even harmonic ratio is used for DC bias magnetic or harmonic current fault judgment; The high-frequency energy ratio is used to judge loose faults; The transformer spectral entropy is used to judge abnormal noise faults of coolers; The resonance demodulation power frequency correlation is used for partial discharge fault judgment.

8. A system for identifying voiceprint defects of power main equipment, characterized in that: include: A data acquisition module, which is configured to: acquire voiceprint data of the device to be identified after the device to be identified is determined to be abnormal; The first recognition module is configured to: when the data entropy value in the feature vector library is lower than the second threshold value, based on the voiceprint data, jointly realize defect recognition through the deep learning model and the mechanism model to obtain a first defect recognition result, use the first defect recognition result as the final defect recognition result, and optimize the deep learning model based on the first defect recognition result; at the same time, extract features from the voiceprint data to obtain a feature vector, perform vector retrieval in the feature vector library, obtain a second defect recognition result, and if the first defect recognition result and the second defect recognition result are inconsistent, add the feature vector to the feature vector library; The second recognition module is configured as follows: when the data entropy value in the feature vector library is higher than the second threshold value, feature extraction is performed on the voiceprint data to obtain a feature vector, vector retrieval is performed in the feature vector library to obtain a second defect recognition result, and the second defect recognition result is used as the final defect recognition result.

9. A power main equipment voiceprint defect recognition system as claimed in claim 8, characterized in that: When the data entropy value 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, otherwise it is less than.

10. A power main equipment voiceprint defect recognition system as claimed in claim 8, characterized in that: When the data entropy value in the feature vector library is higher than the second threshold, if in the vector retrieval, the similarity between the feature vector and the data in the feature vector library is lower than the set value, the defect identification is re-implemented through the deep learning model and the mechanism model, and the feature vector is added to the feature vector library.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a method for identifying voiceprint defects of a power main device as described in any one of claims 1-7 are implemented.

12. 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, the steps of a method for identifying voiceprint defects of a power master device as described in any one of claims 1-7 are implemented.

13. A computer program product, which is a computer program product containing a computer program, characterized in that When the processor executes the computer program, the steps in the method for identifying voiceprint defects of a power master device as described in any one of claims 1-7 are implemented.

Images