Converter transformer external interference source positioning method based on fusion of ultrasonic waves and electromagnetic waves

Abstract

The invention relates to a converter transformer external interference source positioning method based on ultrasonic wave and electromagnetic wave fusion, and belongs to the field of partial discharge detection and positioning. Performing layered identification on the time-frequency structures of the two signals through a T-F fuzzy clustering method, and distinguishing internal discharge and external interference signals of the transformer; inputting the external interference signal into a dual-channel neural network for feature extraction, wherein the dual-channel neural network comprises a first channel based on CNN and a second channel based on BiLSTM; the first channel is used for extracting local frequency change characteristics of an external interference signal; the second channel is used for capturing the dependency relationship and global time sequence characteristics of the external interference signal in the time dimension; a self-attention mechanism is further introduced into the network for weighting processing; and carrying out splicing or weighted fusion on the extracted high-dimensional features through a feature fusion module, and inputting the high-dimensional features into a full-connection layer to realize three-dimensional space positioning of the partial discharge interference source.

Description

Technical Field

[0001] The present invention belongs to the technical field of partial discharge detection and positioning, and relates to a method for positioning an external interference source of a converter transformer based on the fusion of ultrasonic waves and electromagnetic waves. Background Art

[0002] Partial discharge detection mainly relies on the perception of a single physical quantity, such as the ultra-high frequency (UHF) method, the current pulse method, and the ultrasonic method, which capture discharge signals from the perspective of electromagnetic waves or mechanical waves, respectively. These methods perform well in locating discharges inside the transformer body or a closed cavity, but have low accuracy and poor noise resistance in locating external interference sources in complex open spaces. With the development of intelligent sensing and multimodal signal processing, the fusion of multi-source information for spatial positioning of discharge sources has become a research hotspot. As two highly complementary signal carriers, ultrasonic waves and electromagnetic waves have the potential for "high spatial resolution + fast response speed + strong environmental adaptability" through joint analysis. Therefore, external interference source positioning technology based on the fusion of ultrasonic and electromagnetic waves is gradually emerging.

[0003] Hui Enliang proposed an improved sparrow search algorithm based on a combined acoustic and electrical method for localizing partial discharges. This method uses the time at which an ultra-high frequency sensor receives an electromagnetic wave signal as the discharge start time, combined with the time difference information from the acoustic wave received by other ultrasonic sensors. A localization model is established based on the relationship between sound velocity and the sensor's spatial coordinates. The improved sparrow search algorithm optimizes the solution to the time difference equation, accurately locating the spatial location of the partial discharge source. This method integrates electromagnetic and acoustic wave information, improving localization accuracy and anti-interference capabilities. Zhou Dianbo et al. proposed a multi-band joint detection technology for transformer partial discharge, aiming to enhance anti-interference capabilities, pattern recognition accuracy, and localization precision. This method enhances pattern recognition by constructing a three-dimensional discharge spectrum. Using ultra-high frequency signals as a reference time scale, it improves the accuracy of ultrasonic time delay extraction, thereby enhancing localization accuracy. Furthermore, high-frequency signal calibration is used to obtain the apparent discharge amount, enabling accurate assessment of discharge severity. Liu Qiang et al. proposed a method and system for locating external interference in converter transformer partial discharge testing. By collecting electromagnetic and ultrasonic signal sets near the transformer under test and processing them separately to obtain time delay information, a spatial search algorithm is used to calculate the interference location for electromagnetic and ultrasonic positioning. The ultrasonic positioning result is then used to correct the electromagnetic wave positioning result, ultimately obtaining the precise location of the external interference. This method can quickly identify and locate random pulse interference, effectively improving the accuracy of partial discharge detection.

[0004] While a wide variety of partial discharge detection methods exist, each with its own advantages, practical applications lack effective integration between these different approaches, leading to biased detection results and difficulties in accurately identifying and locating the discharge source. Traditional pulsed current methods suffer from low signal-to-noise ratios in strong electromagnetic interference environments, limiting detection accuracy and inability to achieve spatial localization. Ultra-high frequency (UHF) detection methods offer strong anti-interference capabilities and identification accuracy, but due to the rapid propagation of electromagnetic waves, even small time differences can introduce significant errors, limiting their location accuracy. Ultrasonic methods, while offering strong localization capabilities, suffer from low sensitivity and limited detection range due to the complex propagation paths within the equipment. Existing single detection technologies struggle to balance localization accuracy, anti-interference capabilities, and signal extraction reliability, making them difficult to adapt to the actual detection needs of complex power equipment. Therefore, a combined detection method that integrates multiple physical information is urgently needed to improve the accuracy and robustness of partial discharge identification and location. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method for locating external interference sources of converter transformers based on the fusion of ultrasonic and electromagnetic waves.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A method for locating external interference sources of a converter transformer based on the fusion of ultrasonic waves and electromagnetic waves comprises the following steps:

[0008] Synchronously collect original electromagnetic wave signals and ultrasonic signals;

[0009] The time-frequency structures of the two signals are hierarchically identified by using the TF fuzzy clustering method to distinguish internal discharge from external interference signals of the transformer.

[0010] The external interference signal is input into a dual-channel neural network for feature extraction. The dual-channel neural network includes a first channel based on a convolutional neural network (CNN) and a second channel based on a bidirectional long short-term memory (BiLSTM) network. The first channel is used to extract the local frequency variation characteristics of the external interference signal. The second channel is used to capture the dependency relationship and global temporal characteristics of the external interference signal in the time dimension. A self-attention mechanism is also introduced into the first and second channels for weighted processing.

[0011] The extracted high-dimensional features are spliced ​​or weightedly fused through the feature fusion module and input into the fully connected layer to achieve three-dimensional spatial positioning of the partial discharge interference source.

[0012] Furthermore, the TF fuzzy clustering method is used to perform hierarchical identification on the time-frequency structures of the two signals to distinguish between internal discharge of the transformer and external interference signals, which specifically includes the following steps:

[0013] The collected partial discharge time series is subjected to standard normalization processing;

[0014] A partial discharge signal s i (t i ) has n sampling points, and the equivalent time of the partial discharge signal is as follows:

[0015]

[0016] In the formula is the equivalent time position of the signal;

[0017] The equivalent frequency of the partial discharge signal is as follows:

[0018]

[0019] Where f i is the frequency of the partial discharge signal, X i (f i ) is the value of the partial discharge signal frequency;

[0020] The pulse waveform-time series is used instead of the pulse peak-time series to form a TF characteristic spectrum based on the original pulse signal. Different types of discharges will be distributed at different points on the spectrum, thus realizing the classification of 0 partial discharge types. After processing a single partial discharge pulse signal according to the above two equations, a set of equivalent duration T and equivalent bandwidth F are obtained. When multiple sets of TF vectors are projected onto the two-dimensional plane formed by TF, the characteristic TF spectrum of the partial discharge pulse is formed.

[0021] The fuzzy C-means (FCM) clustering method is used to classify the feature points in the TF spectrum. The clustering analysis problem is transformed into a constrained nonlinear programming problem. The optimal partitioning of the data set is obtained through optimization. The fuzzy clustering analysis based on the objective function is described as follows:

[0022]

[0023] Where V=(ν1,ν2,…,ν c ) represents C cluster centers; ||x j -v i || is the Euclidean distance; m∈[1,∞) is the weighted index;

[0024] The FCM specific process includes:

[0025] a) Calculate or update the membership function matrix U according to the following formula b =[u ij ]:

[0026]

[0027] b) Update the cluster center according to the following formula:

[0028]

[0029] c) If ||V b -V b-1 If ||≤ε, the algorithm stops and outputs the membership function matrix U and cluster center V; otherwise, it returns to step a;

[0030] Based on the above steps, the raw data preprocessing of the partial discharge pulse waveform is completed, and the original pulse peak-time series is converted into a TF spectrum. The discharge types are distinguished by FCM, so as to make a preliminary judgment on the number of discharge types. In the clustering step, the number of cluster categories is selected based on two selection methods. One is based on the clustering of the TF spectrum. If there are obvious C types of clusters, the number of cluster categories is set to C; the other is that there is only one type of cluster in the TF spectrum, there are no obvious clusters, but the characteristic duration or characteristic bandwidth distribution is long, then the c value is tried from small to large until an obvious misclassification occurs or the number of pulses is insufficient to compare with the fingerprint library.

[0031] Furthermore, the first channel takes the external interference signals of the electromagnetic wave signal and the ultrasonic signal, as well as the preliminary positioning information obtained by positioning the two signals separately, as input data. It first performs normalization processing, then uses the CNN network to extract the local frequency change characteristics of the external interference signal, and then uses the attention mechanism to perform weighted processing. Finally, it inputs the data into the Dropout layer to prevent overfitting.

[0032] The second channel takes the external interference signals of electromagnetic wave signals and ultrasonic signals, as well as the preliminary positioning information obtained by separately positioning the two signals, as input data. First, the BiLSTM network is used to capture the dependency and global timing characteristics of the external interference signals in the time dimension. Then, normalization is performed, and then weighted processing is performed using the attention mechanism. Finally, the information is input into the Dropout layer to prevent overfitting.

[0033] Furthermore, the features extracted by the dual-channel neural network include amplitude, periodicity, arrival time, frequency characteristics and preliminary positioning results.

[0034] Furthermore, a hybrid hyperparameter search strategy is constructed to dynamically adjust the hyperparameters in the dual-channel neural network. The hybrid hyperparameter search strategy combines the Tent chaotic mapping with the whale optimization algorithm (WOA), takes the minimum positioning error of the dual-channel neural network as the objective function, performs global search and local convergence, and thus constructs a CNN-BiLSTM network with the optimal structure.

[0035] Furthermore, the Tent chaotic map is combined with the whale optimization algorithm WOA, and the calculation process includes:

[0036] (1) Prey encirclement stage, expressed as:

[0037] D=|CX * (t)-X(t)|

[0038] X(t+1)=X * (t)-A×D where X*(t) represents the current whale position closest to the prey, i.e., the optimal whale position; X(t) represents the current position of the whale body; and D represents the distance between the current individual whale and the optimal whale position. A and C are the coefficient vectors for updating the whale position, which are expressed by the following mathematical expressions:

[0039] A=2e·r1-e

[0040] C=2·r2

[0041] e=2-2t / t max

[0042] Where: e represents the convergence function; r1, r2 represent random numbers in [0, 1]; t represents the current number of iterations; tmax represents the maximum number of iterations, and the value of e decreases linearly from 2 to 0;

[0043] (2) In the process of surrounding prey, whales use spiral bubble nets to catch prey, which can be expressed as:

[0044] X(t+1)=D * ·e bl cos(2πl)+X * (t)

[0045] D * =|X * (t)-X(t)|

[0046] Where: D* represents the distance between the current whale individual and the optimal whale position; b represents a constant used to describe the shape of the logarithmic spiral; l represents a random number in [0,1];

[0047] While using a spiral bubble net to hunt, whales use a strategy of spiral swimming and encircling their prey. Assuming that the probability of whales performing both contraction encirclement and spiral swimming is 50%, the process can be expressed by the following mathematical expression:

[0048]

[0049] (3) The whale's search method is determined by adjusting the value of the coefficient vector. When the value of the coefficient vector is greater than a certain threshold, the whale is in the initial stage of foraging. At this time, they will randomly search for prey outside the encirclement and perform random searches based on the positions of other whales around them. When the value of the coefficient vector is less than the threshold, the whale enters the hunting stage and begins to perform two behaviors: prey encirclement and bubble net attack, conducting a small-scale local search. This process is expressed by the following mathematical expression:

[0050] X(t+1)=X r (t)-A×|CX r (t)-X(t)|

[0051] Where: Xr(t) represents the position of a random whale individual.

[0052] (4) Use Tent mapping to initialize the population. The mapping formula is as follows:

[0053]

[0054] Through this mapping method, the traversability of the population and the solution speed of the algorithm are optimized.

[0055] The beneficial effects of the present invention are:

[0056] Significantly improve interference suppression capabilities: This invention integrates the multimodal information of electromagnetic waves and ultrasonic signals, effectively avoiding low-frequency electromagnetic interference in the ultra-high frequency band, and at the same time utilizing the strong anti-interference characteristics of ultrasonic waves to significantly improve the signal-to-noise ratio and stability of the system in complex environments.

[0057] Improved signal recognition and positioning accuracy: A dual-channel deep neural network (CNN-BiLSTM) models different signal types and extracts in-depth features, enabling a more comprehensive capture of the time-frequency and temporal dynamics of discharge signals, enabling accurate classification and three-dimensional spatial location of partial discharge sources. Experimental results demonstrate a significant reduction in positioning error and improvement in spatial positioning accuracy compared to traditional detection methods.

[0058] Enhanced system robustness and intelligence: The introduction of attention mechanisms, normalization, and dropout structures effectively improves the model's adaptability to diverse interference environments, reduces false positives and missed detections, and improves detection reliability. Furthermore, the use of deep learning networks for end-to-end modeling of feature fusion simplifies traditional algorithmic processes and provides higher levels of intelligent processing capabilities.

[0059] Improved adaptability to actual projects: The method of the present invention does not require access to the electrical circuit, avoiding circuit interference caused by high-frequency current detection. It is particularly suitable for scenarios with high requirements on the electromagnetic environment, such as GIS, high-voltage switchgear and converter stations, and has greater practicality and engineering promotion value.

[0060] In summary, the present invention not only solves the problems of weak anti-interference ability and poor positioning accuracy in the existing technology, but also realizes efficient detection and precise positioning of partial discharge interference sources, which has important application significance in ensuring the safe operation of high-voltage direct current transmission systems.

[0061] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0063] Figure 1 This is a flow chart of a method for locating external interference sources of converter transformers based on the fusion of ultrasonic and electromagnetic waves;

[0064] Figure 2 Schematic diagram of the converter transformer external interference identification and positioning model based on IWOA-CNN-LSTM;

[0065] Figure 3 This is the basic framework diagram of the converter transformer external interference source location method based on the fusion of ultrasonic and electromagnetic waves;

[0066] Figure 4 (a) is the processing flow chart of channel 1, and (b) is the processing flow chart of channel 2;

[0067] Figure 5 Schematic diagram of convolutional neural network;

[0068] Figure 6 is the convolution graph of the pooling layer;

[0069] Figure 7 It is a fully connected calculation process diagram;

[0070] Figure 8 Figure 2 is the Dropout structure diagram, where (a) is before adding Dropout, and (b) is after adding Dropout;

[0071] Figure 9 This is the basic structure diagram of RNN;

[0072] Figure 10 This is the long short-term memory network structure diagram;

[0073] Figure 11 Schematic diagram of BiLSTM structure;

[0074] Figure 12 This is the structural diagram of the attention mechanism. DETAILED DESCRIPTION

[0075] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0076] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0077] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the embodiments of the present invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring the embodiments of the present invention.

[0078] Example 1:

[0079] like Figure 1-3 As shown, the present invention provides a method for locating external interference sources of converter transformers based on the fusion of ultrasonic waves and electromagnetic waves. By synchronously collecting electromagnetic wave and ultrasonic signals, the complementarity of the two in the feature dimension is utilized to adopt the TF fuzzy clustering algorithm to identify interference signals, and a dual-channel neural network structure is constructed for feature extraction: channel one uses a convolutional neural network (CNN) to extract its time-frequency features, and channel two uses a bidirectional long short-term memory network (BiLSTM) to extract its time series features. In order to enhance the robustness and expression ability of the model, normalization, attention mechanism and Dropout layer are introduced into both channels. Finally, the high-dimensional features are spliced ​​or weighted fused through the feature fusion module and input into the fully connected layer to realize the three-dimensional spatial positioning of the partial discharge interference source. This method effectively improves the detection accuracy and positioning reliability, and provides strong support for converter transformer status monitoring and fault warning.

[0080] During the data input phase, the original electromagnetic and ultrasonic signals are input. The time-frequency structure of the signals is hierarchically identified using the TF fuzzy clustering method, effectively distinguishing internal transformer discharge from external interference signals. Preliminary localization of the external interference signals is then performed based on the electromagnetic and ultrasonic signals, respectively. The external interference signals and the preliminary localization results of both signals are then used as input data. A hybrid architecture consisting of a convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) network is introduced to extract external interference signal features. The CNN subnetwork extracts local frequency variation features of the input data, enhancing the model's ability to recognize signal edge changes and local spectral line features. The BiLSTM is used to capture the temporal dependencies and global temporal characteristics of the interference signal. Furthermore, to improve the model's ability to determine the importance of different features, a self-attention mechanism is introduced for weighted processing, enhancing the representation of key features. The extracted features primarily include core parameters such as amplitude, periodicity, arrival time, frequency characteristics, and preliminary localization results, forming the data foundation for subsequent modeling.

[0081] The present invention uses the equivalent time-frequency method to convert the peak-time series into the TF characteristic spectrum.First, the collected partial discharge time series is subjected to standard normalization processing.

[0082] Assume a partial discharge signal s i (t i ) has n sampling points, and the equivalent time of the partial discharge signal is shown in formula (1):

[0083]

[0084] In the formula is the equivalent time position of the signal.

[0085] The equivalent frequency of the partial discharge signal is shown in formula (2):

[0086]

[0087] Where f i is the frequency of the partial discharge signal, X i (f i ) is the numerical value of the partial discharge signal frequency.

[0088] The pulse waveform-time series is used instead of the pulse peak-time series to form a TF characteristic spectrum based on the original pulse signal. Different types of discharges will be distributed at different points on the spectrum. This feature can be used to classify the type of partial discharge. After processing a single partial discharge pulse signal according to equations (1) and (2), a set of equivalent duration T and equivalent bandwidth F are obtained. When multiple sets of TF vectors are projected onto the two-dimensional plane formed by TF, the characteristic TF spectrum of the partial discharge pulse can be formed.

[0089] After preprocessing the PD signal using the equivalent time-frequency method, different types of PD pulses are distributed in different regions of the spectrum after TF conversion. This means that TF conversion can cluster pulses with the same characteristics. Classifying these different regions is a key step in distinguishing multiple PD signals. The present invention uses the fuzzy C-means clustering method (FCM) to classify the characteristic points in the TF spectrum. FCM is currently the most sensitive and widely used clustering analysis algorithm. The principle of this algorithm is to transform the clustering analysis problem into a constrained nonlinear programming problem, and to obtain the optimal partitioning of the data set through optimization. Fuzzy clustering analysis based on the objective function can be generally described as follows:

[0090]

[0091] Where V=(ν1,ν2,…,ν c ) represents C cluster centers; ||x j -v i || is the Euclidean distance; m∈[1,∞) is the weighted exponent, also known as the smoothing coefficient.

[0092] Based on the above analysis, the specific process of FCM is as follows.

[0093] Initialization: Give the number of cluster categories c, 2≤c≤n is the total number of samples c; initialize the cluster center V0; set the size of the iteration stop threshold ε.

[0094] Step 1: Calculate or update the membership function matrix U according to formula (4) b =[u ij ].

[0095]

[0096] Step 2: Update the cluster center according to formula (5)

[0097]

[0098] Step 3: If ||V b -V b-1 If ||≤ε, the algorithm stops and outputs the membership function matrix U and cluster center V; otherwise, continue to step 1.

[0099] Based on the above processing, the raw data of the partial discharge pulse waveform has been preprocessed. The raw pulse peak-time series is converted to a TF spectrum, and the discharge types are distinguished through FCM, allowing a preliminary assessment of the number of discharge types. The number of cluster categories in the clustering step can generally be selected based on two methods: first, based on the clustering of the TF spectrum, if there are obvious C types of clusters, the number of cluster categories can be set to C. Second, if there is only one type of cluster in the TF spectrum, but no obvious clusters, but the characteristic duration or characteristic bandwidth distribution is long, the c value can be tried from small to large until a clear misclassification occurs or the number of pulses is insufficient for comparison with the fingerprint library.

[0100] like Figure 4 As shown in the figure, channel 1 uses the external interference signal of the electromagnetic wave as input data. It first performs normalization processing, then uses the CNN network to extract the local frequency variation characteristics of the external interference signal. Then, it uses the attention mechanism to perform weighted processing, and finally inputs it into the Dropout layer to prevent overfitting. Channel 2 uses the external interference signal of the ultrasonic signal as input data. First, it uses the BiLSTM network to capture the time dependency and global temporal characteristics of the external interference signal. It then performs normalization processing, then uses the attention mechanism to perform weighted processing, and finally inputs it into the Dropout layer to prevent overfitting.

[0101] Convolutional Neural Networks (CNNs), one of the core models in deep learning, have demonstrated remarkable performance in tasks such as image recognition and speech processing in recent years, becoming a research hotspot in the field of artificial intelligence. The design concept of CNNs is inspired by biomimetic research on the local perception mechanisms of biological nervous systems. Its core component, artificial neurons, extract features in both spatial and temporal dimensions by responding to local regions of input data. Depending on the dimensionality of the input data, convolution operations can be categorized into various forms, such as one-dimensional and two-dimensional convolutions, providing flexible processing solutions for feature extraction from different data types.

[0102] In the feature extraction process, the convolution kernel (filter) is the core parameter of CNN, and the choice of its size and number has a significant impact on the performance of the model. Users can design the convolution kernel parameters in a targeted manner according to specific task requirements, so as to extract feature information of different levels and dimensions from the original input data. Multiple convolution kernels act in parallel on the same area of ​​the input data, and multiple sets of feature maps are generated through convolution operations to constitute the output results of the convolution layer. Through the weight sharing mechanism, CNN greatly reduces the number of model parameters without significantly increasing the complexity of the model, thereby effectively improving the training efficiency and generalization ability. In addition, the downsampling (pooling) operation plays an important role in the CNN architecture. By compressing the spatial dimension of the feature map, the pooling operation not only enhances the robustness of the model to the translation transformation of the input data, but also reduces the computational complexity, alleviates the overfitting problem, and improves the overall performance of the model. Typical CNN architectures are as follows: Figure 5 As shown in the figure, it usually contains three types of core layers: the convolutional layer is responsible for efficient extraction of local features; the pooling layer improves the generalization ability of the model through feature compression; and the fully connected layer completes the integration of high-level features at the end of the network and realizes classification or regression tasks.

[0103] (1) Convolutional layer

[0104] In a convolutional neural network (CNN), the convolution layer, as a core component, primarily performs feature extraction from raw data. Through convolution operations, the network effectively identifies key features in the input data, enhancing its ability to express salient information while also achieving a certain degree of feature dimensionality compression. The convolution kernel (also known as a filter) performs local perception calculations on the input feature map using a sliding window approach based on preset parameters, ultimately outputting a corresponding feature map.

[0105] When processing time series data, CNNs typically use a one-dimensional convolutional structure. This has the significant advantage of being able to increase or decrease feature dimensionality by adjusting the number of channels without changing the overall structure of the feature matrix. Furthermore, while maintaining the receptive field, one-dimensional convolution can increase network depth and enhance the model's nonlinear expression capabilities, thereby improving the accuracy of sequence prediction or classification. The mathematical expression for one-dimensional convolution is as follows:

[0106]

[0107] Where W k represents the weight parameter of the kth convolution kernel, X t-k+1 is the data value at the corresponding position of the input time series.

[0108] (2) Pooling layer

[0109] The pooling layer is primarily used to further compress and filter the features extracted by the convolutional layer. Its core purpose is to retain key information while reducing feature dimensionality and reducing the computational complexity of subsequent networks. Through pooling, the model can effectively remove locally redundant information, thereby improving overall computational efficiency and significantly reducing the number of network parameters, mitigating the risk of overfitting.

[0110] Common pooling methods mainly include maximum pooling and average pooling. Among them, maximum pooling selects the maximum value in a specified local receptive area as the output to enhance significant features; while average pooling takes the average of all values ​​in the area to smooth the feature distribution. The two pooling strategies can be flexibly selected according to specific task requirements to achieve the best effect. The specific diagram of the pooling operation is as follows Figure 6 shown.

[0111] (3) Fully connected layer

[0112] In a convolutional neural network, the fully connected layer is located at the end of the model and is mainly responsible for integrating and mapping the high-dimensional features after convolution and pooling in the previous stage to complete specific classification or regression tasks. Depending on the specific prediction target, appropriate activation functions or output functions are usually introduced into the fully connected layer to achieve the desired task response. The calculation formula of the fully connected layer is shown below, and the whole calculation process is shown in Figure 7 .

[0113] y=linear(k·x+b) (7)

[0114] Where: k represents the weight matrix; x represents the input vector; b represents the bias vector; linear represents the activation function.

[0115] In order to enhance the generalization ability of the model and prevent overfitting, Dropout technology is often introduced in the fully connected layer. Figure 8 (a) shows the structure before and (b) after adding Dropout. This technique randomly blocks the activation outputs of some neurons during training, forcing the network to learn feature representations in a more redundant and decentralized manner. Dropout not only improves the model's robustness to unknown data but also optimizes computational efficiency and accelerates training convergence.

[0116] Recurrent Neural Network (RNN) is a type of neural network structure that has strong modeling capabilities in processing time series data. Its key features are memory and parameter sharing mechanisms: the network can retain the information extracted at the previous moment in the time dimension and pass it as context to the input of the next moment. RNN is classified as a deep network model, not only because it has a hierarchical structure, but also because its expansion in the time dimension can be regarded as a special deep structure, in which each time step can be equivalent to a network layer. Figure 9 As shown in the figure, the symbol U represents the weight matrix between the input layer and the hidden layer at the current moment, W represents the transfer weight of the hidden state between adjacent time steps, and V is the weight matrix from the hidden layer to the output layer.

[0117] The LSTM model retains memory units based on the RNN model, and introduces a gating mechanism to filter out the data that needs to be retained and the data that needs to be deleted, extracting long-term dependency-related features from variable-length data, effectively solving the problems existing in recurrent neural networks. Figure 10 shown.

[0118] The LSTM model consists of the forget gate F t , input gate I t and output gate O t The three gates interact to allow each neuron to selectively retain, forget, input, and output information, thereby capturing long-term dependencies and effectively solving the gradient vanishing and gradient exploding problems in recurrent neural networks. The functions of each gate are as follows:

[0119] (1) Forget Gate

[0120] The main function of the forget gate is to filter the state C of the previous moment t-1 The information that needs to be forgotten. The output H of the hidden layer at the previous moment t-1 and the current input X t At the same time, enter the activation function σ to get F t , through σ to F t Control it between 0 and 1, judge the probability of needing to retain information, and compare the probability value obtained with the state C at the previous moment t-1 Multiply them together to achieve information forgetting. The calculation formula is as follows:

[0121] F i =σ(W f ×X t +U f ×H t-1 +b f ) (8)

[0122] Where W f and U f is the weight matrix, b f is the bias, and σ is the Sigmoid activation function.

[0123] (2) Input gate

[0124] Input Gate I t The main function is to state the unit C tx X needs to be retained t The amount of information. The value of the hidden layer at the previous moment H t-1 With the current input X t The output I is obtained by the σ activation function and the tanh activation function respectively. t and C tx , then multiply the two and add the information in the forget gate to get the current state unit C t , the calculation formula is as follows:

[0125] C tx =tanh(W c ×X t +U c ×H t-1 +b c ) (9)

[0126] I t =σ(W i ×X t +U i ×H t-1 +b i ) (10)

[0127] C t =F t ×C t-1 +I t ×C tx (11)

[0128] Where W c 、U c 、W i 、U i is the weight matrix, b c 、b i is the bias, σ is the Sigmoid activation function, and tanh is the activation function.

[0129] (3) Output gate

[0130] Output gate O t The main function is to control the value that needs to be output at the current moment. The calculation formula is as follows:

[0131] O t=σ(W o ×X t +U o ×H t-1 +b o ) (12)

[0132] H t =O t ×tanh(C t ) (13)

[0133] Where W o and U o is the weight matrix, b o is the offset.

[0134] In traditional long short-term memory networks (LSTM), information transmission is limited to the forward direction of the time series, that is, from historical time steps to future time steps. This one-way propagation method easily leads to the loss of some input information. To solve this problem, the bidirectional long short-term memory network (BiLSTM) introduces two information flow structures in opposite directions: a forward LSTM and a backward LSTM. This structure can capture features from both the past and future directions of the sequence, which helps to more comprehensively explore the potential correlations in time series data, improve the ability to learn time-dependent characteristics, and thus enhance the model's data utilization efficiency and prediction accuracy. The structure of BiLSTM is as follows: Figure 11 shown.

[0135] from Figure 11 It can be seen that the BiLSTM model inputs the same input sequence into the forward hidden layer h ft With the backward hidden layer h bt After being processed by their respective LSTM units, the hidden states of the two directions are connected and passed to the output layer to finally generate the output h of the BiLSTM model. t The corresponding mathematical expression is shown below.

[0136] h R =f(w1x t +w2h R-1 +b R ) (14)

[0137] h D =f(w3x t +w5h D-1 +b D ) (5)

[0138] h t =g(w4h R +w6h D +b out ) (16)

[0139] Where w1 represents the weight from the input layer to the forward LSTM; w2 is the connection weight between the units of the forward LSTM; w3 and w5 correspond to the weight from the input layer to the backward LSTM and its internal connection weights respectively; w4 and w6 represent the weights from the forward and backward LSTM outputs to the final prediction layer respectively; the bias term b ft 、b bt 、b ot They are used for bias adjustment of the forward, backward and output layers respectively.

[0140] The attention mechanism is a structured model that can dynamically assign attention weights when processing sequence data. It is usually composed of three parts: input representation, weight calculation, and context vector generation. First, the input part converts the original sequence (such as text, time series, etc.) into the corresponding hidden vector representation through the encoder. Then, in the weight calculation stage, the degree of association of each input element with the current query vector is calculated by combining trainable parameters with the activation function to reflect its relative importance in the current context. Finally, by multiplying each input element by the corresponding attention weight and summing them, a context vector is generated, thereby completing the aggregation of key information in the sequence. The basic structure of the attention mechanism is as follows Figure 12 shown.

[0141] The core calculation process of this mechanism includes the following three steps:

[0142] (1) Attention score calculation: For a given query vector q and input vector x, the degree of matching between the two is calculated using a similarity function. Common scoring functions include dot product, additive function, etc., which are uniformly expressed as:

[0143] s(x,q)=x T Wq (17)

[0144] Where: s represents the attention score, and W is the trainable weight parameter.

[0145] (2) Score normalization: All score results are normalized through the Softmax function and converted into probability distribution to obtain the attention weight: expressed as:

[0146] α i =softmax(s(x,q)) (18)

[0147] Where: α i represents the attention weight of the input i-th element on the query vector.

[0148] (3) Weighted sum output: using the calculated attention weight α i , perform weighted summation on the input values ​​and finally generate the aggregated output vector:

[0149]

[0150] Where: att(X,q) represents the output sequence after weighted averaging.

[0151] The Whale Optimization Algorithm (WOA) is a meta-heuristic optimization algorithm inspired by the unique hunting strategy of humpback whales. When hunting, humpback whales form a circle around their prey, gradually shrink the circle in a spiral shape, and finally capture the prey with a bubble net. The WOA algorithm simulates this predation process and aims to find and optimize the solution to the problem through a similar mechanism. During the algorithm's operation, the individual closest to the "prey" (that is, the optimal solution) in the current population is first determined and used as the center of the encirclement. Then, other "whales" (that is, other points in the solution space) will update their positions based on this center. Through repeated iterations and updates, the algorithm gradually approaches the optimal solution and eventually finds the optimal solution to the problem. The process mainly includes the following three operation steps:

[0152] (1) Prey encirclement stage

[0153] During the prey encirclement phase, once a group of whales discovers prey, they identify the individual closest to the prey and consider its position as the optimal encirclement location. Next, the entire group of whales begins adjusting their positions, gradually moving closer to this optimal position to encircle the prey. This process can be expressed mathematically as follows:

[0154] D=|CX * (t)-X(t)| (20)

[0155] X(t+1)=X * (t)-A×D (21)

[0156] Where: X*(t) represents the current position of the whale closest to the prey, i.e., the optimal whale position; X(t) represents the current position of the whale body; and D represents the distance between the current individual whale and the optimal whale position. A and C are the coefficient vectors for updating the whale position, expressed as follows:

[0157] A=2e·r1-e (22)

[0158] C=2·r2 (23)

[0159] e=2-2t / t max (twenty four)

[0160] Where: e represents the convergence function; r1, r2 represent random numbers in [0, 1]; t represents the current number of iterations; tmax represents the maximum number of iterations, and the value of e decreases linearly from 2 to 0.

[0161] (2) Spiral bubble net predation

[0162] To surround their prey, whales employ a unique hunting strategy: they generate bubbles and gradually shrink the encirclement in an ascending spiral. The position of this spiral is updated precisely based on the spiral's width and length, as well as the whale's current position. This spiral update strategy allows whales to more accurately approach and capture their prey. This process can be expressed as follows:

[0163] X(t+1)=D * ·e bl cos(2πl)+X * (t) (25)

[0164] D * =|X * (t)-X(t)| (26)

[0165] Where: D* represents the distance between the current whale individual and the optimal whale position; b represents a constant used to describe the shape of the logarithmic spiral; l represents a random number in [0,1].

[0166] When whales use spiraling bubble nets to hunt, they employ both spiraling and encircling strategies. To more accurately simulate these two parallel behaviors, we assume that the probability of whales performing both contraction and encirclement is 50% each. This process can be expressed mathematically as follows:

[0167]

[0168] (3) Random search for prey

[0169] To enhance the algorithm's global search capabilities and enable it to find optimal solutions over a wider range, the Whale Optimization Algorithm primarily determines the whale's search method by adjusting the value of the coefficient vector. When the coefficient vector is greater than a certain threshold, the whale is in the initial foraging phase, randomly searching for prey outside the encirclement, performing random searches based on the positions of other whales. When the coefficient vector is less than the threshold, the whale enters the hunting phase, performing two actions: encircling the prey and attacking with a bubble net, conducting a small, localized search. This process can be expressed mathematically as follows:

[0170] X(t+1)=X r (t)-A×|CX r (t)-X(t)| (28)

[0171] Where: Xr(t) represents the position of a random whale individual.

[0172] Tent chaotic map initialization population:

[0173] In the original WOA algorithm, the population is usually initialized in a random manner, which may lead to a decrease in population diversity and make the algorithm easily fall into a local optimal solution. In order to improve the global search capability of the algorithm and accelerate the convergence process, it is particularly important to use a high-quality initialization population. Due to its randomness, ergodicity, and initial value sensitivity, the chaotic map performs well in the optimization algorithm, can effectively avoid local optimality, and help the algorithm accelerate convergence. To this end, this study uses the Tent map for population initialization, and its mapping formula is as follows:

[0174]

[0175] Through this mapping method, the traversability of the population and the solution speed of the algorithm are optimized, further enhancing the search efficiency.

[0176] To further improve model performance and achieve structural generalization, this paper combines Tent chaotic mapping with the Whale Optimization Algorithm (WOA) to construct a hybrid hyperparameter search strategy, dynamically adjusting the hyperparameters in the CNN-LSTM model structure. Specific optimizations include convolution kernel parameters, activation function selection, and network depth. Minimizing the positioning error of the CNN-LSTM model is the objective function, and a combined Tent-WOA algorithm performs global search and local convergence to construct an optimal CNN-LSTM network.

[0177] The final constructed IWOA-CNN-LSTM model is iteratively trained by inputting a large amount of labeled data during the training phase. After the model converges stably, it is applied to the test sample set to complete the task of locating the interference source in three-dimensional space and output the three-dimensional spatial coordinate value (x, y, z). The predicted results are compared with the actual interference location and the error indicators such as MAE, RMSE, R 2 and three-dimensional scatter plots to analyze and verify the model positioning performance.

[0178] Compared to traditional combined electro-acoustic location methods (such as ultrasonic location based on pulsed current signals), the combined ultrasonic and ultra-high frequency detection method requires no electrical circuit connection, thus avoiding the impact of circuit interference and strong background noise introduced by high-frequency current sensors on detection results. This method is particularly suitable for equipment with strict electromagnetic environment requirements, such as GIS and high-voltage switchgear. Furthermore, this method avoids interference signals below 300 MHz in substations, effectively enhancing signal extraction capabilities and improving the accuracy and robustness of partial discharge source identification.

[0179] Specifically, combined ultrasonic and ultra-high frequency signal detection and positioning analyzes information such as signal arrival time, intensity changes, time differences, and preliminary positioning results. This allows for the investigation and precise location of partial discharge points over a wide range, offering comprehensive advantages such as high sensitivity, rapid positioning speed, and strong anti-interference capabilities. Based on this, this paper proposes a partial discharge source location method that integrates deep learning. This method utilizes a deep learning model to effectively learn and identify multi-source signal characteristics, optimizes model parameters to improve positioning accuracy, and verifies the effectiveness and reliability of this method in complex interference environments. This method provides a theoretical foundation and engineering support for intelligent diagnosis of partial discharge faults and high-precision location of interference sources.

[0180] Example 2:

[0181] An electronic device comprising a memory and a processor;

[0182] The memory is used to store computer programs;

[0183] The processor is configured to implement the method described in Example 1 when executing the computer program.

[0184] Example 3:

[0185] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described in Example 1 is implemented.

[0186] Example 4:

[0187] A computer program product includes a computer program, which implements the method described in embodiment 1 when executed by a processor.

[0188] In the above embodiments, references to "this embodiment" in the specification indicate that a particular feature, structure, or characteristic described in conjunction with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple occurrences of "this embodiment" do not necessarily refer to the same embodiment.

[0189] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the present invention are intended to encompass all such alternatives, modifications, and variations that fall within the broad scope of the appended claims.

[0190] Regarding the computer-readable storage medium in this embodiment, those skilled in the art will appreciate that all or part of the steps in the aforementioned method embodiments can be implemented using hardware associated with the computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps in the aforementioned method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0191] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication with each other. The memory is used to store computer programs, the communication interface is used for communication, and the processor and the transceiver are used to run computer programs so that the electronic terminal executes the various steps of the above method.

[0192] In this embodiment, the memory may include a random access memory (RAM), and may also include a non-volatile memory (non-volatile memory), such as at least one disk storage.

[0193] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0194] The present invention can be used in a wide variety of general-purpose or special-purpose computing system environments or configurations, such as personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments that include any of the above.

[0195] The present invention may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for locating external interference sources of converter transformers based on the fusion of ultrasonic and electromagnetic waves, characterized by: The following steps are involved: Synchronously collect original electromagnetic wave signals and ultrasonic signals; The time-frequency structures of the two signals are hierarchically identified by using the TF fuzzy clustering method to distinguish internal discharge from external interference signals of the transformer. The external interference signal is input into a dual-channel neural network for feature extraction, wherein the dual-channel neural network includes a first channel based on a convolutional neural network (CNN) and a second channel based on a bidirectional long short-term memory (BiLSTM) network; the first channel is used to extract local frequency variation features of the external interference signal; The second channel is used to capture the dependency of the external interference signal in the time dimension and the global temporal characteristics; the first channel and the first channel also introduce a self-attention mechanism for weighted processing; The extracted high-dimensional features are spliced ​​or weightedly fused through the feature fusion module and input into the fully connected layer to achieve three-dimensional spatial positioning of the partial discharge interference source.

2. The method for locating external interference sources of converter transformers based on the fusion of ultrasonic waves and electromagnetic waves according to claim 1 is characterized in that: The method of hierarchically identifying the time-frequency structures of the two signals by using the TF fuzzy clustering method to distinguish internal discharge of the transformer from external interference signals specifically includes the following steps: The collected partial discharge time series is subjected to standard normalization processing; A partial discharge signal s i (t i ) has n sampling points, and the equivalent time of the partial discharge signal is as follows: In the formula is the equivalent time position of the signal; The equivalent frequency of the partial discharge signal is as follows: Where f i is the frequency of the partial discharge signal, X i (f i ) is the value of the partial discharge signal frequency; The pulse waveform-time series is used instead of the pulse peak-time series to form a TF characteristic spectrum based on the original pulse signal. Different types of discharges will be distributed at different points on the spectrum, thus realizing the classification of 0 partial discharge types. After processing a single partial discharge pulse signal according to the above two equations, a set of equivalent duration T and equivalent bandwidth F are obtained. When multiple sets of TF vectors are projected onto the two-dimensional plane formed by TF, the characteristic TF spectrum of the partial discharge pulse is formed. The fuzzy C-means clustering method FCM is used to classify the feature points in the TF spectrum. The clustering analysis problem is transformed into a constrained nonlinear programming problem. The optimal partitioning of the data set is obtained through optimization. The fuzzy clustering analysis based on the objective function is described as follows: Where V=(ν1,ν2,…,ν c ) represents C cluster centers; ||x j -v i || is the Euclidean distance; m∈[1,∞) is the weighted index; The FCM specific process includes: a) Calculate or update the membership function matrix U according to the following formula b =[u ij ]: b) Update the cluster center according to the following formula: c) If ||V b -V b-1 If ||≤ε, the algorithm stops and outputs the membership function matrix U and cluster center V; otherwise, it returns to step a; Based on the above steps, the raw data preprocessing of the partial discharge pulse waveform is completed, and the original pulse peak-time series is converted into a TF spectrum. The discharge types are distinguished by FCM, so as to make a preliminary judgment on the number of discharge types. In the clustering step, the number of cluster categories is selected based on two selection methods. One is based on the clustering of the TF spectrum. If there are obvious C types of clusters, the number of cluster categories is set to C; the other is that there is only one type of cluster in the TF spectrum, there are no obvious clusters, but the characteristic duration or characteristic bandwidth distribution is long, then the c value is tried from small to large until an obvious misclassification occurs or the number of pulses is insufficient to compare with the fingerprint library.

3. The method for locating external interference sources of converter transformers based on the fusion of ultrasonic waves and electromagnetic waves according to claim 1 is characterized in that: The first channel takes the external interference signals of electromagnetic wave signals and ultrasonic signals, as well as the preliminary positioning information obtained by locating the two signals separately, as input data. It first performs normalization processing, then uses the CNN network to extract the local frequency change characteristics of the external interference signals, and then uses the attention mechanism to perform weighted processing. Finally, it inputs the data into the Dropout layer to prevent overfitting. The second channel takes the external interference signals of electromagnetic wave signals and ultrasonic signals, as well as the preliminary positioning information obtained by separately positioning the two signals, as input data. First, the BiLSTM network is used to capture the dependency and global timing characteristics of the external interference signals in the time dimension. Then, normalization is performed, and then weighted processing is performed using the attention mechanism. Finally, the information is input into the Dropout layer to prevent overfitting.

4. The method for locating external interference sources of converter transformers based on the fusion of ultrasonic waves and electromagnetic waves according to claim 1 is characterized in that: The features extracted by the dual-channel neural network include amplitude, periodicity, arrival time, frequency characteristics and preliminary positioning results.

5. The method for locating external interference sources of converter transformers based on the fusion of ultrasonic waves and electromagnetic waves according to claim 1 is characterized in that: A hybrid hyperparameter search strategy is constructed to dynamically adjust the hyperparameters in a dual-channel neural network. The hybrid hyperparameter search strategy combines the Tent chaotic map with the whale optimization algorithm (WOA), takes the minimum positioning error of the dual-channel neural network as the objective function, performs global search and local convergence, and thus constructs a CNN-BiLSTM network with the optimal structure.

6. The method for locating external interference sources of converter transformers based on the fusion of ultrasonic waves and electromagnetic waves according to claim 1, characterized in that: The Tent chaotic map is combined with the whale optimization algorithm WOA. The calculation process includes: (1) Prey encirclement stage, expressed as: D=|CX * (t)-X(t)| X(t+1)=X * (t)-A×D where X*(t) represents the current whale position closest to the prey, i.e., the optimal whale position; X(t) represents the current position of the whale body; and D represents the distance between the current individual whale and the optimal whale position. A and C are the coefficient vectors for updating the whale position, which are expressed by the following mathematical expressions: A=2e·r1-e C＝2·r2 e=2-2t / t max Where: e represents the convergence function; r1, r2 represent random numbers in [0, 1]; t represents the current number of iterations; tmax represents the maximum number of iterations, and the value of e decreases linearly from 2 to 0; (2) In the process of surrounding prey, whales use spiral bubble nets to catch prey, which can be expressed as: X(t+1)=D * ·e bl cos(2πl)+X * (t) D * =|X * (t)-X(t)| Where: D* represents the distance between the current whale individual and the optimal whale position; b represents a constant used to describe the shape of the logarithmic spiral; l represents a random number in [0,1]; While using a spiral bubble net to hunt, whales use a strategy of spiral swimming and encircling their prey. Assuming that the probability of whales performing both contraction encirclement and spiral swimming is 50%, the process can be expressed by the following mathematical expression: (3) The whale's search method is determined by adjusting the value of the coefficient vector. When the value of the coefficient vector is greater than a certain threshold, the whale is in the initial stage of foraging. At this time, they will randomly search for prey outside the encirclement and perform random searches based on the positions of other whales around them. When the value of the coefficient vector is less than the threshold, the whale enters the hunting stage and begins to perform two behaviors: prey encirclement and bubble net attack, conducting a small-scale local search. This process is expressed by the following mathematical expression: X(t+1)=X r (t)-A×|CX r (t)-X(t)| Where: Xr(t) represents the position of a random whale individual; (4) Use Tent mapping to initialize the population. The mapping formula is as follows: Through this mapping method, the traversability of the population and the solution speed of the algorithm are optimized.

7. An electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is configured to implement the method for locating external interference sources of a converter transformer based on the fusion of ultrasonic waves and electromagnetic waves as described in any one of claims 1 to 6 when executing the computer program.

8. A computer-readable storage medium, characterized in that The storage medium stores a computer program. When the computer program is executed by the processor, the method for locating an external interference source of a converter transformer based on the fusion of ultrasonic waves and electromagnetic waves as described in any one of claims 1 to 6 is implemented.

9. A computer program product, characterized in that: The method comprises a computer program, which, when executed by a processor, implements the method for locating an external interference source of a converter transformer based on the fusion of ultrasonic waves and electromagnetic waves as described in any one of claims 1 to 6.

Images