Power distribution network ground fault feature identification method and computer storage medium

Abstract

The invention relates to a power distribution network grounding fault feature identification method and a computer storage medium, and relates to a tree line discharge fault deep abstract feature mining method based on combination of deep learning and transfer learning. On one hand, the transformer fault features are extracted through the pre-training process of the source task stage, so that various and stable transformer fault data and relatively stable transformer fault discharge signals can be fully utilized, model training is simpler and more efficient, and the training efficiency is improved; more importantly, due to the data similarity and the inducement relevance of the two, the extracted transformer fault features can be highly matched with the tree line discharge feature form and the fault time, and the matching degree is high; and on the other hand, through the post-training process of the target task stage, the special features of the tree line discharge fault are extracted, migration data are specialized, the uniqueness of the target task is reflected, and the model accuracy can be further improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology for power systems, and in particular to a method for identifying grounding fault characteristics in distribution networks and a computer storage medium. Background Technology

[0002] With the increasing complexity of power system structures and the continuous expansion of distribution network coverage, grounding faults such as tree-line discharge have become one of the significant hidden dangers affecting the safe operation of distribution systems. Tree-line discharge refers to the phenomenon where corona discharge, flashover, or even short circuits occur due to conductors being close to trees or other debris for a long time, under the influence of external forces such as wind and rain. These faults are characterized by their suddenness, concealment, and strong nonlinearity, making them difficult to detect or accurately diagnose in advance using traditional methods.

[0003] Traditional tree-line discharge fault diagnosis methods mainly rely on manual inspection and simple feature engineering methods. These methods have the following problems: First, manual inspection is time-consuming and inefficient, and it is easy to miss intermittent fault signals; second, shallow feature extraction methods (such as mean, variance, peak coefficient, etc.) are difficult to capture deep-seated change patterns in fault signals, especially when dealing with low-amplitude, high-noise transient signals.

[0004] In recent years, with the rapid development of artificial intelligence technology, it has shown great potential in feature extraction and fault identification. Compared with traditional methods, deep learning models can automatically learn multi-level, abstract representations from large amounts of raw data, thereby improving the ability to model complex nonlinear features. However, tree-line discharge fault data is scarce and unstable, making model training difficult. How to improve this is a pressing technical problem to be solved in this field. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for identifying grounding fault characteristics in a distribution network, comprising:

[0006] Collect transformer fault signals and construct a source task training set;

[0007] Collect treeline discharge fault signals and construct a training set for the target task;

[0008] The process of constructing and training a tree-line discharge fault feature recognition model includes: Source task stage: Based on the source task training set, a stacked autoencoder network is constructed and trained to obtain a trained transformer fault diagnosis model; Target task stage: Based on the target task training set, the trained transformer fault diagnosis model is transferred to the tree-line discharge fault task through transfer learning to train a tree-line discharge fault feature recognition model to extract tree-line discharge fault features.

[0009] The current signal of the acquisition circuit is input into the tree-line discharge fault feature recognition model to obtain the tree-line discharge fault features.

[0010] Furthermore, a training set for the target task is constructed, including:

[0011] When a tree-line discharge fault occurs, the original transient signal is collected and preprocessed to obtain the preprocessed transient signal.

[0012] Based on the preprocessed transient signal, a two-dimensional coefficient matrix corresponding to the current of each phase is generated;

[0013] Based on the two-dimensional coefficient matrix of each phase current, the time-frequency diagrams of each phase are obtained, and the training set for the target task is constructed.

[0014] Further, preprocessing includes filtering and denoising the original transient signal using wavelet thresholding, and / or normalization preprocessing, to obtain the preprocessed transient signal.

[0015] Further, the filtering and denoising process includes:

[0016] Select a wavelet basis function suitable for transient signals, perform multi-scale wavelet decomposition on the original transient signal, and obtain the decomposed signal;

[0017] Noise in a signal is removed by thresholding high-frequency detail coefficients.

[0018] The denoised signal is obtained by performing an inverse wavelet transform on the coefficients after thresholding.

[0019] Normalization processing includes:

[0020] Calculate the minimum and maximum values ​​of the signal;

[0021] Based on the characteristics of the transient signal of tree-line discharge fault, a minimum-maximum normalization method is used for normalization processing.

[0022] Furthermore, the step of generating the two-dimensional coefficient matrix for each phase current includes:

[0023] Based on the characteristics of the transient signal of tree-line discharge fault, an improved complex Morlet wavelet is selected to perform continuous wavelet transform on the signal, obtaining information containing different scales and time distributions, and generating a two-dimensional coefficient matrix of each phase current.

[0024] Furthermore, the steps for constructing the training set for the target task include:

[0025] The two-dimensional coefficient matrix of each phase current is converted into an image through pseudo-color encoding to obtain the time-frequency diagram of each phase;

[0026] The three-phase time-frequency diagrams A, B, and C are stacked according to their channels to form a three-channel input tensor, which is used to construct the training set for the target task.

[0027] Furthermore, the source task phase includes:

[0028] Build an autoencoder and stack multiple autoencoder layers to generate a stacked autoencoder network;

[0029] Based on the source task training set, a stacked autoencoder is trained layer by layer: the first layer autoencoder directly processes the original input, generates reconstructed data through the encoder and decoder, calculates the difference between the input data and the reconstructed data, and calculates the gradient of the loss function with respect to the weights and biases through backpropagation to update the weights and biases of the autoencoder; after the first layer pre-training is completed, the hidden layer representation of the previous layer is used as the input of the next layer autoencoder, and the iteration is repeated until the predetermined network depth is reached.

[0030] After completing layer-by-layer training, a quasi-output layer is added to the top layer of the stacked autoencoder network; the hidden representation of the last autoencoder layer is connected to the output layer to diagnose transformer faults; the stacked autoencoder network is used for pre-training in transformer fault diagnosis, and then the parameters are globally fine-tuned to update the parameters, resulting in a trained transformer fault diagnosis model.

[0031] Furthermore, the target task phase includes:

[0032] The weights of the encoder part of the trained transformer fault diagnosis model are transferred to the encoder of the target task as a feature extractor.

[0033] After the feature extractor, a total output layer is introduced, and the final feature representation of the feature extractor is connected to the total output layer to construct a tree-line discharge fault feature recognition model.

[0034] Based on the target task training set, the tree-line discharge fault feature recognition model is trained, and the parameters of all layers are updated to obtain the trained tree-line discharge fault feature recognition model.

[0035] Furthermore, it also includes: mapping the obtained tree-line discharge fault feature matrix to a grayscale image to determine the intensity of the fault signal; or / and mapping the obtained tree-line discharge fault feature matrix to a three-dimensional coordinate axis to determine the correlation of the fault signal.

[0036] On the other hand, the present invention also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above methods.

[0037] The distribution network grounding fault feature identification method and computer storage medium provided by this invention are a deep abstract feature mining method for tree-line discharge faults based on a combination of deep learning and transfer learning. On the one hand, through the pre-training process in the source task stage, transformer fault features are extracted, which can make full use of the abundant and stable transformer fault data and the relatively stable transformer fault discharge signals, making model training simpler and more efficient, and improving training efficiency. More importantly, the data similarity and causal correlation between the two make the extracted transformer fault features highly consistent and matched with the tree-line discharge feature form and fault time. On the other hand, through the post-training process in the target task stage, tree-line discharge fault-specific features are extracted, and the transfer data is specialized to reflect the uniqueness of the target task, which can further improve the model accuracy. Attached Figure Description

[0038] Figure 1 This is a flowchart of an embodiment of the distribution network grounding fault feature identification method of the present invention;

[0039] Figure 2 The time-frequency diagram is a combined representation of the currents in phases A, B, and C.

[0040] Figure 3 A schematic diagram of the structure of an embodiment of a stacked autoencoder network;

[0041] Figure 4 A flowchart of a stacked autoencoder network for pushover transfer learning. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0043] It should be noted that if the embodiments of the present invention involve directional indications, such as up, down, left, right, front, back, etc., these directional indications are only used to explain the relative positional relationships and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly. Furthermore, if the embodiments of the present invention involve descriptions such as "first," "second," "S1," "S2," "step one," "step two," etc., these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance, or implicitly indicating the number of technical features indicated or the order of method execution. Those skilled in the art will understand that anything that does not violate the inventive concept and is within the scope of the present invention should be included in the protection scope of the present invention.

[0044] like Figure 1 As shown, the present invention provides a method for identifying grounding fault characteristics in a distribution network, comprising:

[0045] S1: Collect transformer fault signals and construct the source task training set;

[0046] S2: Collect tree-line discharge fault signals and construct a training set for the target task;

[0047] S3: Construct and train a tree-line discharge fault feature recognition model, including:

[0048] Source task phase: Based on the source task training set, construct and train a stacked autoencoder network to obtain a trained transformer fault diagnosis model;

[0049] Target task phase: Based on the target task training set, the trained transformer fault diagnosis model is transferred to the tree-line discharge fault task through transfer learning to train a tree-line discharge fault feature recognition model to extract tree-line discharge fault features.

[0050] S4: Acquire the current signal of the acquisition circuit, input it into the tree-line discharge fault feature recognition model, and obtain the tree-line discharge fault feature matrix.

[0051] This embodiment presents a method for identifying grounding fault features in a distribution network, which is a method for mining deep abstract features of tree-line discharge faults based on a combination of deep learning and transfer learning. Although deep learning and transfer learning are existing technologies, this invention transfers pre-trained model parameters from transformer fault tasks to tree-line discharge fault feature identification. This is not a simple combination or conventional technology selection, but rather the result of creative effort.

[0052] Firstly, the inventors, while analyzing transformer fault signals and tree-line discharge fault signals, creatively discovered that the two possess data similarity and causal correlation:

[0053] A: Data similarity: mainly stems from the core fault characteristics of "partial discharge" or "insulating dielectric breakdown".

[0054] A1: Tree-to-line discharge fault: Essentially, it is caused by the breakdown of the air gap insulation between the overhead power line conductor and the tree (including branches, leaves, and trunk). When the tree approaches the conductor to a critical distance (determined by the line voltage level, such as approximately 1.5m for a 110kV line), the air medium is ionized under a strong electric field, resulting in partial discharge (corona discharge, spark discharge). If the tree directly contacts the conductor, a "tree-to-line" conductive path can be formed, triggering a continuous arc discharge.

[0055] A2: Transformer fault: When the internal insulation structure of a transformer (such as oil-paper insulation, air insulation, solid insulation) ages, becomes damp, or is partially damaged, a strong electric field concentration area will be formed at the weak insulation points (such as between winding turns, poor core grounding, and inside the bushing), causing partial discharge (such as corona discharge, surface discharge, and spark discharge); in severe cases, it will develop into insulation breakdown, which is consistent with the "gap breakdown-discharge" logic of tree line discharge.

[0056] A3: Both produce the following during "partial discharge":

[0057] High-frequency pulse current: A high-frequency pulse current in the nanosecond (ns) range (frequency usually covers tens of kHz to hundreds of MHz) is generated at the moment of discharge. The pulse current of tree line discharge propagates through the line conductor, while the pulse current of transformer partial discharge is conducted to the ground through the transformer shell and grounding lead.

[0058] Broadband electromagnetic radiation: The discharge process radiates broadband electromagnetic waves (covering radio frequency and ultra-high frequency bands) into the surrounding space, which can be detected outside the device by electromagnetic coupling sensors (such as ultra-high frequency UHF sensors and radio frequency CT sensors);

[0059] Harmonics and interference signals: The discharge signals of both will be superimposed on the system fundamental frequency (50 / 60Hz), generating non-integer harmonics (such as the 3rd, 5th, 7th and higher harmonics), accompanied by high-frequency noise interference;

[0060] Therefore, the two sets of data are highly similar, and they are accompanied by rapid charge transfer, resulting in similar electrical signal characteristics.

[0061] B: Correlation of Causes: Transformer fault signals and tree-line discharge faults are not completely independent; they may be correlated through "shared fault causes" and "system disturbance propagation."

[0062] B1: Some external factors can simultaneously increase the probability of two types of failures, forming a "causal association".

[0063] Examples: Thunderstorms: When lightning strikes a power line or transformer, it generates instantaneous overvoltages (which can reach hundreds of thousands of volts). On one hand, this can break down the air gap between the conductor and the tree, inducing tree-wire discharge; on the other hand, it can penetrate the transformer's interior, exacerbating partial discharge at weak points in the insulation, or even directly breaking down the insulation. Strong winds and icing: Strong winds can cause tree branches to sway dramatically, shortening the distance to the conductor (inducing tree-wire discharge); simultaneously, icing can increase line sag or cause ice accumulation on the transformer bushing surface (reducing insulation performance and inducing bushing partial discharge). System voltage rise: When the system experiences voltage rises due to reduced load or excessive reactive power compensation (such as power frequency overvoltage), the breakdown field strength between the conductor and the tree decreases (air insulation strength decreases with increasing voltage), increasing the risk of tree-wire discharge; simultaneously, the electric field strength experienced by the transformer's internal insulation increases, significantly increasing the amplitude of partial discharge pulses, making fault signals easier to detect.

[0064] B2: Impact Conduction: Because power systems are interconnected, line faults and equipment faults can affect each other through voltage and current disturbances; a fault in one system may trigger or exacerbate a fault in another. Example:

[0065] Tree discharge can induce transformer failures: If tree discharge develops into a line short circuit (such as a three-phase short circuit caused by a fallen tree), it will generate a huge short-circuit current. When this current passes through the transformer windings, it will cause electrodynamic impact on the windings (which may lead to winding deformation and insulation wear), indirectly inducing partial discharge faults inside the transformer. The high-frequency harmonic current generated by tree discharge is injected into the system, which will generate additional losses in the transformer windings, leading to local overheating, accelerating insulation aging, and creating hidden dangers for subsequent transformer failures.

[0066] Transformer faults affect tree line discharge detection: If a transformer experiences severe partial discharge (such as a winding short circuit), the high-frequency electromagnetic radiation it generates will propagate along the line and may superimpose with the electromagnetic signal of tree line discharge, interfering with the detection and location of tree line discharge faults (especially in areas where the line and transformer are close together, such as the outgoing end of a substation).

[0067] In summary, the applicant creatively discovered the data similarity and causal correlation between transformer fault signals and tree-line discharge fault signals, leading to similar characteristics and temporal correlation between the two data. The discovery of this problem itself required creative effort.

[0068] Secondly, based on this inventive discovery, the inventors creatively present the inventive points of this invention, which construct a tree-line discharge fault feature recognition model based on two stages: 1. Source task stage: Based on the source task training set constructed from transformer fault signals, a stacked autoencoder network is constructed and trained to obtain a trained transformer fault diagnosis model; 2. Target task stage: Based on the target task training set constructed from tree-line discharge fault signals, the trained transformer fault diagnosis model is transferred to the tree-line discharge fault task through transfer learning to train a tree-line discharge fault feature recognition model for subsequent tree-line discharge fault feature extraction.

[0069] It is important to emphasize that the key to this invention lies in the inventors' decision to train the neural network model in two stages instead of directly using the target task training set constructed from tree-line discharge fault signals. Because tree-line discharge faults are less common than transformer faults, collecting training data is more difficult, leading to training difficulties and hindering the acquisition of a high-precision target model. More importantly, the discharge signals from tree-line discharge faults are unstable (affected by wind, tree swaying, humidity, and large fluctuations in pulse frequency), making training difficult to converge and significantly increasing training time.

[0070] Therefore, a creative approach combining two stages is adopted: On the one hand, the pre-training process in the source task stage extracts transformer fault features, which not only makes full use of the abundant and stable transformer fault data, but also the relatively stable transformer fault discharge signals, making model training simpler and more efficient, thus improving training efficiency; more importantly, the data similarity and causal correlation between the two stages ensure that the extracted transformer fault features are highly consistent and matched with the tree-line discharge feature form and fault time; on the other hand, the post-training process in the target task stage extracts tree-line discharge fault-specific features, specializing the transfer data, reflecting the uniqueness of the target task, and further improving the model accuracy.

[0071] (I) Step S1:

[0072] Specifically, S1: When a transformer fails, it can collect signals such as voltage and current, but this is not a limitation, to construct a source task training set; the method, type, and content of the transformer fault signal collection are not subject to many requirements, but: since the source task of this transformer fault diagnosis serves the target task of tree-line discharge fault feature identification, the fault signal only needs to be compatible with the tree-line discharge transient signal below; if current is collected below, then current is collected here; if voltage is collected below, then voltage is collected here; if current and voltage are collected below, then current and voltage are also collected here; if other signals are collected below, then other signals are also collected here.

[0073] (II) Step S2:

[0074] Specifically, consistent with step S1, S2: when a tree-line discharge fault occurs, signals such as voltage and current can be collected, but this is not a limitation, to construct a training set for the target task; the collection method, type, and content of the transient signals of the tree-line discharge fault are not subject to many requirements.

[0075] Preferably, in steps S1 and S2, the selected fault signals are subjected to noise reduction and amplitude normalization processing to map the signal amplitude to a uniform range.

[0076] More preferably, taking the tree-line discharge fault signal as an example, step S2, constructing the target task training set, includes:

[0077] S21: When a tree-line discharge fault occurs, the original transient signal is collected and preprocessed to obtain the preprocessed transient signal;

[0078] S22: Generate a two-dimensional coefficient matrix corresponding to the current of each phase based on the preprocessed transient signal;

[0079] S23: Based on the two-dimensional coefficient matrix of each phase current, the time-frequency diagrams of each phase are obtained, and the training set for the target task is constructed.

[0080] More preferably, S21: preprocessing, including filtering and denoising the original transient signal using wavelet threshold denoising, and / or normalization preprocessing, to obtain the preprocessed transient signal.

[0081] Specifically, the filtering and noise reduction process includes:

[0082] S211: Select a wavelet basis function suitable for the transient signal, perform multi-scale wavelet decomposition on the original transient signal, and obtain the decomposed signal.

[0083] Specifically, the Daubechies 4 wavelet (db4) can be used, and the formula is as follows:

[0084] Scaling function (φ(t)):

[0085]

[0086] Where N is the support length of the size function, for example: N = 4 means the support length of the size function is 4; h k Here are the coefficients of the low-pass filter, and k is a non-negative integer.

[0087] Wavelet function (ψ):

[0088]

[0089] After wavelet decomposition:

[0090]

[0091] Where x[n] is the original transient signal, cA J For the approximate coefficients (low-frequency components) of the J-th layer, cD J represents the detail coefficients (high-frequency components) of layer J.

[0092] S212: Remove noise from the signal by thresholding the high-frequency detail coefficients, including:

[0093] Soft thresholding can smooth the output and reduce reconstruction oscillations. The formula is as follows:

[0094]

[0095] Hard thresholding preserves mutation characteristics; the formula is as follows:

[0096]

[0097] Where, λ J is the threshold of the Jth layer.

[0098] S213: Perform inverse wavelet transform (IDWT) on the thresholded coefficients to obtain the denoised signal.

[0099]

[0100] Specifically, the normalization process includes:

[0101] S214: Calculate the minimum value of the signal. and maximum value The formula is:

[0102]

[0103] S215: Combining the characteristics of the transient signal of tree-line discharge fault, a minimum-maximum normalization method is adopted to compress the entire signal amplitude range to between [-1, 1]. The formula is as follows:

[0104]

[0105] in, The normalized signal value is the final processed value.

[0106] More preferably, step S22 includes:

[0107] S221: Perform unified sampling verification on the preprocessed transient signal to ensure that the three-phase signals have the same time resolution and number of sampling points; it is worth noting that this step is optional and not necessary.

[0108] S222: Based on the characteristics of the transient signal of tree-line discharge fault, an improved complex Morlet wavelet is selected to perform continuous wavelet transform (CWT) on the signal to obtain information containing different scales and time distributions, and generate a two-dimensional coefficient matrix Z for each phase current.

[0109] Specifically, the complex Morlet wavelet formula is:

[0110]

[0111] In the formula: β is the bandwidth factor, which controls the time-frequency resolution of the wavelet;

[0112] Gaussian envelope, dynamic balance between time and frequency localization capabilities;

[0113] It is a complex sine wave signal.

[0114] For signals Its continuous wavelet transform is defined as:

[0115]

[0116] Where: a is the scale parameter; r is the translation parameter; ψ * It is the conjugate of the wavelet function.

[0117] By adjusting the scale, signal components of different frequencies are obtained; by translation, the instantaneous characteristics of the signal are captured on the time axis. The result of CWT is a two-dimensional coefficient matrix Z, representing the signal intensity at each time point and corresponding frequency (scale).

[0118] More preferably, S23 includes:

[0119] S231: Convert the two-dimensional coefficient matrix of each phase current into an image format through pseudo-color encoding to obtain the time-frequency diagram of each phase;

[0120] S232: Stack the three-phase time-frequency maps A / B / C by channel to form a three-channel input tensor, which is used to construct the training set for the target task. That is, it serves as the input to the subsequent autoencoder network, in the form shown below;

[0121] X input =[Z A (c,d),Z B (c,d),Z C (c,d)]

[0122] Among them, X input ∈R 3×M×N Z is a three-channel input tensor used as the input to the subsequent model. A(c,d) represents the value of the coefficient matrix of phase A at (c,d), and the same applies to phases B and C.

[0123] (III) Step S3:

[0124] Specifically, a stacked auto-encoder (SAE) network is first pre-trained in transformer fault diagnosis to extract deep features. Then, the pre-trained auto-encoder model in transformer fault diagnosis is transferred to the tree-line discharge fault task through push transfer learning to obtain the tree-line discharge fault feature recognition model.

[0125] Preferred, the source task phase includes:

[0126] S31: Build an autoencoder and stack multiple autoencoder layers to generate a stacked autoencoder network (SAE);

[0127] Specifically: First, an auto-encoder (AE) layer is built, which is an unsupervised neural network model. Its core goal is to compress the input data into a low-dimensional latent space without the aid of labels, and to restore the original input as much as possible from this representation, thereby achieving effective feature learning.

[0128] An autoencoder consists of three layers: an input layer, a hidden layer, and an output layer. The number of input neurons is the same as the number of output neurons. The process from input neurons to hidden layer neurons is called encoding, while the process from hidden layer neurons to output layer neurons is called decoding.

[0129] Specifically, including:

[0130] 1) Taking the above time-frequency graph as an example, the pixel values ​​of the time-frequency graph are used as data samples and input into the autoencoder, where the sample is represented as X. input ={X1,X2,…X n}

[0131] 2) The encoder maps high-dimensional input data X to a low-dimensional hidden representation h1. Through the encoder, the input data is gradually compressed into a low-dimensional representation, thereby extracting the key features of the data. The encoder takes the following form:

[0132] h1 = f(W1X + b1)

[0133] Where W1∈R m×n It is the encoder's weight matrix, b1∈R m is the bias vector, f(·) is the non-linear activation function, and m is the hidden layer dimension of the encoder output, which is usually smaller than the input dimension n.

[0134] 3) The decoder restores the low-dimensional representation h1 output by the encoder to reconstructed data that is as close as possible to the original input. The decoder takes the form shown below:

[0135]

[0136] Where W2∈R n×m This is the weight matrix of the decoder, b2∈R n×m is the bias vector, and g(·) is the activation function of the decoder. The output dimension of the decoder is the same as the input dimension n.

[0137] 4) The output layer will reconstruct the signal Output.

[0138] Then, multiple autoencoders are stacked layer by layer to form a stacked autoencoder network.

[0139] S32: Layer-by-layer training of stacked autoencoders: The first layer of the autoencoder directly processes the original input, generates reconstructed data through the encoder and decoder, calculates the difference between the input data and the reconstructed data, and calculates the gradient of the loss function with respect to the weights and biases through backpropagation to update the weights and biases of the autoencoder; after the first layer is pre-trained, the hidden layer representation of the previous layer is used as the input of the next layer of the autoencoder, and the iteration is repeated until the predetermined network depth is reached.

[0140] Specifically: Layer-wise pretraining includes:

[0141] A. Training the first autoencoder

[0142] The first-layer autoencoder directly processes the raw input X, and its training process is as follows:

[0143] a. Forward Propagation

[0144] The input data X is fed into the autoencoder, and the reconstructed data is generated through the encoder and decoder. In the encoder stage, the input data X is used to generate a low-dimensional hidden representation h1. In the decoder stage, the low-dimensional hidden representation h1 is used to recover the reconstructed data.

[0145] b. Loss function calculation

[0146] After forward propagation is complete, it is necessary to calculate the input data X and the reconstructed data. The differences between them. The loss function, optionally the cross-entropy loss function, is defined as:

[0147]

[0148] Where n is the dimension of the input data, X i and These are the i-th input value and the reconstructed value, respectively.

[0149] c. Backpropagation

[0150] To optimize the encoder and decoder parameters, the gradient of the loss function with respect to the weights and biases needs to be calculated via backpropagation. This includes the gradient of the loss function with respect to the encoder parameters and the gradient of the loss function with respect to the decoder parameters.

[0151] 1) Calculate the gradient of the loss function with respect to the encoder parameters.

[0152] The gradient with respect to the weight matrix W1 is:

[0153]

[0154] The gradient with respect to the bias matrix b1 is:

[0155]

[0156] 2) Calculate the gradient of the loss function with respect to the decoder parameters:

[0157] The gradient with respect to the weight matrix W2 is:

[0158]

[0159] The gradient of the bias matrix b2 is:

[0160]

[0161] d. Parameter Update

[0162] The Adam optimization algorithm is used to update the weights and biases of the autoencoder. Adam considers the historical gradient information of each parameter when updating the parameters, and smooths and scales the gradients, avoiding the oscillation or slow convergence problems caused by excessively large or small learning rates in traditional gradient descent algorithms.

[0163] The updated formula is shown below:

[0164]

[0165] Where η is the learning rate, which controls the step size of each update.

[0166] B. Training the second autoencoder

[0167] After the first layer pre-training is completed, the hidden layer representation of the first layer is used as the input of the second layer autoencoder, and similar training steps are repeated:

[0168] a. Positive propagation

[0169] The second-layer autoencoder extracts higher-level features h2 from h1:

[0170] h2 = f(W2h1 + b2)

[0171] Decode the hidden layer representation h2 into reconstructed data.

[0172]

[0173] b. Loss function calculation

[0174] Calculate hidden layer h1 and reconstructed Losses between:

[0175]

[0176] c. Backpropagation and parameter update

[0177] The parameters of the second-layer autoencoder are updated through backpropagation and optimization algorithms.

[0178] C. Pre-training of subsequent layers

[0179] Following a similar process, the third layer and subsequent autoencoders are pre-trained sequentially. Each layer is trained independently, with the hidden layer representation of the previous layer serving as the input to the current layer, until the predetermined network depth is reached.

[0180] S33: Fine-tuning to obtain a trained stacked autoencoder network: After completing layer-by-layer training, a quasi-output layer is added to the top layer of the stacked autoencoder network; the hidden representation of the last autoencoder layer is connected to the output layer to diagnose transformer faults; the stacked autoencoder network is used for pre-training in transformer fault diagnosis, and then the parameters are updated globally to obtain a trained transformer fault diagnosis model.

[0181] Specifically, after step S32, the stacked autoencoder network has obtained the initial network weights. At this point, only supervised overall fine-tuning is needed to further optimize the network performance.

[0182] A. Introducing a quasi-output layer

[0183] Specifically, since transformer fault diagnosis is a classification or regression task, an output layer can be added to the top layer of the stacked autoencoder network, and the hidden representation h of the last autoencoder layer can be passed to the output layer. nConnect to the output layer. For transformer fault diagnosis tasks, the option is:

[0184] y = softmax(W y h n +b y )

[0185] Among them, W y and b y These are the parameters of the output layer; the softmax function is used for the output of multi-class classification tasks.

[0186] B. Global Fine-tuning

[0187] Global forward propagation: Treat the entire network as a whole, pass the input data to the output layer, and obtain the predicted value y.

[0188] Calculate the global loss: For transformer fault diagnosis, which is a classification task, cross-entropy loss is used.

[0189] y = softmax(W y h n +b y )

[0190] Global backpropagation: The loss is backpropagated to all parameters of the entire network, updating the parameters of each layer, including the encoder, decoder, and output layer.

[0191] Global parameter update: Update the network parameters using the Adam optimization algorithm.

[0192] Through overall fine-tuning, stacked autoencoder networks can further optimize network weights for specific tasks, thereby improving model performance.

[0193] A more preferred approach: In the target task phase, a direct transfer learning mechanism is introduced to transfer some weights of the SAE model trained in the transformer fault diagnosis task to the target task, followed by overall fine-tuning; including:

[0194] S34: Transfer the weights of the encoder part of the trained transformer fault diagnosis model (i.e., the trained stacked autoencoder network) to the encoder of the target task as a feature extractor.

[0195] Specifically, in the initial stage, the encoder weights are fixed, and only subsequent layers for the target task are trained. At this time, the target task model is initialized with the weights of the source task model, i.e.:

[0196]

[0197] S35: After the feature extractor, a total output layer is introduced, and the last feature representation of the feature extractor is connected to the total output layer to build a tree-line discharge fault feature recognition model.

[0198] Specifically, for the target task, a new overall output layer is introduced, connecting the high-level feature representations to the target task's output layer. Let the input features be h. n These are features extracted from the hidden layer of the last autoencoder. The calculation process of the output layer is as follows:

[0199] y=σ(W y h n +b y )

[0200] Among them: W y It is the weight matrix of the output layer, b y σ is the bias vector of the output layer, and σ is the sigmoid activation function, defined as...

[0201] S36: Based on the target task training set, train the tree-line discharge fault feature recognition model, update the parameters of all layers, and obtain the trained tree-line discharge fault feature recognition model.

[0202] Specifically, in the target task, the entire network is fine-tuned using the target task's label y to optimize its parameters. The goal of this fine-tuning is to update the parameters of all layers using supervised learning methods, allowing the model to better adapt to the target task data. This includes:

[0203] a: Forward propagation, input target task data x target The predicted value y is calculated through the entire network.

[0204] b: Calculate the loss function. For feature mining tasks, the Mean Squared Error (MSE) loss function is usually used to measure the difference between the network's output y and the target feature y. true The differences between them are defined as follows:

[0205]

[0206] Where n is the dimension of the output feature, y i and y true,i These are the predicted and actual values ​​of the i-th feature, respectively.

[0207] c: Backpropagation and parameter update: The backpropagation algorithm is used to calculate the gradient of the loss function with respect to the parameters of each layer, and then the optimization algorithm (Adam) is used to update the model parameters.

[0208] d: The formula for updating the weights and biases of the l-th layer is:

[0209]

[0210] Where η is the learning rate. and These represent the gradients of the loss function with respect to the weights and biases of the l-th layer, respectively.

[0211] (IV) Step S4:

[0212] After completing the above preparations, in the fault feature identification stage, the current transient signal of the circuit is collected, further processed according to step S2, and then input into the trained tree-line discharge fault feature identification model to extract the tree-line discharge fault features.

[0213] More preferably, to further improve the intuitiveness of fault display and analysis, it also includes:

[0214] S5: Map the obtained tree-line discharge fault feature matrix to a grayscale image to determine the intensity of the fault signal; or / and map the obtained tree-line discharge fault feature matrix to a three-dimensional coordinate axis to determine the correlation of the fault signal.

[0215] Specifically, the deep feature matrix F extracted by the model can be mapped to a grayscale image, using color brightness to visually display feature information, thereby quickly identifying local high-intensity features. Specifically: the deep feature matrix F extracted by the neural network is used as input to the image pixel values, and the matplotlib.pyplot.imshow() module in Python can be used to render this matrix as a grayscale image; for the feature matrix F... ij When F ij When the value is large, the corresponding area on the grayscale will be displayed as a brighter color, which represents the component with higher intensity in the fault signal; conversely, it will be darker, in order to determine the intensity of the fault signal.

[0216] Or / and, the deep feature matrix F ij This is displayed to form a 3D image, allowing for the analysis of interactions between multiple features. Specifically, the deep feature matrix F is mapped to three key dimensions onto a three-dimensional coordinate system, optionally using the `plotly` module in Python to generate the 3D image. The O-axis represents time, the P-axis represents frequency, and the Q-axis represents the magnitude of the feature values. In the 3D image, certain feature points stand out significantly, indicating a sudden increase in the feature values ​​of the fault signal within that time or frequency range. By observing the relative positions of the points in the image, we can determine which features are highly correlated and which are independent, thus establishing the correlation degree of the fault signal.

[0217] This invention achieves efficient extraction and visualization analysis of tree-line discharge fault features by introducing a feature mining strategy that combines deep learning and transfer learning. First, wavelet thresholding denoising and amplitude normalization effectively suppress high-frequency noise in transient signals and unify feature scales. Although established methods are used, they are all selected based on the characteristics of tree-line discharge fault signals, providing high-quality input for subsequent modeling. Second, a creative approach utilizes continuous wavelet transform to generate time-frequency maps of three-phase currents, and then constructs a three-channel input tensor as the specific form of model input. This is not a simple selection but fully reveals the local characteristics of the signal in the time-frequency domain, making the features more comprehensive and distinguishable. Furthermore, a stacked autoencoder network is used for layer-by-layer pre-training of transformer fault diagnosis, which can automatically learn multi-layer abstract features and avoid the limitations of manual feature engineering. Combined with direct transfer learning, the model pre-trained in transformer fault diagnosis is transferred to the tree-line discharge fault task, which not only significantly reduces the dependence of the target task on large-scale samples but also accelerates model convergence and improves diagnostic accuracy and robustness in small-sample scenarios. Finally, by extracting the intensity and correlation of fault signals, the fault diagnosis results can be displayed more intuitively, possessing interpretability and engineering usability. Therefore, this method achieves efficient mapping from raw signals to deep abstract features in tree-line discharge fault diagnosis, providing strong support for improving the safety and stability of power system operation.

[0218] On the other hand, the present invention also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above-mentioned distribution network grounding fault feature identification methods.

[0219] On the other hand, the present invention also provides a terminal device, including a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to execute any of the above-mentioned distribution network grounding fault feature identification methods.

[0220] For example, the program code can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the program code in the terminal device.

[0221] The terminal device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the terminal device may also include input / output devices, network access devices, buses, etc.

[0222] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0223] The memory can be an internal storage unit of the terminal device, such as a hard drive or RAM. The memory can also be an external storage device of the terminal device, such as a plug-in hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units of the terminal device. The memory is used to store the program code and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output.

[0224] The aforementioned computer storage medium and terminal equipment are created based on the aforementioned method for identifying grounding fault characteristics in power distribution networks. Their technical functions and beneficial effects will not be elaborated here. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0225] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for identifying grounding fault characteristics in a distribution network, characterized in that, include: Collect transformer fault signals and construct a source task training set; Collect treeline discharge fault signals and construct a training set for the target task; The process of constructing and training a tree-line discharge fault feature recognition model includes: Source task stage: Based on the source task training set, a stacked autoencoder network is constructed and trained to obtain a trained transformer fault diagnosis model; Target task stage: Based on the target task training set, the trained transformer fault diagnosis model is transferred to the tree-line discharge fault task through transfer learning to train a tree-line discharge fault feature recognition model to extract tree-line discharge fault features. The current signal of the acquisition circuit is input into the tree-line discharge fault feature recognition model to obtain the tree-line discharge fault features.

2. The method according to claim 1, characterized in that, Construct the training set for the target task, including: When a tree-line discharge fault occurs, the original transient signal is collected and preprocessed to obtain the preprocessed transient signal. Based on the preprocessed transient signal, a two-dimensional coefficient matrix corresponding to the current of each phase is generated; Based on the two-dimensional coefficient matrix of each phase current, the time-frequency diagrams of each phase are obtained, and the training set for the target task is constructed.

3. The method according to claim 2, characterized in that, Preprocessing includes filtering and denoising the original transient signal using wavelet thresholding, and / or normalization preprocessing, to obtain the preprocessed transient signal.

4. The method according to claim 3, characterized in that, Filtering and noise reduction processing includes: Select a wavelet basis function suitable for transient signals, perform multi-scale wavelet decomposition on the original transient signal, and obtain the decomposed signal; Noise in a signal is removed by thresholding high-frequency detail coefficients. The denoised signal is obtained by performing an inverse wavelet transform on the coefficients after thresholding. Normalization processing includes: Calculate the minimum and maximum values ​​of the signal; Based on the characteristics of the transient signal of tree-line discharge fault, a minimum-maximum normalization method is used for normalization processing.

5. The method according to claim 2, characterized in that, The steps for generating the two-dimensional coefficient matrix of each phase current include: Based on the characteristics of the transient signal of tree-line discharge fault, an improved complex Morlet wavelet is selected to perform continuous wavelet transform on the signal, obtaining information containing different scales and time distributions, and generating a two-dimensional coefficient matrix of each phase current.

6. The method according to claim 2, characterized in that, The steps for constructing a training set for the target task include: The two-dimensional coefficient matrix of each phase current is converted into an image through pseudo-color encoding to obtain the time-frequency diagram of each phase; The three-phase time-frequency diagrams A, B, and C are stacked according to their channels to form a three-channel input tensor, which is used to construct the training set for the target task.

7. The method according to claim 1, characterized in that, The source task phase includes: Build an autoencoder and stack multiple autoencoder layers to generate a stacked autoencoder network; Based on the source task training set, a stacked autoencoder is trained layer by layer: the first layer autoencoder directly processes the original input, generates reconstructed data through the encoder and decoder, calculates the difference between the input data and the reconstructed data, and calculates the gradient of the loss function with respect to the weights and biases through backpropagation to update the weights and biases of the autoencoder; after the first layer pre-training is completed, the hidden layer representation of the previous layer is used as the input of the next layer autoencoder, and the iteration is repeated until the predetermined network depth is reached. After completing layer-by-layer training, a quasi-output layer is added to the top layer of the stacked autoencoder network; the hidden representation of the last autoencoder layer is connected to the output layer to diagnose transformer faults; the stacked autoencoder network is used for pre-training in transformer fault diagnosis, and then the parameters are globally fine-tuned to update the parameters, resulting in a trained transformer fault diagnosis model.

8. The method according to claim 1, characterized in that, The target task phase includes: The weights of the encoder part of the trained transformer fault diagnosis model are transferred to the encoder of the target task as a feature extractor. After the feature extractor, a total output layer is introduced, and the final feature representation of the feature extractor is connected to the total output layer to construct a tree-line discharge fault feature recognition model. Based on the target task training set, the tree-line discharge fault feature recognition model is trained, and the parameters of all layers are updated to obtain the trained tree-line discharge fault feature recognition model.

9. The method according to any one of claims 1 to 8, characterized in that, Also includes: The obtained tree-line discharge fault feature matrix is ​​mapped to a grayscale image to determine the intensity of the fault signal; Or / and, the treeline discharge fault feature matrix will be obtained, which can be mapped to the three-dimensional coordinate axis to determine the correlation of the fault signal.

10. A computer storage medium, characterized in that, It stores executable program code; the executable program code is used to perform the method of any one of claims 1 to 9.

Images