Fault detection method and device for power distribution network transformer, and electronic equipment

Abstract

The invention discloses a fault detection method and device for a power distribution network transformer and electronic equipment. The method comprises the following steps: acquiring a voiceprint detection signal generated during operation of a power distribution network transformer; the voiceprint detection signals are processed through multiple detection models, multiple voiceprint detection results are obtained, the variance between the voiceprint detection results output by different detection models is larger than a preset threshold value, and the multiple detection models are obtained in an integrated learning mode; and summarizing the plurality of voiceprint detection results to obtain a target detection result of the power distribution network transformer, the target detection result being used for representing whether the power distribution network has a fault. According to the invention, the technical problem of low accuracy of fault detection of the power distribution network transformer in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the field of power grid systems, and in particular to a fault detection method and device for a distribution network transformer, and electronic equipment. Background Art

[0002] Distribution network transformers are an indispensable part of the power system, and the healthy operation of transformers is crucial to ensuring the stability and safety of power supply. With the continuous expansion of the scale and increase in complexity of the power system, higher requirements are placed on the real-time monitoring and fault warning of distribution network transformers. Traditional detection methods, such as voltage, current monitoring and infrared thermal imaging technology, can reflect the operating status of the transformer to a certain extent, but they are often only effective when the fault develops to a more serious stage, and are not sensitive enough to early or potential fault information.

[0003] DL / T 573-2021 (China Electric Power Industry Standard) Power Transformer Maintenance Guidelines mentions that manual listening is an effective auxiliary means to evaluate the operating status of transformers. However, this method relies on the experience and subjective judgment of inspectors and has obvious limitations. For example, manual listening is difficult to standardize, diagnostic results may vary from person to person, and it is inefficient and cannot achieve continuous monitoring, which is particularly evident in large or dense distribution network systems. In addition, the complexity of the outdoor environment, including natural environmental noise (such as wind and rain) and interference from other power equipment, further limits the accuracy and reliability of manual listening.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present invention provide a method and device for fault detection of a distribution network transformer, and electronic equipment, so as to at least solve the technical problem of low accuracy in fault detection of a distribution network transformer in the related art.

[0006] According to one aspect of an embodiment of the present invention, a fault detection method for a distribution network transformer is provided, comprising: obtaining a voiceprint detection signal generated by the distribution network transformer during operation; using multiple detection models to process the voiceprint detection signal respectively to obtain multiple voiceprint detection results, wherein the variance between the voiceprint detection results output by different detection models is greater than a preset threshold, and the multiple detection models are obtained by an integrated learning method; summarizing the multiple voiceprint detection results to obtain a target detection result of the distribution network transformer, wherein the target detection result is used to characterize whether a fault occurs in the distribution network.

[0007] Furthermore, multiple detection models are used to process the voiceprint detection signal respectively to obtain multiple voiceprint detection results, including: obtaining multiple model identifiers, wherein multiple detection models and multiple model identifiers correspond one to one; performing feature extraction on the voiceprint detection signal based on the multiple model identifiers to obtain multiple voiceprint features, wherein different model identifiers correspond to different voiceprint features; and inputting the corresponding voiceprint features into the corresponding detection model based on the multiple model identifiers to obtain multiple detection results.

[0008] Furthermore, feature extraction is performed on the voiceprint detection signal based on multiple model identifiers to obtain multiple voiceprint features, including: acquiring multiple feature extraction methods, wherein multiple model identifiers correspond to multiple feature extraction methods one-to-one; feature extraction is performed on the voiceprint detection signal based on multiple feature extraction methods to obtain multiple first features, wherein different feature extraction methods correspond to different first features; in response to the multiple first features containing multiple second features of the same feature type, the multiple second features are matched based on the feature type to obtain feature matching results, wherein the feature matching results are used to reflect the accuracy between different second features; the multiple first features are updated based on the feature matching results to obtain multiple voiceprint features.

[0009] Furthermore, in response to the inclusion of a particle swarm optimization model in multiple detection models, feature extraction is performed on the voiceprint detection signal based on multiple feature extraction methods to obtain multiple first features, including: identifying the voiceprint detection signal to obtain a transient voiceprint signal and a steady-state voiceprint signal contained in the voiceprint detection signal; filtering the transient voiceprint signal based on a preset filter to obtain a transient filtered signal; performing feature extraction on the steady-state voiceprint signal to obtain a steady-state voiceprint feature, and performing feature extraction on the transient filtered signal to obtain a transient voiceprint feature, wherein the multiple first features include steady-state voiceprint features and transient voiceprint features.

[0010] Furthermore, feature extraction is performed on the steady-state voiceprint signal to obtain steady-state voiceprint features, including: framing the steady-state voiceprint signal according to a preset frequency to obtain a steady-state framed signal; performing power identification on the steady-state framed signal to obtain an auto-power spectrum of the steady-state framed signal; performing amplitude identification on the auto-power spectrum to determine the fundamental frequency amplitude and harmonic wave amplitude of the steady-state framed signal; processing the average value of the fundamental frequency amplitude and the average value of the harmonic wave amplitude to obtain the harmonic distortion value of the steady-state voiceprint signal.

[0011] Furthermore, in response to the inclusion of a convolutional neural model in multiple detection models, feature extraction is performed on the voiceprint detection signal based on multiple feature extraction methods to obtain multiple first features, including: framing the voiceprint detection signal to obtain a voiceprint frame signal; windowing the voiceprint frame signal to obtain at least one voiceprint signal segment; performing short-time discrete Fourier transform on the voiceprint signal segment to obtain initial time-frequency spectrum characteristics of the voiceprint detection signal; filtering and transforming the initial time-frequency spectrum characteristics to obtain target time-frequency spectrum characteristics, wherein the multiple first features include the target time-frequency spectrum characteristics.

[0012] Furthermore, the initial time-frequency spectrum features are filtered and transformed to obtain target time-frequency spectrum features, including: obtaining a component matrix of a filter component, wherein the filter component is used to filter and transform the initial time-frequency spectrum features; performing a matrix transformation on the initial time-frequency spectrum features to obtain an initial matrix; multiplying the initial matrix with the component matrix to obtain a filter matrix; and performing logarithm processing on the filter matrix to obtain the target time-frequency spectrum features.

[0013] Furthermore, the method also includes: resampling the training data set based on multiple model identifiers to obtain a resampled data set; training multiple detection models based on the resampled data set to obtain model training results; and assigning multiple detection weights to the multiple detection models based on the model training results, wherein the multiple detection weights are used to aggregate multiple voiceprint detection results to obtain target detection results.

[0014] Furthermore, obtaining a voiceprint detection signal generated by the distribution network transformer during operation includes: performing voiceprint detection on at least one transformer component included in the distribution network transformer through a voiceprint sensor to obtain an initial detection signal, wherein the voiceprint sensor is deployed on the transformer component, the voiceprint sensor corresponds to the transformer one-to-one, and the transformer component includes at least: a transformer body and a cooling component; performing noise preprocessing on the initial detection signal to obtain a voiceprint detection signal.

[0015] According to another aspect of an embodiment of the present invention, a fault detection device for a distribution network transformer is also provided, including: a signal acquisition module, used to obtain a voiceprint detection signal generated by the distribution network transformer during operation; a voiceprint detection module, used to use multiple detection models to process the voiceprint detection signal respectively, to obtain multiple voiceprint detection results, wherein the variance between the voiceprint detection results output by different detection models is greater than a preset threshold, and the multiple detection models are obtained by an integrated learning method; a result aggregation module, used to aggregate the multiple voiceprint detection results to obtain a target detection result of the distribution network transformer, wherein the target detection result is used to characterize whether a fault occurs in the distribution network.

[0016] According to another aspect of an embodiment of the present invention, there is further provided an electronic device, comprising: a memory storing an executable program; and a processor for running the program, wherein the method in each embodiment of the present invention is executed when the program is running.

[0017] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium includes a stored executable program, wherein when the executable program is running, the device where the computer-readable storage medium is located is controlled to execute the methods in various embodiments of the present invention.

[0018] According to another aspect of an embodiment of the present invention, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the method in each embodiment of the present invention is implemented.

[0019] According to another aspect of an embodiment of the present invention, a computer program product is provided, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method in each embodiment of the present invention is implemented.

[0020] According to another aspect of the embodiments of the present invention, a computer program is further provided. When the computer program is executed by a processor, the methods in the embodiments of the present invention are implemented.

[0021] In an embodiment of the present invention, a method is adopted in which a voiceprint detection signal generated by a distribution network transformer during operation is obtained; the voiceprint detection signal is processed respectively by multiple detection models to obtain multiple voiceprint detection results; and the multiple voiceprint detection results are summarized to obtain a target detection result of the distribution network transformer. The detection system can pre-train multiple detection models, use the above multiple detection models to process the acquired voiceprint detection signal, and summarize the detection results of the above multiple detection models through an integrated learning strategy to obtain the final target detection result, thereby achieving the purpose of accurately and efficiently identifying early equipment failures in the distribution network, thereby achieving the technical effect of improving the accuracy and efficiency of distribution network fault detection, and further solving the technical problem of low accuracy of fault detection of distribution network transformers in related technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0023] Figure 1 is a flow chart of a fault detection method for a distribution network transformer according to an embodiment of the present invention;

[0024] Figure 2is an optional processing flow chart of bone conduction voiceprint sensing signals according to an embodiment of the present invention;

[0025] Figure 3 is a flow chart of an optional harmonic distortion calculation according to an embodiment of the present invention;

[0026] Figure 4 is an optional training flow chart of a distribution network transformer voiceprint detection model based on PSO-SVM according to an embodiment of the present invention;

[0027] Figure 5 is an optional training flow chart of a distribution network transformer voiceprint detection model based on a convolutional neural network according to an embodiment of the present invention;

[0028] Figure 6 is an optional training flow chart of a voiceprint detection model based on a semi-supervised generative adversarial network according to an embodiment of the present invention;

[0029] Figure 7 is a structural diagram of an optional bone conduction voiceprint detection sensor according to an embodiment of the present invention;

[0030] Figure 8 is a schematic diagram of an optional constant current source conditioning circuit according to an embodiment of the present invention;

[0031] Fig. 9 is a schematic diagram of an optional 24-bit dual ADC acquisition technology according to an embodiment of the present invention;

[0032] Fig.10 is an optional voiceprint detection flow chart according to an embodiment of the present invention;

[0033] Fig.11 is an optional fault detection flow chart based on a semi-supervised generative adversarial network according to an embodiment of the present invention;

[0034] Fig.12 is a schematic diagram of the structure of an optional generator and a discriminator according to an embodiment of the present invention;

[0035] Fig.13 is a schematic diagram of a fault detection device for a distribution network transformer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0037] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0038] According to an embodiment of the present invention, an embodiment of a fault detection method for a distribution network transformer is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] Figure 1 is a flow chart of a fault detection method for a distribution network transformer according to an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0040] Step S102, obtaining a voiceprint detection signal generated by the distribution network transformer during operation.

[0041] The above-mentioned voiceprint detection signal may refer to a sound or vibration signal captured by a specific device. In an optional embodiment, considering that the internal components of the transformer may show early signs of faults such as wear, looseness, and cracks, the above-mentioned fault signs will be reflected in the sound or vibration mode, but the traditional detection method may not be able to detect them in time. Therefore, the distribution network abnormality detection system (hereinafter referred to as the detection system) can obtain the voiceprint detection signal generated by the distribution network transformer during operation through a sensor, so as to capture the characteristics of the above-mentioned early faults, achieve early warning, and avoid further deterioration of the fault, thereby improving the fault detection accuracy of the distribution network transformer.

[0042] For example, considering the negative impact of environmental noise on voiceprint detection signals, the above sensor can be a bone conduction sensor. The staff can first fix the above bone conduction sensor on the transformer body and cooling module. When the transformer is running, its internal components will make noises, which will be transmitted to the device housing through the solid medium inside the device and then captured by the above bone conduction sensor. Then, the bone conduction sensor can transmit the captured voiceprint detection signal to the detection system, so that the detection system can obtain the voiceprint detection signal.

[0043] For another example, in order to improve the real-time performance of fault detection, the above sensor can be a vibration sensor. The staff can first place the vibration sensor at the key parts of the distribution network transformer, such as the core, winding, cooling module, etc., to accurately capture the vibration generated by the above components during operation and convert it into an electrical signal. Then, the vibration sensor can transmit the captured voiceprint detection signal to the detection system, so that the detection system can obtain the voiceprint detection signal.

[0044] Step S104, using multiple detection models to process the voiceprint detection signal respectively to obtain multiple voiceprint detection results, wherein the variance between the voiceprint detection results output by different detection models is greater than a preset threshold, and the multiple detection models are obtained by integrated learning.

[0045] The above-mentioned multiple detection models can be multiple models for processing the above-mentioned voiceprint detection signals. For example, the above-mentioned multiple detection models can include: PSO-SVM (Particle Swarm Optimization-Support Vector Machine) model, convolutional neural network model, semi-supervised generative adversarial network model, but not limited to this.

[0046] The above-mentioned voiceprint detection result may be the calculation result of the above-mentioned detection model on the above-mentioned voiceprint detection signal. The above-mentioned preset threshold may be a value representing the accuracy of the detection result. The above-mentioned ensemble learning may be a machine learning strategy combining multiple learners.

[0047] In an optional embodiment, considering the diversity of voiceprint signals and sample imbalance, as well as the lack of accuracy, generalization, and robustness of existing algorithms, if a single artificial intelligence algorithm is used to detect voiceprint anomalies of distribution network transformers, the detection accuracy may be low. Therefore, the detection system can first obtain multiple detection models based on ensemble learning, and then input the above voiceprint detection signals into multiple detection models. Each detection model processes the input voiceprint detection signal and outputs the voiceprint detection result. It should be noted that in order for the ensemble learning model to learn new and correct a posteriori knowledge, the variance between the multiple different voiceprint detection results output by the above different detection models needs to be greater than the preset threshold.

[0048] For example, in order to ensure the stability of the detection model, the above-mentioned integrated learning can adopt the Bagging strategy, and the detection system can first obtain the PSO-SVM model, the convolutional neural network model and the semi-supervised generative adversarial network model based on the Bagging strategy. Then, the detection system can perform pre-processing such as noise reduction and feature extraction on the acquired voiceprint detection signal to ensure that the data input to the model is clean and contains valid information. Finally, the detection system can input the acquired voiceprint detection signal into the above-mentioned multiple detection models respectively, and the multiple detection models respectively output the voiceprint detection results.

[0049] Step 106, summarizing the multiple voiceprint detection results to obtain a target detection result of the distribution network transformer, wherein the target detection result is used to characterize whether a fault occurs in the distribution network.

[0050] In an optional embodiment, the above-mentioned multiple voiceprint detection models have output multiple voiceprint detection results. Considering the integration of the voiceprint detection results of multiple models, the detection system can more comprehensively evaluate the operating status of the transformer and accurately identify early faults and potential problems, thereby improving maintenance efficiency and equipment reliability. The detection system can count and summarize the above-mentioned multiple voiceprint detection results to obtain a target detection result. The detection system can further determine whether a distribution network fault occurs based on the target detection result.

[0051] For example, in order to count and summarize the above multiple voiceprint detection results, the voiceprint detection results of different detection models can be as follows: the voiceprint detection result output by the PSO-SVM model is "abnormal" with a confidence of 0.9, the voiceprint detection result output by the convolutional neural network model is "normal" with a confidence of 0.7, and the voiceprint detection result output by the semi-supervised generative adversarial network model is "abnormal" with a confidence of 0.85. The detection system can count the models with abnormal detection results and obtain a summary result of 2 / 3, that is, the detection results of most models are abnormal, then the summary result is abnormal, which indicates that the distribution network has a fault.

[0052] In another optional embodiment, considering that the weights corresponding to the above-mentioned multiple detection models may be different, which means that different detection models have different degrees of influence on the target detection results, the detection system can use the above-mentioned different weights to perform weighted summation on the above-mentioned multiple voiceprint detection results to obtain the target detection result. The detection system can further determine whether the distribution network is faulty based on the target detection result.

[0053] For example, in order to perform weighted summation on the above multiple voiceprint detection results, the voiceprint detection results of different detection models can be as follows: the voiceprint detection result output by the PSO-SVM model is "abnormal", with a confidence of 0.8, the voiceprint detection result output by the convolutional neural network model is "normal", with a confidence of 0.6, and the voiceprint detection result output by the semi-supervised generative adversarial network model is "abnormal", with a confidence of 0.7. Among them, the normal detection result has a low contribution to fault diagnosis, and its weight can be set to 0. For the abnormal detection results of different models, their weights can be set as follows: the weight of the PSO-SVM model is set to 4, the weight of the convolutional neural network model is set to 3, and the weight of the semi-supervised generative adversarial network model is set to 2. The detection system can pre-set the abnormal judgment threshold to 3. The detection system can perform weighted summary of the detection results based on the weights of the above multiple models. The calculation result is 4.6, which is greater than the above abnormal judgment threshold. The summary result is abnormal, which indicates that the distribution network has a fault.

[0054] In an embodiment of the present invention, a method is adopted in which a voiceprint detection signal generated by a distribution network transformer during operation is obtained; the voiceprint detection signal is processed respectively by multiple detection models to obtain multiple voiceprint detection results; and the multiple voiceprint detection results are summarized to obtain a target detection result of the distribution network transformer. The detection system can pre-train multiple detection models, use the above multiple detection models to process the acquired voiceprint detection signal, and summarize the detection results of the above multiple detection models through an integrated learning strategy to obtain the final target detection result, thereby achieving the purpose of accurately and efficiently identifying early equipment failures in the distribution network, thereby achieving the technical effect of improving the accuracy and efficiency of distribution network fault detection, and further solving the technical problem of low accuracy of fault detection of distribution network transformers in related technologies.

[0055] Furthermore, multiple detection models are used to process the voiceprint detection signal respectively to obtain multiple voiceprint detection results, including: obtaining multiple model identifiers, wherein multiple detection models and multiple model identifiers correspond one to one; performing feature extraction on the voiceprint detection signal based on the multiple model identifiers to obtain multiple voiceprint features, wherein different model identifiers correspond to different voiceprint features; and inputting the corresponding voiceprint features into the corresponding detection model based on the multiple model identifiers to obtain multiple detection results.

[0056] The multiple model identifiers may be identifiers used to distinguish multiple different detection models. The multiple voiceprint features may be features in the voiceprint detection signal that are helpful for fault diagnosis.

[0057] In an optional embodiment, considering that the original voiceprint signal may contain invalid data, directly using these original data may result in a waste of computing resources. Therefore, the detection system may filter out data or noise components that are not related to the detection task through feature extraction to reduce the waste of computing resources, thereby improving the operating efficiency of the detection system. First, the detection system may obtain multiple model identifiers, and the multiple detection models may correspond one-to-one to the multiple model identifiers. Then, the detection system may perform feature extraction on the voiceprint detection signal based on the multiple model identifiers to eliminate redundant data and convert the data into a form suitable for a specific model, thereby obtaining multiple voiceprint features corresponding to the different model identifiers. Finally, the monitoring system inputs the multiple voiceprint features into the detection model corresponding to the multiple model identifiers, so as to output multiple detection results by calculation using the detection model.

[0058] For example, the detection system can first obtain the model identification of the PSO-SVM model, the convolutional neural network model, and the semi-supervised generative adversarial network model, and perform preprocessing such as denoising and standardization on the above voiceprint detection signal. Then, the detection system can perform feature extraction on the preprocessed voiceprint detection signal based on the above model identification, thereby extracting different voiceprint features corresponding to the above model identification. The detection system can input the voiceprint features obtained based on the PSO-SVM model into the PSO-SVM model, input the voiceprint features obtained based on the convolutional neural network model into the convolutional neural network model, and input the voiceprint features obtained based on the semi-supervised generative adversarial network model into the semi-supervised generative adversarial network model. Finally, the above multiple models are calculated and output multiple detection results.

[0059] Furthermore, feature extraction is performed on the voiceprint detection signal based on multiple model identifiers to obtain multiple voiceprint features, including: acquiring multiple feature extraction methods, wherein multiple model identifiers correspond to multiple feature extraction methods one-to-one; feature extraction is performed on the voiceprint detection signal based on multiple feature extraction methods to obtain multiple first features, wherein different feature extraction methods correspond to different first features; in response to the multiple first features containing multiple second features of the same feature type, the multiple second features are matched based on the feature type to obtain feature matching results, wherein the feature matching results are used to reflect the accuracy between different second features; the multiple first features are updated based on the feature matching results to obtain multiple voiceprint features.

[0060] The feature extraction method may be a specific method for the detection model to extract features from the voiceprint detection signal. The first feature may be a feature extracted based on the attribute of the voiceprint detection signal, for example, it may be a spectrum feature, an image feature, a sound feature, but not limited thereto. The second feature may be a feature of the same type among the multiple first features, for example, it may be a spectrum feature extracted by a convolutional neural network model, or a spectrum feature extracted by a semi-supervised generative adversarial network model, but not limited thereto. The feature matching result may be a comparison result of the accuracy of the multiple second features.

[0061] In an optional embodiment, the detection system has obtained the model identifications of the above-mentioned multiple detection models. Considering that multiple identical features may be extracted between different models, in order to ensure the accuracy of the above-mentioned voiceprint detection results, the above-mentioned multiple identical features need to be processed. The detection system can first obtain multiple feature extraction methods corresponding to the above-mentioned multiple model identifications. Then, the detection system can perform feature extraction on multiple attributes of the voiceprint detection signal based on the above-mentioned multiple feature extraction methods to obtain multiple first features corresponding to the above-mentioned multiple feature extraction methods. The above-mentioned multiple first features may include multiple second features of the same feature type. In order to obtain feature values ​​with higher accuracy in the above-mentioned second features, the detection system can match the above-mentioned multiple second features based on the feature type to obtain feature matching results. Finally, the detection system can determine multiple voiceprint features with higher accuracy based on the above-mentioned feature matching results for use in determining the voiceprint detection results.

[0062] For example, the detection system can perform feature extraction on the voiceprint detection signal based on the feature extraction methods of the PSO-SVM model, the convolutional neural network model, and the semi-supervised generative adversarial network model, and obtain multiple different extraction results respectively. The extraction results of the above different models all contain time domain statistical features. The detection system can perform feature matching on multiple feature values ​​of the above time domain statistical features. The feature matching results show that the time domain statistical features extracted by the PSO-SVM model are relatively accurate. Therefore, the detection system can directly retain the time domain statistical features extracted by the PSO-SVM model as the final time domain statistical features.

[0063] For another example, both the convolutional neural network model and the semi-supervised generative adversarial network model can convert the original signal into a Mel time-frequency spectrum (Mel time-frequency spectrum) to obtain the time-frequency domain features of the voiceprint detection signal. However, the time-frequency domain features obtained by the above two models may still not meet the accuracy requirements of the detection system. Therefore, the detection system can use a model that specifically extracts time-frequency domain features to re-extract the time-frequency domain features of the voiceprint detection signal to improve the accuracy of the feature value. For example, the detection system can use a time-frequency convolutional neural network model, but is not limited to this. The detection system can use the re-extracted results as the final time-frequency domain features.

[0064] Furthermore, in response to the inclusion of a particle swarm optimization model in multiple detection models, feature extraction is performed on the voiceprint detection signal based on multiple feature extraction methods to obtain multiple first features, including: identifying the voiceprint detection signal to obtain a transient voiceprint signal and a steady-state voiceprint signal contained in the voiceprint detection signal; filtering the transient voiceprint signal based on a preset filter to obtain a transient filtered signal; performing feature extraction on the steady-state voiceprint signal to obtain a steady-state voiceprint feature, and performing feature extraction on the transient filtered signal to obtain a transient voiceprint feature, wherein the multiple first features include steady-state voiceprint features and transient voiceprint features.

[0065] The above-mentioned transient voiceprint signal may be a signal that changes rapidly in a short period of time and then reaches a steady state or disappears. The above-mentioned steady-state voiceprint signal may be a signal that remains basically unchanged or changes slowly for a long time. The above-mentioned preset filter may refer to a device or circuit that can process and control the voiceprint detection signal. For example, the above-mentioned preset filter may be a low-pass filter, a high-pass filter, or a band-pass filter, but is not limited to this. The above-mentioned transient filtered signal may be a signal obtained after the above-mentioned transient voiceprint signal is processed by the above-mentioned preset filter. The above-mentioned steady-state voiceprint feature may be a signal obtained after feature extraction of the above-mentioned steady-state voiceprint signal. The above-mentioned transient voiceprint feature may be a signal obtained after feature extraction of the above-mentioned transient filtered signal.

[0066] In an optional embodiment, considering that the voiceprint detection model may have arbitrariness and uncertainty when selecting parameters, resulting in poor model performance, for example, overfitting or underfitting scenarios may occur, which further leads to reduced robustness and stability of the model. Therefore, the detection system needs to optimize the input parameters of the model to improve the accuracy of the detection results output by the model. Based on this, the detection system can first identify the voiceprint detection signal to obtain the transient voiceprint signal and the steady-state voiceprint signal contained in the voiceprint detection signal. Among them, for the transient voiceprint signal, the detection system can process it based on a preset filter to obtain a transient filtered signal, and then the detection system can further perform feature extraction on the above-mentioned transient filtered signal to obtain a transient voiceprint feature. For the steady-state voiceprint signal, the detection system can perform feature extraction on it to obtain a steady-state voiceprint feature.

[0067] For example, the above-mentioned transient soundprint signal can be a vibration acceleration signal discrete in the time domain, and the above-mentioned steady-state soundprint signal can be a harmonic distortion signal. The detection system can first identify the soundprint detection signal, and obtain the vibration acceleration signal and the harmonic distortion signal contained in the soundprint detection signal. For the vibration acceleration signal, the detection system can perform a 10Hz-10kHz bandpass filter on it based on a bandpass filter, so as to obtain the transient filtered signal of the vibration acceleration signal, and then the detection system can perform feature extraction on the above-mentioned transient filtered signal to obtain the transient soundprint feature of the vibration acceleration signal. For the harmonic distortion signal, the detection system can perform feature extraction on it to obtain the steady-state soundprint feature of the harmonic distortion signal.

[0068] For ease of understanding, Figure 2 is an optional processing flow chart of a bone conduction voiceprint sensing signal according to an embodiment of the present invention, such as Figure 2As shown, the bone conduction voiceprint signal is passed through a 10Hz-10kHz bandpass filter, and corresponding processing is performed according to the sound characteristics of the above bone conduction voiceprint signal to distinguish the sound characteristics of the bone conduction voiceprint information, thereby obtaining the corresponding voiceprint characteristics, wherein the obtained voiceprint characteristics may include: sharpness calculation, loudness calculation, effective value calculation, kurtosis calculation, roughness calculation, and jitter calculation. Specifically, for the voiceprint detection signal containing the continuous high-frequency sharp sound of the transformer body and the uniform periodic "cluck" metal friction sound of the oil pump, the sharpness calculation can be performed after band-pass filtering; for the voiceprint detection signal containing abnormally increased and obvious abnormal sound, the loudness and effective value calculation can be performed after band-pass filtering; for the voiceprint detection signal containing "squeaking" or "crackling" sound and the irregular non-metallic friction sound of the oil pump, the kurtosis calculation can be performed after band-pass filtering; for the "hissing" sound contained in the voiceprint detection signal, the roughness calculation can be performed after band-pass filtering; for the "bubble" boiling sound and "wow" sound contained in the voiceprint detection signal, the jitter calculation can be performed after band-pass filtering.

[0069] Furthermore, feature extraction is performed on the steady-state voiceprint signal to obtain steady-state voiceprint features, including: framing the steady-state voiceprint signal according to a preset frequency to obtain a steady-state framed signal; performing power identification on the steady-state framed signal to obtain an auto-power spectrum of the steady-state framed signal; performing amplitude identification on the auto-power spectrum to determine the fundamental frequency amplitude and harmonic wave amplitude of the steady-state framed signal; processing the average value of the fundamental frequency amplitude and the average value of the harmonic wave amplitude to obtain the harmonic distortion value of the steady-state voiceprint signal.

[0070] The above-mentioned preset frequency may refer to the signal frequency used to define the frequency range of signal acquisition and analysis. The above-mentioned steady-state framed signal may be a signal obtained after framing the steady-state voiceprint signal. The above-mentioned autopower spectrum may refer to the spectrum used to describe the energy distribution of the above-mentioned steady-state framed signal at different frequencies. The above-mentioned fundamental frequency amplitude may be the intensity of the fundamental frequency component of the above-mentioned steady-state framed signal in the spectrum. The above-mentioned harmonic amplitude may refer to the intensity of other frequency components contained in the above-mentioned steady-state framed signal in addition to the above-mentioned fundamental frequency amplitude. The above-mentioned harmonic distortion value may be the power ratio of the harmonic component to the fundamental frequency component of the above-mentioned steady-state framed signal.

[0071] In an optional embodiment, considering the periodic changes or small fluctuations that may exist inside the transformer of the distribution network, the signal can be divided into a plurality of sufficiently short time periods through frame processing, thereby facilitating the calculation of harmonic distortion. Figure 3 is a flow chart of an optional harmonic distortion calculation according to an embodiment of the present invention, such as Figure 3As shown, the detection system can first intercept the bone conduction voiceprint signal during the steady-state operation of the transformer, and perform frame processing on the above-mentioned steady-state voiceprint signal according to a preset frequency to obtain a steady-state framed signal, and the frequency resolution of the above-mentioned framed signal is not less than 1Hz. Then, the detection system can obtain the autopower spectrum of the steady-state framed signal by calculation, and obtain the fundamental frequency amplitude and harmonic amplitude of the steady-state framed signal based on the autopower spectrum. Finally, the detection system can calculate the average value of the fundamental frequency amplitude and the harmonic amplitude of the above-mentioned steady-state framed signal, and calculate the harmonic distortion value of the steady-state voiceprint signal. The calculation formula of the above-mentioned harmonic distortion value can be shown as follows:

[0072]

[0073] Among them, THD represents the harmonic distortion value, Aharmonic n It represents the average value of the amplitude of the nth harmonic, and Abasic represents the average value of the amplitude of the fundamental frequency.

[0074] For example, in order to calculate the harmonic distortion rate of a steady-state voiceprint signal, the detection system can first divide the continuous steady-state voiceprint signal into multiple frames with a duration of 3 seconds and a frame length of 16384 points, thereby obtaining a steady-state framed signal of the above-mentioned steady-state voiceprint signal. Then, the detection system can calculate the autopower spectrum of the above-mentioned steady-state framed signal, and identify the fundamental frequency amplitude and harmonic amplitude based on the calculated autopower spectrum. Then, for each frame signal, the detection system calculates the average value of the fundamental frequency amplitude to be 50Hz, and calculates the average values ​​of the four harmonic amplitudes to be 100Hz, 150Hz, 200Hz, and 250Hz, respectively. Finally, based on the above formula, the detection system can calculate the harmonic distortion value of the steady-state voiceprint signal to be 7.34, which can be used to evaluate the relative size of the harmonic component and the fundamental frequency component.

[0075] It should be noted that the support vector machine (SVM) in the PSO-SVM model can minimize the risk of the entire structure under the condition that the sample does not approach infinity, and can complete classification and recognition well in the case of small samples. At the same time, the support vector machine introduces some kernel functions with special properties to transform the actual problem into a high-dimensional feature space through nonlinear transformation, and constructs a linear decision function in the high-dimensional space to realize the nonlinear decision function in the original space, so that the process of solving the support vector machine is only related to the number of training samples, and has nothing to do with the sample dimension, that is, the algorithm complexity has nothing to do with the sample dimension.

[0076] For example, suppose the separable data set is (x i ,y i ),i=1,2,…,l, where x i ∈R d, y∈{-1,+1), y is the classification label, d is the dimension of the input vector, solving the hyperplane problem can be transformed into the following optimization problem:

[0077] Among them, y i [(ω*x i )+b]≥1-ξ i ,i=1,…l.

[0078] In the above formula, ω is the normal vector of the hyperplane, b is the constant term, and ξ i is the ith relaxation factor, C is the penalty factor, and l is the number of data in the data set. The classification function of SVM is defined as follows:

[0079]

[0080] Among them, f(x) is the SVM classification function, a i is the Lagrangian operator, b is a constant term, K(x i ,x) is a kernel function, which is used to transform the input data into a high-dimensional feature space. The commonly used kernel function radial basis function (RBF) can be shown as follows:

[0081]

[0082] Among them, σ is the radius of the radial basis function. In order to achieve better fault prediction results, it is very important to select appropriate support vector machine model parameters. In this embodiment, the particle swarm optimization algorithm (PSO) is used to optimize the support vector machine parameters. It should be pointed out that the parameters that need to be optimized in the SVM model are mainly the penalty factor C and the radial basis function radius σ. Furthermore, the detection system can use PSO to optimize SVM, which can greatly improve the classification effect of SVM. Assuming that the search is in an n-dimensional space, a population is composed of m particles, and the position of the i-th particle can be expressed as shown in the following formula:

[0083]

[0084] in, is the position of the ith particle, k is the kth set of eigenvectors, and n is the dimension of the eigenvectors. Assuming that the position of each particle may be a potential solution, Substituting into the objective function f(x) we can get the fitness Fitness is used to measure the quality of the solution. Assume that the flying speed of the i-th particle is as follows:

[0085]

[0086] in, represents the flight speed of the i-th particle, k represents the k-th set of eigenvectors, and n represents the dimension of the eigenvectors. According to the above formula, the optimal solution when reaching the i-th particle can be expressed as follows:

[0087]

[0088] in, represents the optimal solution when the i-th particle is reached, k represents the k-th group of eigenvectors, and n represents the dimension of the eigenvectors. The optimal solution searched in the entire particle swarm can be expressed as follows:

[0089]

[0090] in, represents the optimal solution searched in the entire particle swarm, k represents the kth group of eigenvectors, and n represents the dimension of the eigenvectors. In the process of PSO algorithm optimization, the search position is initialized first. and speed V i 0 , calculate the fitness of each particle and get the optimal solution of the current particle swarm. Then update the particle swarm according to the speed, and continuously update the optimal solution according to the set number of iterations. The update iterative formula for the particle speed and position can be shown as follows:

[0091]

[0092] in, represents the particle velocity of the i-th particle after iteration, represents the position of the i-th particle after iteration, w k is the inertia factor, and w k >0, c1 and c2 are learning factors, θ1 and θ2 are random numbers, and θ1∈[0,1], θ2∈[0,1].

[0093] For ease of understanding, Figure 4 is an optional training flow chart of a distribution network transformer voiceprint detection model based on PSO-SVM according to an embodiment of the present invention, such as Figure 4As shown, the detection system first processes the bone conduction voiceprint signal, and calculates the effective value and loudness, sharpness, kurtosis, roughness, jitter and harmonic distortion based on the voiceprint signal. Then, the detection system extracts features based on the above calculation results and obtains feature vectors. The feature vectors can be used for training the SVM classifier. If the training results show that the current SVM classifier cannot accurately identify voiceprint defects, the parameters of the current model are optimized based on PSO. If the training results show that the current SVM classifier can accurately identify voiceprint defects, the SVM classifier can be deployed and deployed for application. The deployed SVM classifier can accurately classify the above feature vectors and output a classification result, indicating whether the operating status of the distribution network transformer is normal or there is some abnormality.

[0094] Furthermore, in response to the inclusion of a convolutional neural model in multiple detection models, feature extraction is performed on the voiceprint detection signal based on multiple feature extraction methods to obtain multiple first features, including: framing the voiceprint detection signal to obtain a voiceprint frame signal; windowing the voiceprint frame signal to obtain at least one voiceprint signal segment; performing short-time discrete Fourier transform on the voiceprint signal segment to obtain initial time-frequency spectrum characteristics of the voiceprint detection signal; filtering and transforming the initial time-frequency spectrum characteristics to obtain target time-frequency spectrum characteristics, wherein the multiple first features include the target time-frequency spectrum characteristics.

[0095] The above-mentioned voiceprint frame signal may be a signal obtained after framing the voiceprint detection signal. The above-mentioned windowing processing may refer to changing the edge effect of the signal by multiplying the original signal with a window function. For example, the above-mentioned window function may be a Hamming window or a Hann window, but is not limited thereto. The above-mentioned voiceprint signal segment may refer to a signal obtained after windowing the above-mentioned voiceprint frame signal. The above-mentioned initial time-frequency spectrum feature may be a feature that describes the change of the frequency components of the voiceprint detection signal at different time points. The above-mentioned target time-frequency spectrum feature may be a feature obtained after filtering and transforming the above-mentioned initial time-frequency spectrum feature.

[0096] In an optional embodiment, the detection system may first perform frame processing on the voiceprint detection signal to obtain a voiceprint frame signal, and then the detection system multiplies the window function with the voiceprint signal of each frame to obtain the voiceprint signal segment of the voiceprint frame signal, and performs a short-time discrete Fourier transform on the current voiceprint signal segment to obtain the spectrum information of the current voiceprint signal segment, that is, to obtain the initial time-frequency spectrum feature. Finally, the detection system may filter and transform the initial time-frequency spectrum feature through a filter to obtain the target time-frequency spectrum feature.

[0097] For example, in order to reduce the leakage effect in spectrum analysis, that is, to reduce the spectrum aliasing caused by signal truncation, the above windowing operation can be completed based on the Hamming window function, and the calculation formula of the Hamming window can be as follows:

[0098]

[0099] Among them, ω(n) represents the value of the Hamming window function at the sample point with index n, π represents pi, n represents the index of the sample point in the sequence, and N represents the length of the window. express The cosine value of .

[0100] The detection system can cut the above voiceprint detection signal into samples of 3s in length, and use the overlapping segmentation method to divide the frames, with the frame length set to 16384 frames and the frame shift set to 8192 frames, so as to obtain the voiceprint frame signal of the voiceprint detection signal. Then, the detection system can multiply the above Hamming window function with the voiceprint frame signal of each frame, and the calculation result is the voiceprint signal segment of the current voiceprint frame signal. The detection system can further perform a short-time discrete Fourier transform on the current voiceprint signal segment to obtain the time-frequency spectrum of the current voiceprint signal segment. The calculation process of the above short-time discrete Fourier transform can be shown as follows:

[0101] 0≤n and k≤N-1.

[0102] Among them, W(k) represents the time-frequency spectrum of the current voiceprint signal segment, e is the base of the natural logarithm, ω(n) represents the value of the Hamming window function at the sample point with index n, k represents the current voiceprint signal segment, and m represents the index of the component. Finally, the detection system can use Mel filtering to filter and transform the above initial time-frequency spectrum features to obtain the target time-frequency spectrum features of the voiceprint detection signal.

[0103] Furthermore, the initial time-frequency spectrum features are filtered and transformed to obtain target time-frequency spectrum features, including: obtaining a component matrix of a filter component, wherein the filter component is used to filter and transform the initial time-frequency spectrum features; performing a matrix transformation on the initial time-frequency spectrum features to obtain an initial matrix; multiplying the initial matrix with the component matrix to obtain a filter matrix; and performing logarithm processing on the filter matrix to obtain the target time-frequency spectrum features.

[0104] The filter component may be a combination of multiple filters, for example, multiple Mel filters, multiple triangular filters, multiple Mel filters and triangular filters, but is not limited thereto. The component matrix may be a matrix composed of the multiple filters.

[0105] In an optional embodiment, the detection system needs to map the spectrum to a frequency scale that is closer to human ear perception, so as to extract features that are more important for identification. The detection system can first obtain the component matrix of the filter component to facilitate the subsequent filtering conversion of the above-mentioned initial time-frequency spectrum features. The detection system can convert the above-mentioned initial time-frequency spectrum features based on the obtained component matrix to obtain an initial matrix, and then multiply the initial matrix with the above-mentioned component matrix to obtain a filter matrix. Finally, the detection system can perform logarithmic processing on the obtained filter matrix to obtain the target time-frequency spectrum features.

[0106] For example, the detection system can first transform the voiceprint time-frequency spectrum into the Mel time-frequency spectrum under the Mel scale based on Mel filtering. The conversion relationship can be shown as follows:

[0107] Mel(k)=25951g(1+f / 700).

[0108]

[0109] Where f is the frequency of the normal scale, 0≤f≤48000, k is the Mel scale frequency, and the unit is Hz. The detection system can use the Mel filter group to perform triangular filtering and amplification processing on the frequency domain features of low frequency and medium-low frequency, and perform triangular filtering and weight reduction processing on the medium-high frequency and high frequency features. The above operation is conducive to describing the energy characteristics of the stable sound wave signal in the operation state of the equipment. The Mel filter group uses an equal-height triangular filter group to filter the linear time-frequency spectrum. The above triangular filter group is characterized by having equal-width filters, and its function is shown in the following formula:

[0110]

[0111] Wherein, m is the filter group number, f(m) represents the center frequency of the mth filter, and in this embodiment, m can be set to the following range: 1≤m≤100. The center frequency of the triangular filter can be calculated according to the following formula:

[0112]

[0113] Among them, f max Indicates the maximum frequency of the filter’s filtering range, f min Indicates the minimum frequency of the filter’s filtering range, f s is the sampling frequency of the sensor, and N is the frame length of the discrete Fourier transform. The detection system can obtain the Mel time-frequency spectrum by multiplying the time-frequency spectrum matrix with the Mel filter bank matrix and then taking the logarithm.

[0114] It should be noted that, in this embodiment, the Mel time-frequency spectrum can be used as the input of the convolutional neural network for deep learning to form a Mel time-frequency spectrum-convolutional neural network recognition model, thereby realizing the extraction of voiceprint features and pattern recognition of different operating states.

[0115] The Mel time-frequency spectrum input image size is [376×100×3], where 376 represents the time domain component, 100 represents the frequency domain component under the Mel scale, and 3 represents the RGB color channel of the power spectrum density of the voiceprint signal. By inputting the Mel time-frequency spectrum graph into the convolutional neural network, the training of the weights in the network structure is completed, and the pattern recognition of different voiceprint signals is realized. In this embodiment, according to the characteristics of the distribution network transformer voiceprint data set, the traditional convolutional neural network structure is modified to design a neural network for voiceprint recognition. The network structure can be shown in Table 1 below:

[0116] Table 1

[0117]

[0118]

[0119] The ReLU in the above table is an activation function used in artificial neural networks, Dropout is a regularization technique used to prevent model overfitting, and softmax is a commonly used mathematical function for solving multi-classification problems.

[0120] For ease of understanding, Figure 5 is an optional training flow chart of a distribution network transformer voiceprint detection model based on a convolutional neural network according to an embodiment of the present invention, such as Figure 5As shown, first, the detection system can collect the voiceprint signal generated during the operation of the distribution network transformer through the bone conduction voiceprint sensor. The detection system can perform denoising, framing, windowing and short-time discrete Fourier transform (Fast Fourier Transform, FFT) on the collected voiceprint signal to obtain the voiceprint time-frequency spectrum of the above voiceprint signal. Then, the detection system can convert the above voiceprint time-frequency spectrum into Mel time-frequency spectrum by Mel filtering, taking logarithm, and discrete cosine transform (Discrete Cosine Transform, DCT). Further, the detection system can perform dimensionality reduction processing on the above Mel time-frequency spectrum through principal component analysis (Principal Component Analysis, PCA) and extract features from it through feature matrix. The detection system can train the voiceprint detection model based on the extracted features of the deep learning of the convolutional neural network algorithm. In addition, the detection system can also train the above voiceprint detection model based on the normal sample library of the equipment and the abnormal sample library of the equipment to help the model understand the characteristics and patterns of normal operation, and enable the model to distinguish and predict abnormal states. The trained voiceprint detection model can perform difference comparison and voiceprint recognition based on the classification output of the convolutional neural network model. If the above voiceprint detection model predicts an abnormal voiceprint signal, it can prompt the system with the current status, fault type and fault location of the equipment, as well as whether there are unknown abnormalities. Experts can judge its operating status by manual listening and mark and classify the above voiceprint data, thereby improving the above-mentioned equipment normal sample library and equipment abnormal sample library.

[0121] Furthermore, the method also includes: resampling the training data set based on multiple model identifiers to obtain a resampled data set; training multiple detection models based on the resampled data set to obtain model training results; and assigning multiple detection weights to the multiple detection models based on the model training results, wherein the multiple detection weights are used to aggregate multiple voiceprint detection results to obtain target detection results.

[0122] The resampled data set may be a data set obtained by resampling the training data set in order to construct a more balanced training set. The detection weight may be a numerical value reflecting the different importance or influence of the detection model.

[0123] In an optional embodiment, considering that a single detection model may have advantages in some aspects but shortcomings in other aspects, the detection system can use the advantages of multiple different models to make up for each other's weaknesses, thereby obtaining more accurate and robust voiceprint detection results. The detection system can use the Bagging integration strategy to implement the training of the above multiple detection models. The basic framework pseudo code of the Bagging integration strategy is shown in Table 2 below:

[0124] Table 2

[0125]

[0126] in, represents a data set, represents the sub-model algorithm, T represents the number of sub-models and the number of training rounds, Represents the dataset Generate Boostrap sampling set, h m Represents the submodel, y = h m (x) represents the detection result of the sub-model, and H(x) represents the final integration result.

[0127] The detection system can first identify the above training data set based on different models Perform bootstrap sampling to obtain a resampled data set Then, the detection system can be based on the above resampled dataset The above-mentioned multiple detection models are trained to obtain model training results. Finally, the detection system can assign multiple detection weights to the multiple detection models based on the above-mentioned model training results, so as to obtain the above-mentioned target detection results.

[0128] For example, the above-mentioned multiple detection models can be a PSO-SVM model, a convolutional neural network model, and a semi-supervised generative adversarial network model. In order to summarize the voiceprint detection results of multiple detection models, the detection system can first generate multiple Boostrap sampling subsets from the original training set by bootstrapping sampling, wherein the size of each of the above subsets can be the same as the original training set, but the samples contained can be different, so that the above three detection models can learn different data features. Then, the detection system can use the above-mentioned generated subsets to train the above three detection models separately. It should be noted that the detection system can train each of the above detection models independently to ensure that each detection model will learn different decision boundaries based on different data sets, thereby increasing the diversity between models. Finally, the prediction probability of each detection model will be weighted according to its performance (accuracy, recall rate, etc.) on the validation set, and the detection model with a higher weight will have a greater impact on the final prediction result.

[0129] Specifically, here we take the semi-supervised generative adversarial network model as an example. Figure 6 is an optional training flow chart of a voiceprint detection model based on a semi-supervised generative adversarial network according to an embodiment of the present invention, such as Figure 6As shown in the figure, the semi-supervised generative adversarial network can include a generator and multiple discriminators, wherein the generator can generate fake samples through random noise, and the discriminator can not only distinguish true from false, but also classify real data. The generator generates fake samples based on random noise, and then trains the discriminator 2 together with real samples with K categories. The discriminator 2 is a standard classifier, including supervised and unsupervised parts. The supervised part is used to realize the classification of categories in K, and the unsupervised part is used to realize the classification of true and false. Therefore, its output result is a K+1-dimensional vector, and the extra category is the classification of the fake samples generated by the generator. Because the discriminator 1 and the discriminator 2 share parameters, the discriminator 1 will automatically obtain the parameters of the discriminator 2 after training. Then the discriminator 1 is used to complete the training of the generator. At this time, the parameters of the discriminator 1 are kept unchanged, and the parameters of the generator are adjusted to make the generated results meet the standards of the discriminator 1. After multiple alternating training, when the generator and the discriminator can no longer improve their own capabilities, it can be considered that they have reached a Nash equilibrium. At this time, the final trained model will be output.

[0130] Furthermore, obtaining a voiceprint detection signal generated by the distribution network transformer during operation includes: performing voiceprint detection on at least one transformer component included in the distribution network transformer through a voiceprint sensor to obtain an initial detection signal, wherein the voiceprint sensor is deployed on the transformer component, the voiceprint sensor corresponds to the transformer one-to-one, and the transformer component includes at least: a transformer body and a cooling component; performing noise preprocessing on the initial detection signal to obtain a voiceprint detection signal.

[0131] The above-mentioned voiceprint sensor may be a sensor for capturing and converting sound or vibration signals into electronic signals, for example, a bone conduction voiceprint sensor, a vibration voiceprint sensor, but not limited thereto. The above-mentioned initial detection signal may be an original voiceprint signal captured directly from the above-mentioned voiceprint sensor from the power distribution network transformer.

[0132] For ease of understanding, Figure 7 is a structural diagram of an optional bone conduction voiceprint detection sensor according to an embodiment of the present invention, such as Figure 7 As shown, the sensor specification is 5 / 8-24UNEF-2A, with a total length of 51.00 mm, of which the sleeve length is 34.70 mm. The base is hexagonal, with a base height of 6.00 mm, a base inscribed circle radius of 22.00 mm, a base hexagonal maximum width of 24.8 mm, and M6 means a thread diameter of 6 mm.

[0133] In an optional embodiment, considering that the above initial detection signal may contain a large amount of interference information, the detection system may perform noise preprocessing on the above initial detection signal. Specifically, the detection system may first perform voiceprint detection on the distribution network transformer component through a voiceprint sensor deployed on the distribution network transformer body or cooling component to obtain the above initial detection signal. Then, the detection system may use filtering technology to remove noise irrelevant to fault detection and retain signal components related to fault characteristics, thereby obtaining a voiceprint detection signal.

[0134] For example, in order to meet the needs of voiceprint detection and analysis of distribution network transformers, an industrial-grade single-axis IEPE (Integrated Electronics Piezoelectric) voltage output bone conduction voiceprint sensor can be used. The specific performance indicators of the bone conduction voiceprint sensor are shown in Table 3 below:

[0135] Table 3

[0136]

[0137]

[0138]

[0139] The detection system can capture the sound on the distribution network transformer body through the bone conduction voiceprint sensor and convert it into an electrical signal to generate an initial detection signal. Then, the detection system can remove the noise in the above initial detection signal through bandpass filtering to obtain the voiceprint detection signal.

[0140] For example, in addition to meeting the basic voiceprint detection function, it is also necessary to ensure that the entire device coordinates to handle the problems of no network and no power supply interface. The detection system can be based on an edge computing distribution network transformer voiceprint detection device. The device includes a dual-channel high-precision data acquisition module, a solar power supply module, and an industrial wireless gateway. Through structured packaging, it can ensure that it can realize the sound detection of distribution network transformers without a network interface or a power supply interface.

[0141] For ease of understanding, Figure 8 is a schematic diagram of an optional constant current source conditioning circuit according to an embodiment of the present invention, such as Figure 8As shown, the power input voltage is set to 24V. The positive pole of the power supply is usually connected to the power rail in the circuit, while the negative pole is connected to the common ground of the circuit, and the output end is connected to the bone conduction voiceprint sensor ICP_POWER1. Q4 is a transistor, the base of which is connected to a reference voltage, and the emitter is connected to the power rail of the circuit to receive the power supply voltage. R98 is a reference resistor, connected between the base and emitter of transistor Q4, used to set the reference voltage of the circuit. R97 is a load resistor, and R97 or the entire load circuit is connected to the collector of transistor Q4 to ensure that the sensor can work stably. C86 is a signal isolation capacitor, connected between the collector of Q4 and the load R97. U24 TL431 is a precision reference voltage source, which provides a stable constant current source for ICP_POWER1. The protection circuit is connected between the power supply and transistor Q4 or between the load R97 and the common ground to prevent power supply abnormalities or sudden load changes from damaging the circuit or sensor. C61 is a power supply filter capacitor, used to remove high-frequency noise and fluctuations in the power supply.

[0142] Fig. 9 is a schematic diagram of an optional 24-bit dual ADC (Analog-to-Digital Converter) acquisition technology according to an embodiment of the present invention, such as Fig. 9 As shown, ICP (Integrated Circuit Piezoelectric) represents an integrated electronic piezoelectric sensor, which is used to monitor the vibration and noise of the transformer during operation. KA1 is a programmable switch used to control the on and off of the sensor signal. K1 is a relay switch used to control the on and off of the signal. FPGA (Field-Programmable Gate Array) is a programmable logic device, and the FPGA control and data processing unit is used for data processing and control. The 1G network port is a 1 Gigabit per second Ethernet interface used to connect the device to a local area network or other network devices. For the bone conduction voiceprint signal, it is connected to the input end of the vibration or noise sensor, and the ICP power supply voltage is loaded to the sensor through the programmable switch KA1, providing the sensor with a voltage of 24VDC and a constant current of 4mA. After sensing the vibration or noise signal, the internal piezoelectric crystal generates a weak voltage, which is amplified and output by the amplification circuit in the sensor, and the 24V DC voltage of the power supply is superimposed, and output together to one end of the capacitor. At this time, the program control K1 is in an open circuit state, and the signal amplified by the sensor superimposed on the DC 24VDC is isolated at the capacitor input end after being blocked by the capacitor. The output of the capacitor is the alternating vibration signal amplified by the sensor.

[0143] For ease of understanding, Fig.10 is an optional voiceprint detection flow chart according to an embodiment of the present invention, such as Fig.10As shown, in general, the process is divided into three steps: voiceprint detection database construction, voiceprint detection model construction optimization, and voiceprint detection model deployment and application. Specifically, the bone conduction sensor is responsible for the collection of bone conduction voiceprint signals, and professionals perform voiceprint audio listening and marking to form a voiceprint database containing normal and abnormal states. Then, the bone conduction signal is preprocessed to obtain the steady-state signal time-frequency diagram for building a convolutional neural network model. At the same time, transient signal features are extracted to build a PSO-SVM model. Through an integrated learning strategy, an integrated learning detection model is constructed based on the above convolutional neural network model and PSO-SVM model. The above bone conduction voiceprint signal can be input into the above integrated learning detection model for real-time analysis. The model will output the transformer voiceprint detection result according to the characteristics of the voiceprint signal. If the detection result indicates that the voiceprint signal is abnormal, the voiceprint audio file and the detection result will be sent to the PC and Android terminals through the 4G cloud platform so that the operation and maintenance personnel can take timely measures.

[0144] In order to facilitate a more detailed understanding of the embodiments of the present invention, Fig.11 is an optional fault detection flow chart based on a semi-supervised generative adversarial network according to an embodiment of the present invention, such as Fig.11 As shown, first, the detection system can extract the time-frequency feature image by wavelet transforming the vibration signal collected by the bone conduction voiceprint sensor. Then, the detection system can divide the extracted time-frequency image into two parts: a training set and a test set, and normalize each part so that it can converge quickly during the subsequent network training. Then, the detection system can generate the generator and discriminator of the adversarial network, that is, determine what type of neural network is used by the generator and the discriminator, and their specific network structures. The detection system can alternately train the discriminator and the generator until they reach the above-mentioned Nash equilibrium. Finally, the detection system can input the test set into the trained discriminator for fault diagnosis and output the results of the fault diagnosis.

[0145] Fig.12 is a schematic diagram of the structure of an optional generator and discriminator according to an embodiment of the present invention, such as Fig.12 As shown in the figure, the upper part is the generator, which uses a deconvolution structure, and the lower part is the discriminator, which uses a convolution structure. In the generator, a 100-dimensional uniform distribution Z is reconstructed into a small spatial range convolution representation through many feature maps, and finally converted into a 128×128×3 image. The size of the deconvolution kernel is 5×5. In the discriminator, there are 5 convolution layers, the size of the convolution kernel is 3×3, and the output layer is a fully connected layer.

[0146] According to an embodiment of the present invention, an embodiment of a fault detection device for a distribution network transformer is provided. It should be noted that the device can be used to execute the above-mentioned fault detection method for a distribution network transformer. The specific implementation method and application scenario are the same as those of the above-mentioned embodiment, and will not be described in detail here.

[0147] Fig.13 is a schematic diagram of a fault detection device for a distribution network transformer according to an embodiment of the present invention, such as Fig.13 As shown, the device comprises:

[0148] The signal acquisition module 1302 is used to acquire the voiceprint detection signal generated by the distribution network transformer during operation.

[0149] The voiceprint detection module 1304 is used to use multiple detection models to process the voiceprint detection signal respectively to obtain multiple voiceprint detection results, wherein the variance between the voiceprint detection results output by different detection models is greater than a preset threshold, and the multiple detection models are obtained through integrated learning.

[0150] The result aggregation module 1306 is used to aggregate multiple voiceprint detection results to obtain a target detection result of the distribution network transformer, wherein the target detection result is used to indicate whether a fault occurs in the distribution network.

[0151] Furthermore, the voiceprint detection model is also used to: use multiple detection models to process the voiceprint detection signal respectively to obtain multiple voiceprint detection results, including: obtaining multiple model identifiers, wherein the multiple detection models and the multiple model identifiers correspond one to one; based on the multiple model identifiers, feature extraction is performed on the voiceprint detection signal to obtain multiple voiceprint features, wherein different model identifiers correspond to different voiceprint features; based on the multiple model identifiers, the corresponding voiceprint features are input into the corresponding detection model to obtain multiple detection results.

[0152] Furthermore, the voiceprint detection model is also used to: perform feature extraction on the voiceprint detection signal based on multiple model identifiers to obtain multiple voiceprint features, including: obtaining multiple feature extraction methods, wherein multiple model identifiers correspond to multiple feature extraction methods one by one; perform feature extraction on the voiceprint detection signal based on multiple feature extraction methods to obtain multiple first features, wherein different feature extraction methods correspond to different first features; in response to multiple first features containing multiple second features of the same feature type, match the multiple second features based on the feature type to obtain feature matching results, wherein the feature matching results are used to reflect the accuracy between different second features; and update the multiple first features based on the feature matching results to obtain multiple voiceprint features.

[0153] Furthermore, the voiceprint detection model is also used to: in response to the inclusion of a particle swarm optimization model in multiple detection models, perform feature extraction on the voiceprint detection signal based on multiple feature extraction methods to obtain multiple first features, including: identifying the voiceprint detection signal to obtain a transient voiceprint signal and a steady-state voiceprint signal contained in the voiceprint detection signal; filtering the transient voiceprint signal based on a preset filter to obtain a transient filtered signal; performing feature extraction on the steady-state voiceprint signal to obtain a steady-state voiceprint feature, and performing feature extraction on the transient filtered signal to obtain a transient voiceprint feature, wherein the multiple first features include steady-state voiceprint features and transient voiceprint features.

[0154] Furthermore, the voiceprint detection model is also used to: extract features from steady-state voiceprint signals to obtain steady-state voiceprint features, including: framing the steady-state voiceprint signals according to a preset frequency to obtain steady-state framed signals; performing power identification on the steady-state framed signals to obtain the autopower spectrum of the steady-state framed signals; performing amplitude identification on the autopower spectrum to determine the fundamental frequency amplitude and harmonic wave amplitude of the steady-state framed signals; processing the average value of the fundamental frequency amplitude and the average value of the harmonic wave amplitude to obtain the harmonic distortion value of the steady-state voiceprint signal.

[0155] Furthermore, the voiceprint detection model is also used to: in response to the inclusion of a convolutional neural model in multiple detection models, perform feature extraction on the voiceprint detection signal based on multiple feature extraction methods to obtain multiple first features, including: performing frame processing on the voiceprint detection signal to obtain a voiceprint frame signal; performing windowing processing on the voiceprint frame signal to obtain at least one voiceprint signal segment; performing short-time discrete Fourier transform on the voiceprint signal segment to obtain an initial time-frequency spectrum feature of the voiceprint detection signal; performing filtering conversion on the initial time-frequency spectrum feature to obtain a target time-frequency spectrum feature, wherein the multiple first features include the target time-frequency spectrum feature.

[0156] Furthermore, the voiceprint detection model is also used to: perform filtering transformation on the initial time-frequency spectrum features to obtain target time-frequency spectrum features, including: obtaining a component matrix of a filter component, wherein the filter component is used to perform filtering transformation on the initial time-frequency spectrum features; perform matrix transformation on the initial time-frequency spectrum features to obtain an initial matrix; multiply the initial matrix by the component matrix to obtain a filtering matrix; and perform logarithm processing on the filtering matrix to obtain the target time-frequency spectrum features.

[0157] Furthermore, the result summary module is also used to: resample the training data set based on multiple model identifiers to obtain a resampled data set; train multiple detection models based on the resampled data set to obtain model training results; assign multiple detection weights to multiple detection models based on the model training results, wherein the multiple detection weights are used to summarize multiple voiceprint detection results to obtain target detection results.

[0158] Furthermore, the signal acquisition module is also used to: obtain the voiceprint detection signal generated by the distribution network transformer during operation, including: performing voiceprint detection on at least one transformer component included in the distribution network transformer through a voiceprint sensor to obtain an initial detection signal, wherein the voiceprint sensor is deployed on the transformer component, the voiceprint sensor corresponds to the transformer one-to-one, and the transformer component includes at least: a transformer body and a cooling component; performing noise preprocessing on the initial detection signal to obtain a voiceprint detection signal.

[0159] An embodiment of the present application further provides an electronic device, comprising: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of the present invention when running.

[0160] An embodiment of the present application further provides a computer-readable storage medium, which includes a stored executable program, wherein when the executable program is running, the device where the computer-readable storage medium is located is controlled to execute the methods in various embodiments of the present invention.

[0161] An embodiment of the present application further provides a computer program product, including a computer program, which implements the methods in various embodiments of the present invention when executed by a processor.

[0162] An embodiment of the present application further provides a computer program product, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium is used to store a computer program, and when the computer program is executed by a processor, the method in each embodiment of the present invention is implemented.

[0163] The embodiments of the present application further provide a computer program, which implements the methods in the above-mentioned embodiments of the present invention when executed by a processor.

[0164] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0165] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0166] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0167] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0168] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0169] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting a fault of a distribution network transformer, characterized in that: include: Acquire the voiceprint detection signal generated by the distribution network transformer during operation; Using multiple detection models to process the voiceprint detection signal respectively to obtain multiple voiceprint detection results, wherein the variance between the voiceprint detection results output by different detection models is greater than a preset threshold, and the multiple detection models are obtained by an integrated learning method; The multiple voiceprint detection results are summarized to obtain a target detection result of the distribution network transformer, wherein the target detection result is used to characterize whether a fault occurs in the distribution network.

2. The method according to claim 1, characterized in that The voiceprint detection signal is processed using multiple detection models to obtain multiple voiceprint detection results, including: Acquire a plurality of model identifiers, wherein the plurality of detection models correspond to the plurality of model identifiers one by one; Extracting features of the voiceprint detection signal based on the multiple model identifiers to obtain multiple voiceprint features, wherein different model identifiers correspond to different voiceprint features; Based on the multiple model identifiers, the corresponding voiceprint features are input into the corresponding detection model to obtain multiple detection results.

3. The method according to claim 2, characterized in that Feature extraction is performed on the voiceprint detection signal based on the multiple model identifiers to obtain multiple voiceprint features, including: Acquire multiple feature extraction methods, wherein the multiple model identifiers correspond to the multiple feature extraction methods one by one; Performing feature extraction on the voiceprint detection signal based on the multiple feature extraction methods to obtain multiple first features, wherein different feature extraction methods correspond to different first features; In response to the plurality of first features including a plurality of second features of the same feature type, matching the plurality of second features based on the feature type to obtain a feature matching result, wherein the feature matching result is used to reflect the accuracy between different second features; The multiple first features are updated based on the feature matching results to obtain the multiple voiceprint features.

4. The method according to claim 3, characterized in that In response to the multiple detection models including a particle swarm optimization model, feature extraction is performed on the voiceprint detection signal based on the multiple feature extraction methods to obtain multiple first features, including: Identifying the voiceprint detection signal to obtain a transient voiceprint signal and a steady-state voiceprint signal contained in the voiceprint detection signal; Filtering the transient voiceprint signal based on a preset filter to obtain a transient filtered signal; Feature extraction is performed on the steady-state voiceprint signal to obtain a steady-state voiceprint feature, and feature extraction is performed on the transient filtered signal to obtain a transient voiceprint feature, wherein the multiple first features include the steady-state voiceprint feature and the transient voiceprint feature.

5. The method according to claim 4, characterized in that Extracting features from the steady-state voiceprint signal to obtain steady-state voiceprint features includes: Performing frame processing on the steady-state voiceprint signal according to a preset frequency to obtain a steady-state frame signal; Performing power identification on the steady-state framed signal to obtain an autopower spectrum of the steady-state framed signal; Perform amplitude recognition on the autopower spectrum to determine the fundamental frequency amplitude and harmonic amplitude of the steady-state frame signal; The average value of the fundamental frequency amplitude and the average value of the harmonic amplitude are processed to obtain the harmonic distortion value of the steady-state voiceprint signal.

6. The method according to claim 3, characterized in that In response to the multiple detection models including a convolutional neural model, feature extraction is performed on the voiceprint detection signal based on the multiple feature extraction methods to obtain multiple first features, including: Performing frame processing on the voiceprint detection signal to obtain a voiceprint frame signal; Performing windowing processing on the voiceprint frame signal to obtain at least one voiceprint signal segment; Performing short-time discrete Fourier transform on the voiceprint signal segment to obtain initial time-frequency spectrum characteristics of the voiceprint detection signal; The initial time-frequency spectrum feature is filtered and transformed to obtain a target time-frequency spectrum feature, wherein the plurality of first features include the target time-frequency spectrum feature.

7. The method according to claim 6, characterized in that The initial time-frequency spectrum features are filtered and converted to obtain target time-frequency spectrum features, including: Acquire a component matrix of a filter component, wherein the filter component is used to perform filtering conversion on the initial time-frequency spectrum feature; Performing matrix transformation on the initial time-frequency spectrum features to obtain an initial matrix; Multiplying the initial matrix by the component matrix to obtain a filter matrix; The filter matrix is ​​logarithmically processed to obtain the target time-frequency spectrum characteristics.

8. The method according to claim 2, characterized in that: The method further comprises: Resampling the training data set based on the multiple model identifiers to obtain a resampled data set; Training the multiple detection models based on the resampled data set to obtain model training results; Based on the model training results, multiple detection weights are allocated to the multiple detection models, wherein the multiple detection weights are used to aggregate the multiple voiceprint detection results to obtain the target detection result.

9. The method according to claim 1, characterized in that: Acquire the voiceprint detection signal generated by the distribution network transformer during operation, including: Performing voiceprint detection on at least one transformer component included in the distribution network transformer through a voiceprint sensor to obtain an initial detection signal, wherein the voiceprint sensor is deployed on the transformer component, the voiceprint sensor corresponds to the transformer one by one, and the transformer component at least includes: a transformer body and a cooling component; The initial detection signal is subjected to noise preprocessing to obtain the voiceprint detection signal.

10. A fault detection device for a distribution network transformer, characterized in that: include: A signal acquisition module is used to acquire the voiceprint detection signal generated by the distribution network transformer during operation; A voiceprint detection module, configured to process the voiceprint detection signal using a plurality of detection models to obtain a plurality of voiceprint detection results, wherein the variance between the voiceprint detection results output by different detection models is greater than a preset threshold, and the plurality of detection models are obtained by an integrated learning method; The result aggregation module is used to aggregate the multiple voiceprint detection results to obtain the target detection result of the distribution network transformer, wherein the target detection result is used to characterize whether the distribution network has a fault.

Images