A transformer abnormal sound detection method, device, system and storage medium

Abstract

The application relates to a transformer abnormal sound detection method, device, system and storage medium, the method comprising: acquiring audio information of a to-be-tested transformer and a voltage grade to which the to-be-tested transformer belongs, and extracting an audio feature based on the audio information; inputting the audio feature into a trained first network, outputting a plurality of first posterior probabilities from the first network, the first network being a voltage grade multi-classification network, and each first posterior probability being a posterior probability that the audio feature corresponds to a voltage grade; taking the voltage grade to which the to-be-tested transformer belongs as a target voltage grade, determining a first posterior probability corresponding to the target voltage grade from the plurality of first posterior probabilities as a first target posterior probability; and determining whether the audio information is abnormal based on the first target posterior probability. The application has the effect of not needing abnormal training data, and is more accurate in judging positive abnormalities in a subdivided scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of audio detection, and more particularly to a transformer abnormal sound detection method, device, system and storage medium. BACKGROUND

[0002] At present, the industrial intelligent manufacturing in China has entered an era of rapid development, and the normal operation of industrial equipment is becoming increasingly important. Once a sudden failure of industrial equipment occurs, it will not only bring great loss to the industrial manufacturing of a region, but also have great secondary safety hazards, causing significant economic losses, and even may lead to catastrophic casualties or adverse social impact. In the power and other energy scenarios, with the increasing demand for high-voltage and ultra-high-voltage power grid construction, the country has put forward higher and higher requirements for the reliability of the power system, especially the transformer (such as economic losses and even large-scale social shutdown caused by failure). Therefore, it is more and more important to warn the failure and abnormality of the transformer.

[0003] At present, the failure of the transformer can be detected by machine recognition of the sound emitted during the failure, that is, the abnormal sound detection technology. The specific detection principle of the abnormal sound detection is that the transformer is subjected to electric, magnetic, mechanical and other stress to produce vibration, and the mechanical wave formed by the vibration is converted into an audio signal through solid-liquid-gas medium transmission and is captured by an acoustic sensor (i.e. a microphone). The audio signal contains a large amount of time-frequency domain feature information. After the device is abnormal, the sound also changes, and combined with artificial intelligence means such as deep learning, the device state can be effectively judged.

[0004] In the field of transformer abnormal sound detection, the current common scheme is mainly to train an auto-encoder (AE) network based on the acoustic features of audio. After training, the acoustic features of the audio sample to be tested are inferred through the network to obtain the reconstructed features, and the residual error before and after feature reconstruction is used to determine whether it is abnormal. If the residual error is large, it is considered to be abnormal, otherwise it is normal.

[0005] The abnormal sound detection method in the related art only needs normal training data, and the labeling cost is small, but it depends on the labeling accuracy of the training data. If a small amount of abnormal data is mixed in the training data, the model may incorrectly adapt to the abnormal data, and the training is prone to overfitting, resulting in poor fitting effect of the algorithm on other scene samples, and the coverage of the normal training data. If other normal situation data not covered in the normal sample space appears, the residual error will increase, and the false alarm of abnormality will increase.

[0006] Based on the above defects, the present application provides a transformer abnormal sound detection scheme to solve at least one of the above problems. SUMMARY

[0007] The present application is proposed to address the above-mentioned problems. According to one aspect of the present application, a method for detecting abnormal noise in a transformer is provided, the method comprising: obtaining audio information of a transformer to be tested and the voltage level of the transformer to be tested, extracting audio features based on the audio information; inputting the audio features into a trained first network, having the first network output multiple first posterior probabilities, the first network being a voltage level multi-classification network, each first posterior probability being a posterior probability that the audio feature corresponds to a voltage level; taking the voltage level of the transformer to be tested as a target voltage level, determining a first posterior probability corresponding to the target voltage level from the multiple first posterior probabilities as a first target posterior probability; and determining whether the audio information is abnormal based on the first target posterior probability.

[0008] In one embodiment of the present application, after determining the first target posterior probability, the method further includes: inputting the audio feature into multiple trained second networks, each second network outputting a second posterior probability to obtain multiple second posterior probabilities, and selecting the maximum value from the multiple second posterior probabilities as the second target posterior probability; wherein each second network is a binary classification network corresponding to a voltage level in the first network, and each second posterior probability is the posterior probability that the audio feature is a normal audio feature at a voltage level; determining whether the audio information is abnormal based on the first target posterior probability includes: determining whether the audio information is abnormal based on the first target posterior probability and the second target posterior probability.

[0009] In one embodiment of the present application, the training data of the second network includes the training data of the first network and interference sample data, wherein the training data of the first network includes normal sample data, and the interference sample data is obtained by screening the data in the open source set through the trained first network.

[0010] In one embodiment of the present application, the training data of the second network includes the training data of the first network and interference sample data, wherein the training data of the first network includes normal sample data, and the interference sample data is obtained by screening the data in the open source set through the trained first network.

[0011] In one embodiment of the present application, the second network is trained using data of the voltage level corresponding to the second network as normal class training data, and using data of other voltage levels and the interference class sample data as abnormal class pseudo-label training data.

[0012] In one embodiment of the present application, the data in the open source set is screened by the trained first network to obtain the interference class sample data, including: extracting features from each data to be screened in the open source set, inputting the extracted features into the trained first network, and when the first posterior probability output by the first network for the extracted features is greater than a first threshold, using the data to be screened as the interference class sample data.

[0013] In one embodiment of the present application, determining whether the audio information has any abnormality based on the first target posterior probability and the second target posterior probability includes: performing a weighted summation of the first target posterior probability and the second target posterior probability to obtain a third target posterior probability; comparing the third target posterior probability with a second threshold; when the third target posterior probability is greater than or equal to the second threshold, determining that the audio information has no abnormality, otherwise determining that the audio information has an abnormality.

[0014] In one embodiment of the present application, determining whether the audio information has any abnormality based on the first target posterior probability includes: comparing the first target posterior probability with a first threshold; when the first target posterior probability is greater than or equal to the first threshold, determining that the audio information has no abnormality, otherwise determining that the audio information has an abnormality.

[0015] In one embodiment of the present application, the audio feature is a filter bank feature.

[0016] According to another aspect of the present application, a transformer abnormal sound detection device is provided, the device comprising: an audio acquisition module for acquiring audio information of a transformer to be tested and a voltage level to which the transformer to be tested belongs, and extracting audio features based on the audio information; a detection module for inputting the audio features into a trained first network, whereby the first network outputs a plurality of first posterior probabilities, the first network being a voltage level multi-classification network, and each first posterior probability being a posterior probability that the audio feature corresponds to a voltage level; a matching module for taking the voltage level to which the transformer to be tested belongs as a target voltage level, and determining a first posterior probability corresponding to the target voltage level from the plurality of first posterior probabilities as a first target posterior probability; and an evaluation module for determining whether the audio information has any abnormality based on the first target posterior probability.

[0017] According to another aspect of the present application, a transformer abnormal sound detection system is provided, the system including a memory and a processor, the memory storing a computer program executed by the processor, and the computer program, when executed by the processor, causing the processor to execute the above-mentioned transformer abnormal sound detection method.

[0018] According to another aspect of the present application, a storage medium is provided, on which a computer program is stored. When the computer program is run, the method for detecting abnormal noise of a transformer is executed.

[0019] Based on the method, device and system for detecting abnormal noise in a transformer according to the embodiments of the present application, a first posterior probability of the audio to be detected is obtained through a first network, and finally a first target posterior probability of the corresponding voltage level is obtained. The positive anomaly of the audio to be detected is judged by the first target posterior probability. This method does not require abnormal training data, does not overly rely on the annotation accuracy of the training data, is not prone to problems such as training overfitting and reliance on normal training data coverage, and subdivides equipment attributes and application scenarios such as the voltage level of the transformer, so the positive anomaly judgment in the subdivided scenarios is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0021] Figure 1 The figure shows a flow chart of transformer abnormal noise detection in the related art.

[0022] Figure 2 A flow chart of a transformer abnormal noise detection method in an embodiment of the present application is shown.

[0023] Figure 3 A schematic diagram of a process for the second network to detect the posterior probability of the second target in an embodiment of the present application is shown.

[0024] Figure 4 The figure shows a structural block diagram of a transformer abnormal noise detection device in an embodiment of the present application.

[0025] Figure 5 The structural block diagram of the transformer abnormal sound detection system in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the present application more apparent, example embodiments according to the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein. Based on the embodiments of the present application described in this application, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of this application.

[0027] like Figure 1 Figure 2 shows the training process for audio to be detected based on an AE network in related technology. AE, also known as an autoencoder, is a type of artificial neural network used in semi-supervised and unsupervised learning. Its function is to represent input information by using it as a learning target. AE can serve as a powerful feature extractor suitable for pre-training deep neural networks. Furthermore, it can randomly generate data similar to the training data, effectively expressing important information from the original data. Therefore, it is often considered a generative model.

[0028] An autoencoder consists of two parts. The first part is the encoder, which is generally a multi-layer network that compresses the input data into a vector of low dimension. This vector is called the bottleneck (Code). The second part is the decoder, which takes the bottleneck as input and outputs the data, which is called the reconstructed input data. The loss function is to minimize the distance between the reconstructed data and the original data. The encoder and decoder of an autoencoder can be built using a variety of neural network components. The most classic component is the Multi-Layer Perceptron (MLP). Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and transformers can also be used to implement it. The objective function can be described by the following formula:

[0029] Minimize||f D (f E (x))-X|| 2

[0030] Among them, f D and f E The acoustic features of the audio sample to be tested are inferred through the network to obtain the reconstructed features, and the residual size before and after the feature reconstruction is used to determine whether it is abnormal. The residual E calculation formula is as follows:

[0031]

[0032] Among them, M represents the number of frames of acoustic features, N represents the dimension of acoustic features, and X mn Represents the nth dimension of the mth frame of the input acoustic feature, X' mn Represents the nth dimension of the mth frame of the reconstructed feature output. A large residual value is considered abnormal, otherwise it is normal. Figure 1As shown in the figure, after training, the audio data with acoustic feature extraction passes through the AE network model, and the reconstructed acoustic features are obtained through reasoning. The reconstructed residual value is then calculated, and the abnormality is judged by comparing the residual value with the threshold to obtain the detection result.

[0033] The above method has problems such as over-reliance on the annotation accuracy of training data, easy overfitting of training, and reliance on normal training data coverage. Based on the above problems, this application is proposed.

[0034] like Figure 2 As shown, the present application provides a method for detecting abnormal noise in a transformer, comprising:

[0035] S21. Acquire audio information of the transformer to be tested and the voltage level of the transformer to be tested, and extract audio features based on the audio information.

[0036] S22. Input the audio feature into a trained first network, and the first network outputs multiple first posterior probabilities. The first network is a voltage level multi-classification network, and each first posterior probability is a posterior probability that the audio feature corresponds to a voltage level.

[0037] S23. Taking the voltage level of the transformer to be tested as the target voltage level, and determining a first posterior probability corresponding to the target voltage level from the multiple first posterior probabilities as the first target posterior probability.

[0038] S24. Determine whether the audio information is abnormal based on the first target posterior probability.

[0039] Based on the method for detecting abnormal noise in a transformer according to the embodiment of the present application, a first posterior probability of the audio to be detected is obtained through a first network, and finally a first target posterior probability of the corresponding voltage level is obtained. The positive anomaly of the audio to be detected is judged by the first target posterior probability. This method does not require abnormal training data and does not overly rely on the accuracy of the annotation of the training data. In addition, the device attributes and application scenarios such as the voltage level of the transformer are subdivided, so the positive anomaly judgment in the subdivided scenarios is more accurate.

[0040] Exemplarily, for the first network, it is necessary to obtain its training data. The method used to obtain the training data is: collect audio with a microphone 5 cm outside the normally operating transformer casing, and record the voltage level of the collected transformer, where the voltage level can be: 110KV, 220KV, 500KV, 800KV, 1000KV. For the collected audio, the filter bank features of the audio are extracted, because the filter bank features have more DCT (discrete pre-transform) than the Mel-frequency cepstral coefficients (MFCC), which can be considered as a loss of audio information. Exemplarily, for the audio information to be detected, when feature extraction is performed in step S11, the extracted audio features are also filter bank features.

[0041] After obtaining the training data for the first network, the first network is trained. Exemplarily, the processed training data is used to train a voltage level classification network in a supervised manner. The network structure used is a deep residual network (resnet34), and the loss function used in training is cross-entropy loss (CE Loss). After training, the audio features are input into the trained first network, which outputs multiple first posterior probabilities. The first network is a voltage level multi-classification network, and each first posterior probability is the posterior probability of the audio feature corresponding to a voltage level. Exemplarily, the voltage level classification network is a 5-classification network. The first network is divided into 5 voltage nodes based on the 5 voltage levels, and each first posterior probability is the posterior probability of the audio feature corresponding to one of the 5 voltage nodes. After the voltage nodes are divided, the first posterior probability corresponding to the target voltage level (voltage node) is determined from the multiple first posterior probabilities, thereby obtaining the first target posterior probability Score1. After obtaining the first target posterior probability Score1, the presence of an abnormality in the audio information can be determined based on the first target posterior probability. Exemplarily, Score1 is compared with a first threshold thresh1. When Score1 is greater than or equal to thresh1, it is determined that the audio information has no abnormality; otherwise, it is determined that the audio information has an abnormality.

[0042] Illustratively, the method further includes, after determining the first target posterior probability Score1, inputting the audio feature into multiple trained second networks, each second network outputting a second posterior probability to obtain multiple second posterior probabilities, and selecting a maximum value from the multiple second posterior probabilities as the second target posterior probability Score2. Each second network is a binary classification network corresponding to a voltage level in the first network, and each second posterior probability Score2 is a posterior probability that the audio feature is a normal audio feature at that voltage level.

[0043] For the second network, its training data includes the training data of the first network and interference sample data. The above content introduces the acquisition of the training data of the first network, and the training data of the first network includes normal sample data. The interference sample data is obtained by filtering the data in the open source set through the trained first network. Exemplarily, for the interference sample data, the data in the open source set is filtered by the trained first network to obtain the interference sample data, including: extracting features for each data to be filtered in the open source set, inputting the extracted features into the trained first network, and when the first posterior probability output by the first network for the extracted features is greater than the first threshold, the data to be filtered is used as the interference sample data. The first posterior probability obtained for the interference sample data is greater than the first threshold, which is a difficult interference sample that is more easily confused with normal sound.

[0044] After obtaining the interference class sample data, the voltage level data corresponding to the second network is used as the normal class training data of the second network, and the data of other voltage levels and the interference class sample data are used as the abnormal class pseudo-label training data to train the second network. The network structure used by the second network is also a deep residual network (resnet34). The positive and abnormal binary classification model is obtained by using the normal class training data and the abnormal class pseudo-label training data to form the second network. At the same time, the corresponding positive and abnormal binary classification model is obtained based on each voltage level. For example, there are 5 voltage levels. The 5 second networks corresponding to these 5 positive and abnormal binary classification models are combined to obtain a second network group.

[0045] For example, Figure 3 FIG. 1 is a schematic diagram showing detection through a second network. Figure 3 Multiple second networks are shown, including binary positive anomaly classification models corresponding to five different voltage nodes: 110 kV, 220 kV, 500 kV, 800 kV, and 1000 kV. The posterior probabilities of the normal nodes for the binary positive anomaly classification models corresponding to different voltage nodes are calculated to obtain five second posterior probabilities. The maximum of these five second posterior probabilities is then taken as the second target posterior probability Score2.

[0046] For example, when determining whether the audio information is abnormal based on the first target posterior probability Score1, the second target posterior probability Score2 also needs to be considered. That is, the determination of whether the audio information is abnormal is based on the first target posterior probability Score1 and the second target posterior probability Score1. Specifically, after obtaining the first target posterior probability Score1 and the second target posterior probability Score2, the first target posterior probability Score1 and the second target posterior probability Score2 are weighted and summed to obtain the final third target posterior probability Score3. The process of obtaining Score3 is as follows:

[0047] Score3=a*Score1+b*Score2

[0048] In the above formula, a and b are both constants, and a + b = 1. After obtaining the third target posterior probability Score3, the third target posterior probability Score3 is compared with the preset second threshold thresh2. If the third target posterior probability Score3 is greater than or equal to the second threshold thresh2, the audio information is determined to be normal; otherwise, the audio information is determined to be abnormal. After receiving feedback indicating that the audio information is abnormal, the abnormal sound of the transformer to be tested can be detected.

[0049] Based on the transformer abnormal sound detection method of the embodiment of the present application, the posterior probability of the corresponding voltage level is selected through the trained first network, indirectly realizing the judgment of positive anomalies. The first network is then used to filter the open source collection to obtain interference samples. The second network is trained based on the data and interference samples trained by the first network. The third target posterior probability Score3 is then obtained by using the first target posterior probability Score1 and the second target posterior probability Score2 obtained by the first network. The audio features are then evaluated using Score3, and Score3 is compared with a second threshold to determine whether the audio information is abnormal, thereby realizing the detection of transformer abnormal sounds. Compared with the detection methods in the related art, this solution does not require abnormal training data and subdivides the equipment properties and application scenarios of the transformer, thereby making the judgment of positive anomalies in the subdivided scenarios more accurate.

[0050] Furthermore, the transformer noise detection method based on the embodiment of this application employs a second network trained with data containing interference samples. Based on the first network's indirect positive and abnormal judgment scheme, the second network is trained using interference samples that the first network cannot identify. The two networks are then combined to form a complementary score. This combined decision-making strategy further enhances performance and improves the robustness and generalization of the overall system, resulting in greater robustness in transformer noise detection tasks.

[0051] The present application also provides a transformer abnormal sound detection device, such asFigure 4 As shown, the transformer abnormal noise detection device 40 includes an audio acquisition module 41, a detection module 42, a matching module 43, and an evaluation module 44. The transformer abnormal noise detection device 40 can be used to execute the transformer abnormal noise detection method described above. The specific operations of the transformer abnormal noise detection method have been described in detail above. For the sake of brevity, the specific details are omitted here, and only the main operations are described.

[0052] The audio acquisition module 41 is used to obtain the audio information of the transformer under test and the voltage level of the transformer under test, and extract audio features based on the audio information. The detection module 42 is used to input the audio features into a trained first network, which outputs multiple first posterior probabilities. The first network is a voltage level multi-classification network, and each first posterior probability is the posterior probability of the audio feature corresponding to a voltage level. The matching module 43 is used to use the voltage level of the transformer under test as the target voltage level and determine the first posterior probability corresponding to the target voltage level from the multiple first posterior probabilities as the first target posterior probability. The evaluation module 44 is used to determine whether the audio information is abnormal based on the first target posterior probability.

[0053] In an embodiment of the present application, the detection module 42 can also be used to: after determining the first target posterior probability, input the audio feature into multiple trained second networks, and each second network outputs a second posterior probability to obtain multiple second posterior probabilities, wherein each second network is a binary classification network corresponding to a voltage level in the first network, and each second posterior probability is the posterior probability that the audio feature is a normal audio feature at a voltage level; the matching module 43 can also be used to select the maximum value from the multiple second posterior probabilities as the second target posterior probability; the judgment module 44 can also be used to determine whether the audio information is abnormal based on the first target posterior probability and the second target posterior probability.

[0054] In an embodiment of the present application, the training data of the second network includes the training data of the first network and interference sample data, wherein the training data of the first network includes normal sample data, and the interference sample data is obtained by screening the data in the open source set through the trained first network.

[0055] In an embodiment of the present application, the second network is trained using data of a voltage level corresponding to the second network as normal class training data, and using data of other voltage levels and interference class sample data as abnormal class pseudo-label training data.

[0056] In an embodiment of the present application, data in an open source set is screened by a trained first network to obtain interference sample data, including: extracting features from each data to be screened in the open source set, inputting the extracted features into the trained first network, and when the first posterior probability output by the first network for the extracted features is greater than a first threshold, the data to be screened is used as interference sample data.

[0057] In an embodiment of the present application, the evaluation module 44 determines whether the audio information is abnormal based on the first target posterior probability and the second target posterior probability, which may include: performing a weighted summation on the first target posterior probability and the second target posterior probability to obtain a third target posterior probability; comparing the third target posterior probability with the second threshold; when the third target posterior probability is greater than or equal to the second threshold, determining that the audio information is normal, otherwise determining that the audio information is abnormal.

[0058] In an embodiment of the present application, the evaluation module 44 determines whether the audio information is abnormal based on the first target posterior probability, which may include: comparing the first target posterior probability with a first threshold; when the first target posterior probability is greater than or equal to the first threshold, determining that the audio information is normal, otherwise determining that the audio information is abnormal.

[0059] In an embodiment of the present application, the audio feature is a filter bank feature.

[0060] Based on the device for detecting abnormal sound of transformers in the embodiment of the present application, the first posterior probability of the audio to be detected is obtained through the first network, and finally the first target posterior probability of the corresponding voltage level is obtained, and the positive anomaly of the audio to be detected is judged by the first target posterior probability. This method does not require abnormal training data, does not overly rely on the accuracy of the annotation of the training data, and subdivides the equipment attributes and application scenarios such as the voltage level of the transformer, so the positive anomaly judgment in the subdivided scenario is more accurate. In addition, based on the device for detecting abnormal sound of transformers in the embodiment of the present application, a second network is used for training with data including interference samples, and an indirect positive anomaly judgment scheme is adopted based on the first network. When training the second network, interference samples that cannot be identified by the first network are used, and the scores of the two are fused to form a good complement. By adopting the above-mentioned joint judgment strategy, the effect is further improved, and the robustness and generalization of the overall system are improved, and the robustness in the task of detecting abnormal sound of transformers is better.

[0061] like Figure 5As shown, the embodiment of the present application further provides a transformer abnormal sound detection system 50, which comprises a memory 51 and a processor 52, the memory 51 stores a computer program which is run by the processor 52, and the computer program, when being run by the processor, causes the processor 52 to execute the transformer abnormal sound detection method described above. The transformer abnormal sound detection system 50 can be used to execute the transformer abnormal sound detection method described above, and the specific operations of the transformer abnormal sound detection method have been described in detail above, and thus will not be described here again for the sake of brevity.

[0062] The embodiment of the present application further provides a storage medium, which stores a computer program, and the computer program, when being run, executes the transformer abnormal sound detection method described above.

[0063] The transformer abnormal sound detection method, device, system and storage medium disclosed by the embodiment of the present application indirectly realize the judgment of positive anomaly by selecting the posterior probability corresponding to the voltage level through the first network after training. And the interference samples are obtained by screening the open source set through the first network, the second network after training is obtained according to the data trained by the first network and the interference samples, the third target posterior probability Score3 is obtained through the first target posterior probability Score1 and the second target posterior probability Score2 obtained by the first network, the audio features are judged again through Score3, Score3 and the second threshold are compared, whether the audio information is abnormal is judged, and the detection of the transformer abnormal sound is realized. Compared with the detection method in the related art, the present scheme does not need abnormal training data, and the device properties and application scenarios of the transformer are subdivided, so that the positive anomaly judgment in the subdivided scenarios is more accurate. And the problems such as over-reliance on the labeling accuracy of training data, easy overfitting of training, and dependence on the coverage of normal training data in the related art are avoided.

[0064] Although the example embodiments have been described herein with reference to the accompanying drawings, it is to be understood that the example embodiments are only exemplary and are not intended to limit the scope of the present application. Those of ordinary skill in the art can make various changes and modifications without departing from the scope and spirit of the present application. All these changes and modifications are intended to be included within the scope of the present application as claimed in the appended claims.

[0065] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, or in a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those of ordinary skill in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0066] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units described is merely a logical function division. In actual implementation, other division methods may be used, such as combining or integrating multiple units or components into another device, or ignoring or not performing some features.

[0067] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known methods, structures, and techniques are not shown in detail so as not to obscure the understanding of this description.

[0068] Similarly, it should be understood that in order to streamline the present application and aid in understanding one or more of the various inventive aspects, in the description of the exemplary embodiments of the present application, the various features of the present application are sometimes grouped together into a single embodiment, figure, or description thereof. However, this approach of the present application should not be interpreted as reflecting the intention that the application claimed for protection requires more features than those explicitly recited in each claim. More precisely, as reflected in the corresponding claims, the inventive point is that the corresponding technical problem can be solved with fewer features than all the features of a single disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into the detailed description, with each claim itself serving as a separate embodiment of the present application.

[0069] It will be understood by those skilled in the art that, except where mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or apparatus disclosed herein may be combined in any combination. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature providing the same, equivalent, or similar purpose.

[0070] Furthermore, those skilled in the art will appreciate that although some embodiments described herein include certain features included in other embodiments but not other features, combinations of features from different embodiments are intended to be within the scope of this application and to form different embodiments. For example, in the claims, any of the claimed embodiments may be used in any combination.

[0071] The various component embodiments of the present application can be implemented in hardware, or in a software module running on one or more processors, or in a combination thereof. Those skilled in the art will appreciate that a microprocessor or digital signal processor (DSP) can be used in practice to implement some or all of the functions of some modules according to the embodiments of the present application. The application can also be implemented as a part or all of a device program (e.g., a computer program and a computer program product) for performing the method described herein. Such a program implementing the present application can be stored on a computer-readable medium, or can have the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0072] It should be noted that the above embodiments illustrate rather than limit the present application, and that a person skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbols placed between brackets should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present application may be implemented by means of hardware comprising several different elements and by means of appropriately programmed computers. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0073] The above description is merely a specific embodiment or illustration of a specific embodiment of the present application, and the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. The scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for detecting abnormal noise in a transformer, characterized in that: The method comprises: Acquire audio information of the transformer to be tested and the voltage level of the transformer to be tested, and extract audio features based on the audio information; Inputting the audio feature into a trained first network, and having the first network output a plurality of first posterior probabilities, wherein the first network is a voltage level multi-classification network, and each first posterior probability is a posterior probability of the audio feature corresponding to a voltage level; Taking the voltage level of the transformer to be tested as the target voltage level, and determining a first posterior probability corresponding to the target voltage level from the multiple first posterior probabilities as a first target posterior probability; The feature is input into multiple trained second networks, each of which outputs a second posterior probability to obtain multiple second posterior probabilities, and the maximum value is selected from the multiple second posterior probabilities as the second target posterior probability; each of the second networks is a binary classification network corresponding to a voltage level in the first network, and each of the second posterior probabilities is a posterior probability that the audio feature is a normal audio feature at a voltage level; Whether the audio information is abnormal is determined based on the first target posterior probability and the second target posterior probability.

2. The method according to claim 1, characterized in that The training data of the second network includes the training data of the first network and interference sample data, wherein the training data of the first network includes normal sample data, and the interference sample data is obtained by screening the data in the open source set through the trained first network.

3. The method according to claim 2, characterized in that The second network is trained using data of the voltage level corresponding to the second network as normal class training data, and using data of other voltage levels and the interference class sample data as abnormal class pseudo-label training data.

4. The method according to claim 2, characterized in that The interference sample data is obtained by screening the data in the open source set through the trained first network, including: Extract features from each data to be screened in the open source set, input the extracted features into the trained first network, and when the first posterior probability output by the first network for the extracted features is greater than a first threshold, use the data to be screened as the interference sample data.

5. The method according to claim 4, characterized in that The determining whether the audio information is abnormal based on the first target posterior probability and the second target posterior probability includes: Performing a weighted summation on the first target posterior probability and the second target posterior probability to obtain a third target posterior probability; comparing the third target posterior probability with a second threshold; When the third target posterior probability is greater than or equal to the second threshold, it is determined that the audio information has no abnormality; otherwise, it is determined that the audio information has an abnormality.

6. The method according to claim 1, characterized in that The determining whether the audio information is abnormal based on the first target posterior probability includes: comparing the first target posterior probability with a first threshold; When the first target posterior probability is greater than or equal to the first threshold, it is determined that the audio information has no abnormality; otherwise, it is determined that the audio information has an abnormality.

7. The method according to any one of claims 1 to 6, characterized in that The audio features are filter bank features.

8. A transformer abnormal noise detection device, characterized in that: The device comprises: An audio acquisition module, configured to acquire audio information of the transformer to be tested and the voltage level of the transformer to be tested, and extract audio features based on the audio information; a detection module, configured to input the audio feature into a trained first network, and have the first network output a plurality of first posterior probabilities, wherein the first network is a voltage level multi-classification network, and each first posterior probability is a posterior probability that the audio feature corresponds to a voltage level; A matching module is configured to use the voltage level of the transformer to be tested as a target voltage level, determine a first posterior probability corresponding to the target voltage level from the multiple first posterior probabilities as a first target posterior probability; input the audio feature into multiple trained second networks, each of the second networks outputting a second posterior probability to obtain multiple second posterior probabilities, and select a maximum value from the multiple second posterior probabilities as the second target posterior probability; wherein each second network is a binary classification network corresponding to a voltage level in the first network, and each second posterior probability is a posterior probability that the audio feature is a normal audio feature at a voltage level; An evaluation module is configured to determine whether the audio information is abnormal based on the first target posterior probability and the second target posterior probability.

9. A transformer abnormal sound detection system, characterized in that: The system includes a memory and a processor, wherein a computer program executed by the processor is stored in the memory, and when the computer program is executed by the processor, the processor executes the transformer abnormal sound detection method according to any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is run, the method for detecting abnormal noise of a transformer according to any one of claims 1 to 7 is executed.

Images