Early warning model training and early warning method and device for distribution transformer inter-turn short circuit

Abstract

The application discloses a power distribution transformer inter-turn short circuit early warning model training and early warning method and device. The method generates simulation supplementary data by obtaining real sample data and introducing a transformer simulation model based on a time-varying transition resistance model of a Sigmoid function to generate simulation supplementary data, and fuses and constructs a high-dimensional multi-channel electrical characteristic matrix. According to a dynamic evolution track, four evolution stages of normal operation, extremely early latency, accelerated degradation and complete breakdown are divided, and auxiliary correction and state coding labeling are carried out based on a zero sequence sensitive multi-physical field coupling degradation index to construct a hybrid state data set. Then, an early intelligent warning model is constructed, and the hybrid state data set is used for training and optimization to obtain an optimal network model. During early warning, real-time diagnosis data is collected and input into the optimal network model for forward reasoning classification, so that online monitoring and early warning of insulation degradation are realized. The application effectively solves the problem of insufficient weak fault samples, and improves the accuracy and reliability of early warning.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring and intelligent fault early warning technology, specifically involving a training method and equipment for an early warning model of inter-turn short-circuit faults in distribution transformers based on multi-physics field features and hybrid deep learning. Background Technology

[0002] Distribution transformers are key equipment in power distribution networks, responsible for the transformation and distribution of electrical energy. Their operating status directly affects the reliability of power supply and the quality of electricity for end users. During long-term operation, the winding insulation of distribution transformers is subjected to the combined effects of electrical stress, thermal stress, mechanical vibration, short-time overvoltage, and environmental aging, gradually leading to problems such as localized deterioration, decreased insulation strength, and even inter-turn breakdown. Among these, inter-turn short-circuit faults are characterized by high frequency, strong concealment, and rapid development, making them one of the most harmful types of internal faults in distribution transformers.

[0003] In the early stages of inter-turn short-circuit faults, the fault point typically exhibits high impedance or intermittent conduction, with a small short-circuit circulating current and minimal distortion in the three-phase voltage and current at the external ports. The distribution transformer may still appear to be operating normally. However, the localized, weak short-circuit current will continuously increase local winding losses and temperature, further accelerating the thermal aging of the insulation material. When the insulation deteriorates rapidly from a high-resistance state to a low-resistance state, the fault may develop into a severe inter-turn short circuit within a short period, leading to winding burnout, overheating of the insulating oil, equipment shutdown, or even power outages in the distribution network. Therefore, timely identification of weak fault characteristics during the latent stage of inter-turn insulation deterioration is a crucial issue in the condition monitoring and proactive maintenance of distribution transformers.

[0004] Existing fault monitoring methods for distribution transformers mainly include protection methods based on electrical quantities such as current and voltage, chemical diagnostic methods based on dissolved gas analysis in oil, condition detection methods based on partial discharge, vibration, acoustic, or infrared thermal imaging, and data-driven diagnostic methods based on machine learning or deep learning. Traditional electrical quantity protection methods such as overcurrent protection and differential protection usually rely on macroscopic current amplitude or differential threshold, and can only reliably operate after the fault has developed to a certain severity. They also have dead zones for high impedance and weak inter-turn insulation degradation. Although dissolved gas analysis in oil can reflect the internal insulation degradation and overheating of the transformer, it usually has shortcomings such as offline detection, response lag, and difficulty in real-time early warning. Partial discharge, acoustic, vibration, and infrared detection methods require additional dedicated sensors, which are easily affected by on-site electromagnetic noise, environmental noise, installation location, and changes in operating conditions, resulting in high engineering and promotion costs.

[0005] In recent years, with the development of artificial intelligence technology, data-driven transformer fault diagnosis methods have been gradually applied. These methods can automatically learn fault characteristics from multi-dimensional monitoring data, overcoming to some extent the problems of relying on manual threshold settings and expert experience. However, early-stage inter-turn short-circuit fault samples in distribution transformers are extremely scarce in real-world operating scenarios, and the early fault characteristics are weak and easily masked by load fluctuations, three-phase imbalance, and strong electromagnetic interference. Models trained solely on actual operating data are prone to problems such as insufficient samples, weak generalization ability, and high false alarm / false negative rates; models trained solely on simulation data may not be entirely consistent with the actual equipment operating characteristics. Furthermore, shallow machine learning models have limited ability to extract local transient distortions and long-distance evolution dependencies in high-dimensional time-series electrical signals, making it difficult to achieve refined identification of the entire process of inter-turn insulation from normal state, high-resistance latent state, medium-low resistance deterioration state to low-resistance breakdown state.

[0006] Therefore, there is an urgent need for a new early warning technology for inter-turn short-circuit faults in distribution transformers, which can achieve more advanced, accurate, and engineering-generalizable intelligent early warning under conditions of strong noise, small sample size, and early weak faults. Summary of the Invention

[0007] To address the shortcomings of existing technologies, such as the blind spot in monitoring early and weak faults caused by traditional macroscopic protection thresholds, the susceptibility of weak characteristic signals to being masked by strong electromagnetic noise in the field, and the extreme scarcity of real early fault samples, this invention provides a training method and device for early warning models of inter-turn short circuits in distribution transformers based on multi-physics field features and hybrid deep learning. This method provides more accurate and timely warnings with stronger engineering generalization capabilities.

[0008] In a first aspect, this invention proposes a training method for an early warning model of inter-turn short-circuit faults in distribution transformers, comprising: acquiring real sample data, wherein the real sample data includes the zero-sequence component of multi-channel electrical quantities calculated under actual operating scenarios of distribution transformers and experimental fault data; using the real sample data as a physical calibration reference, generating simulation supplementary data by introducing a transformer simulation model based on a time-varying transition resistance model with a Sigmoid function; fusing the real sample data and the simulation supplementary data to construct a high-dimensional multi-channel electrical feature matrix; dividing the high-dimensional multi-channel electrical feature matrix into four evolution states—normal operation, very early latency, accelerated degradation, and complete breakdown—accelerated according to the dynamic evolution trajectory of the time-varying transition resistance model, and performing auxiliary correction and state encoding labeling on the stage division results based on the coupling degradation index of zero-sequence sensitive multiphysics field to construct a hybrid state dataset; constructing an early intelligent warning model based on a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a Bayesian optimization algorithm, and training and optimizing the early intelligent warning model using the hybrid state dataset to obtain the optimal network model.

[0009] In some examples, the time-varying transition resistance model is defined as:

[0010] In the formula, Indicates time-varying transition resistance. and These represent the high resistance value when the insulation is normal and the minimum resistance value when it is completely broken down, respectively. The shape parameter characterizing the degradation rate and Greater than 0, The time offset center for accelerated insulation aging is used; the time-varying transition resistance model is substituted as a dynamic coupling variable into the control equation of the distribution transformer winding to generate simulated high and low voltage three-phase voltage and current signals under each weak degradation stage, and the corresponding simulated zero-sequence component is extracted simultaneously as simulation supplementary data.

[0011] In some examples, the channels of the high-dimensional multi-channel electrical feature matrix include high-voltage side three-phase voltage, high-voltage side three-phase current, low-voltage side three-phase voltage, low-voltage side three-phase current, high-voltage side zero-sequence voltage, high-voltage side zero-sequence current, low-voltage side zero-sequence voltage, and low-voltage side zero-sequence current. When constructing the high-dimensional multi-channel electrical feature matrix, Gaussian white noise with a preset signal-to-noise ratio is superimposed on the simulation supplementary data, and the high-dimensional multi-channel electrical feature matrix is ​​expanded by a time-series translation enhancement method to form feature matrix sub-samples with the same fault state but different start times.

[0012] In some examples, the criterion for dividing the stages according to the dynamic evolution trajectory of the time-varying transition resistance model is: when When, it is determined to be in normal operating condition; when At that time, it was determined to be in a very early latent state; when When, it is judged to be in an accelerated deterioration state; when When the time is right, it is determined to be a complete breakdown state.

[0013] In some examples, the rules for auxiliary correction and state coding labeling include: determining the initial state label of the high-dimensional multi-channel electrical feature matrix corresponding to each time window based on the interval of the time-varying transition resistance model, and determining the adjacent state judgment threshold based on the statistical distribution of the coupling degradation index of the zero-sequence sensitive multiphysics field in the high-dimensional multi-channel electrical feature matrix corresponding to different degradation stages; when the state mapped by the coupling degradation index is consistent with the initial state, the initial state label remains unchanged; when the two are inconsistent, a review is performed in conjunction with auxiliary features, which include the changing trends of short-circuit circulating current, winding loss, local temperature rise, port current distortion, and zero-sequence sensitive factor; if the coupling degradation index exceeds the adjacent state judgment threshold for multiple consecutive time windows, and at least two of the auxiliary features show a continuous changing trend consistent with the direction of insulation degradation, the initial state is corrected to the adjacent more severe degradation state by one level; if the coupling degradation index does not exceed the adjacent state judgment threshold consecutively, and the auxiliary features do not show a continuous degradation trend, the original state label is maintained; a single correction does not cross two or more state levels.

[0014] In some examples, the coupling degradation index is calculated as follows: zero-sequence energy is calculated by extracting zero-sequence channels from the high-dimensional multi-channel electrical feature matrix, and the total port energy is calculated using the three-phase voltage and current channels on the high and low voltage sides. The zero-sequence sensitivity factor is determined based on the ratio of zero-sequence energy to the total port energy. The multi-physics trend coupling factor is calculated based on one or more characteristic parameters among short-circuit circulating current, winding loss, local temperature rise, and port current distortion. The zero-sequence sensitivity factor and the multi-physics trend coupling factor are fused to obtain the coupling degradation index.

[0015] In some examples, in the early intelligent warning model, a one-dimensional convolutional neural network is used to perform convolution operations on the local waveform distortion and high-frequency disturbances of each electrical channel and to reduce dimensionality and noise; a bidirectional long short-term memory network is used to extract long-distance temporal dependency features through the forward and reverse temporal propagation directions; and a Bayesian optimization algorithm uses the validation set classification accuracy, loss function value, or macro-average F1 value as the objective function to adaptively optimize the hyperparameters of the early intelligent warning model through a Gaussian process.

[0016] Secondly, this invention proposes an early warning method for inter-turn short-circuit faults in distribution transformers, comprising: real-time acquisition of multi-channel electrical quantities on the high and low voltage sides of the distribution transformer to be diagnosed and calculation of zero-sequence components; construction of a high-dimensional multi-channel electrical feature matrix to be diagnosed based on the acquired electrical quantities and zero-sequence components; inputting the high-dimensional multi-channel electrical feature matrix to be diagnosed into the optimal network model trained by the early warning model training method for inter-turn short-circuit faults in distribution transformers for forward inference, mapping and classifying to the corresponding evolution state, thereby realizing online monitoring and early intelligent warning of insulation degradation of distribution transformers.

[0017] In some examples, early intelligent warnings include: no warning is triggered when the output of the optimal network model is in normal operation; a level 1 warning signal is issued when the output is in an extremely early latent state; a level 2 warning signal is issued when the output is in an accelerated degradation state; a level 3 warning signal is issued when the output is in a complete breakdown state; and a corresponding level of warning signal is confirmed to be triggered when the output of the optimal network model for multiple consecutive time windows all point to the same abnormal state, or when the probability of the abnormal state exceeds a preset threshold. Otherwise, the monitoring state continues.

[0018] Thirdly, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the model training method or the steps of the early warning method.

[0019] This invention can fully combine actual operating data, experimental fault samples, and multi-physics simulation supplementary samples, utilize zero-sequence components to enhance the sensitivity to local asymmetric degradation of windings, and simultaneously extract local waveform distortion features and long-distance time-series evolution features through a deep learning model, thereby achieving more proactive, accurate, and engineering-generalizable intelligent early warning under conditions of strong noise, small sample size, and early weak faults.

[0020] This invention overcomes the problem of scarce early-stage inter-turn short-circuit fault samples in distribution transformers by fusing real-world operational data, small-sample experimental fault data, and multi-physics simulation supplementary data. It enhances the model's sensitivity to local asymmetric degradation of windings by introducing zero-sequence voltage and current characteristics from both high and low voltage sides. Gaussian white noise injection and time-sequence shift enhancement improve the model's robustness to complex electromagnetic noise and variations in sampling starting points. Joint modeling with 1D-CNN and BiLSTM enables the collaborative extraction of local waveform distortion features and long-distance temporal evolution features. Adaptive optimization of network hyperparameters via Bayesian optimization reduces the risk of local optima caused by manual experience-based settings, thus enabling more proactive and accurate intelligent early warning in the very early stages of inter-turn insulation degradation in distribution transformers. Attached Figure Description

[0021] Figure 1 The flowchart below shows the overall process of an early warning method for inter-turn short-circuit faults in distribution transformers based on multi-physics field features and hybrid deep learning, as provided in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the branch circuit for inter-turn short circuit faults in a distribution transformer and the connection of an external time-varying transition resistor.

[0023] Figure 3This is a flowchart illustrating the process of fusing real sample data and simulation supplementary data in one embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the construction of a high-dimensional multi-channel electrical feature matrix in one embodiment of the present invention.

[0025] Figure 5 This is a schematic diagram illustrating the four-stage division of insulation degradation based on time-varying transition resistance and multi-parameter trend inflection points in one embodiment of the present invention.

[0026] Figure 6 This is a schematic diagram of the BOA-CNN-BiLSTM early intelligent warning model structure in one embodiment of the present invention.

[0027] Figure 7 This is a flowchart of online monitoring and graded early warning of distribution transformers in one embodiment of the present invention. Detailed Implementation

[0028] Example 1: A training method for an early warning model of inter-turn short-circuit faults in distribution transformers.

[0029] This embodiment provides a training method for an early warning model of inter-turn short-circuit faults in distribution transformers. The method includes: acquiring real sample data, which includes the zero-sequence component of multi-channel electrical quantities calculated under actual operating scenarios of the distribution transformer and experimental fault data; using the real sample data as a physical calibration reference, generating supplementary simulation data by introducing a transformer simulation model based on a time-varying transition resistance model using the Sigmoid function; fusing the real sample data with the supplementary simulation data to construct a high-dimensional multi-channel electrical feature matrix; dividing the high-dimensional multi-channel electrical feature matrix into four evolution states—normal operation, very early latency, accelerated degradation, and complete breakdown—accelerated according to the dynamic evolution trajectory of the time-varying transition resistance model, and performing auxiliary correction and state coding labeling on the stage division results based on the zero-sequence sensitive multi-physics coupling degradation index to construct a hybrid state dataset; constructing an early intelligent warning model based on a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a Bayesian optimization algorithm; training and optimizing the early intelligent warning model using the hybrid state dataset to obtain the optimal network model.

[0030] Example 2: An early warning method for inter-turn short-circuit faults in distribution transformers.

[0031] This embodiment provides an early warning method for inter-turn short-circuit faults in distribution transformers, including: real-time acquisition of multi-channel electrical quantities on the high and low voltage sides of the distribution transformer to be diagnosed and calculation of zero-sequence components; construction of a high-dimensional multi-channel electrical feature matrix to be diagnosed based on the acquired electrical quantities and zero-sequence components; inputting the high-dimensional multi-channel electrical feature matrix to be diagnosed into the optimal network model trained by the method described in Embodiment 1 for forward inference, mapping and classifying it to the corresponding evolution state, thereby realizing online monitoring and early intelligent warning of insulation degradation of the distribution transformer.

[0032] Figure 1 This paper presents the overall process of an early warning method for inter-turn short-circuit faults in distribution transformers based on multi-physics field features and hybrid deep learning. The following section combines... Figure 1 The above embodiments 1 and 2 will be described in detail.

[0033] Step 1: Collect multi-channel electrical quantities under actual operation scenarios of distribution transformers and calculate zero-sequence component data as field reference data; at the same time, acquire small sample experimental fault data, and use both as real sample data.

[0034] Multi-channel electrical quantities on the high-voltage and low-voltage sides of the distribution transformer are acquired in real time using voltage and current measuring devices. The measurement locations are at the high-voltage side input terminal and the low-voltage side output terminal of the transformer, respectively. The acquired electrical quantities specifically include the three-phase voltage, three-phase current, and three-phase voltage and current on the high-voltage side. By calculating the average value of the three-phase instantaneous sampled values ​​of the above three-phase electrical quantities in real time, the true zero-sequence voltage and zero-sequence current components, which characterize the local asymmetric damage features of the windings, are extracted. The formula for calculating the zero-sequence component is as follows:

[0035] in and These represent the instantaneous sampled values ​​of the three-phase voltage and three-phase current on the high-voltage side at time t, respectively. and These represent the instantaneous sampled values ​​of the three-phase voltage and three-phase current on the low-voltage side at time t, respectively.

[0036] The zero-sequence voltage on the high-voltage side calculated above Zero-sequence current on the high-voltage side Low-voltage side zero-sequence voltage and low-voltage side zero-sequence current This serves as the zero-order channel input to the high-dimensional multi-channel electrical feature matrix in step 3, and as the zero-order sensitivity factor in step 4. The computational basis.

[0037] Simultaneously, a transformer physical experimental platform was used to conduct simulation experiments via physical taps on the windings and an external adjustable resistor. To realistically reproduce the entire physical process of insulation degradation and calibrate subsequent transformer simulation models, an external adjustable resistor was used. Several typical high-resistance, medium-low-resistance, and low-resistance fault samples are provided as calibration data for the simulation model; the boundary of the continuous degradation stage is determined by the time-varying transition resistance model, multi-parameter trend inflection point, and zero-sequence sensitive multi-physics coupling degradation index in step 4.

[0038] In summary, the final collected real sample data includes information on the actual normal operation status of the transformer on site, as well as information on the very early and weak fault status obtained under experimental conditions.

[0039] Step 2: Using the real sample data output in Step 1 as a physical calibration reference, generate corresponding state-specific simulation supplementary data through the transformer simulation model.

[0040] The specific method for generating supplementary simulation data is as follows: A three-dimensional electromagnetic-thermal-hydrodynamic field-circuit cooperative model (hereinafter referred to as the simulation model) of the distribution transformer is constructed using multiphysics simulation software; the real sample data obtained in step 1 is extracted as a benchmark, and the boundary conditions and equivalent circuit parameters of the simulation model are iteratively calibrated in reverse to ensure that the transient electrical quantities output by the simulation model match the physical experimental quantities, thereby improving the high fidelity of the model; in the calibrated simulation model, a time-varying transition resistance model based on the Sigmoid function is introduced to characterize the nonlinear dynamic process of inter-turn insulation degradation. For example... Figure 2 As shown, to simulate the inter-turn insulation degradation and inter-turn short-circuit fault of a distribution transformer, a fault branch is set in a local part of the winding, and a time-varying transition resistor is introduced into the fault branch. Figure 2 In the diagram, A, B, and C are the three-phase terminals on the high-voltage side, a, b, and c are the three-phase terminals on the low-voltage side, and o is the neutral point; L1~L3 and R1~R3 are the inductance and resistance of the high-voltage side windings, and L4~L6 and R4~R6 are the inductance and resistance of the low-voltage side windings. This is a time-varying transition resistance. The time-varying transition resistance is used to characterize the dynamic process of inter-turn insulation evolving from a high-resistance latent state to a low-resistance breakdown state. By changing the fault branch connection location, the number of short-circuit turns, and the variation law of the transition resistance, the deterioration conditions of inter-turn insulation under different fault locations, different degrees of degradation, and different development rates can be simulated.

[0041] Time-varying transition resistance The mathematical evolution model is defined as follows:

[0042]

[0043]

[0044] In the formula, and These represent the high resistance value when the insulation is normal and the minimum resistance value when it is completely broken down, respectively. The shape parameter characterizing the degradation rate and Greater than 0, This is the time offset center for accelerated aging of insulation. Among them, Indicates the first The time-varying transition resistance values ​​corresponding to each time window are used as the main criterion for the four-stage division in step 4 and the multiphysics trend coupling factor. The calculation of the time-varying transition resistance. As a dynamic coupling variable, it is substituted into the control equation of the transformer winding.

[0045] By setting different levels of fault conditions through full traversal and adjusting the attenuation coefficient of the Sigmoid curve, the simulated three-phase voltage and current signals of the high and low voltage sides of the transformer under each weak deterioration stage are solved and output in a targeted and large-scale manner. Finally, using the same vector sum formula as in step 1, the corresponding simulated zero-sequence voltage and zero-sequence current components of the high and low voltage sides are extracted synchronously as supplementary simulation data.

[0046] Step 3: As Figure 3 , Figure 4 As shown, the real sample data obtained in step 1 and the simulation supplementary data generated in step 2 are fused and spliced ​​together according to a unified sampling frequency, a unified time window length and a unified channel arrangement to construct a high-dimensional multi-channel electrical feature matrix.

[0047] The real sample data obtained in step 1 and the simulation supplementary data generated in step 2 are truncated, aligned, and normalized according to the same sampling frequency, the same time window length, and the same channel order; let the first... Each time window contains For each sampling point, the multi-channel electrical quantities within that time window are constructed into a high-dimensional multi-channel electrical feature matrix:

[0048] in:

[0049] t The sampling point index is (t=1,2,…,T); , , They represent the first Within the first time window The high-voltage side three-phase voltage at each sampling point , , These represent the three-phase voltages on the low-voltage side. , , This indicates the three-phase current on the high-voltage side. , , These represent the three-phase currents on the low-voltage side. This is the zero-sequence voltage on the high-voltage side. This is the zero-sequence voltage on the low-voltage side. This is the zero-sequence current on the high-voltage side. This is the zero-sequence current on the low-voltage side.

[0050] To improve the adaptability of the network model input data in step 5 to the complex electromagnetic environment in the field, when constructing the high-dimensional multi-channel electrical feature matrix, Gaussian white noise with a preset signal-to-noise ratio (e.g., 20dB to 40dB) is superimposed on the simulation supplementary data to simulate random measurement noise, electromagnetic interference, and load disturbances in the actual operation of the distribution transformer. Simultaneously, a time-series shift enhancement method is used to expand the high-dimensional multi-channel electrical feature matrix. That is, while keeping the fault state label unchanged, the original time series is truncated with different time lengths to form multiple feature matrix sub-samples with the same fault state but different start times, thereby enhancing the model's adaptability to fault occurrence phase, sampling start point, and load fluctuations. The aforementioned high-dimensional multi-channel electrical feature matrix and its feature matrix sub-samples are collectively referred to as time-window electrical feature samples (hereinafter referred to as samples).

[0051] Step 4: According to the dynamic evolution trajectory of the time-varying transition resistance and the distortion law of multi-physics field, the high-dimensional multi-channel electrical feature matrix is ​​divided into four evolution stages: normal operation, very early latency, accelerated degradation and complete breakdown; and the stage division results are assisted and the state coding is labeled based on the zero-order sensitive multi-physics field coupling degradation index to construct a hybrid state dataset.

[0052] With time-varying transition resistance Using short-circuit circulating current, winding loss, zero-sequence component, and port current distortion characteristics as auxiliary criteria, the inter-turn insulation degradation process of distribution transformers is divided into four stages.

[0053] Specifically, during the construction phase of the real sample data and simulation supplementary data, a multi-parameter feature sequence under different transition resistance conditions is obtained. The multi-parameter features include one or more of the following: short-circuit circulating current, winding loss, high-voltage side input current, zero-sequence voltage, and zero-sequence current. The normalized deviation of each feature parameter relative to the normal operating state is calculated, and the rate of change of the normalized deviation with the logarithmic value of the transition resistance is further calculated. When the rate of change shows a continuous increase or abrupt change in a certain transition resistance range, the boundary corresponding to the transition resistance range is determined as the dividing point of the insulation degradation stage.

[0054] Figure 5 This paper presents a schematic diagram illustrating the four-stage division of insulation degradation based on time-varying transition resistance and multi-parameter trend inflection points. According to the multi-parameter trend inflection point analysis results, the inter-turn insulation degradation state of the distribution transformer is divided into four stages. When, it is determined to be in normal operating condition; when At that time, it was determined to be in a very early latent state; when When, it is judged to be in an accelerated deterioration state; when When the time is right, it is determined to be a complete breakdown state.

[0055] To improve the adaptability of state partitioning to different load conditions, different fault locations, and early weak asymmetric degradation characteristics, a zero-order sensitive multiphysics coupling degradation index is constructed. Used for time-varying transition resistance The obtained state partitioning results are then used for auxiliary correction. For the first... A time window, from the high-dimensional multi-channel electrical feature matrix obtained in step 3. Channels 13 through 16 are extracted as zero-sequence channels, and their zero-sequence energies are calculated. Simultaneously, channels 1 to 12 are extracted as the three-phase voltage and current channels on the high and low voltage sides, and the total energy at the ports is calculated. This leads to the zero-order sensitivity factor. :

[0056]

[0057]

[0058] in, The number of sampling points within the time window. t For the sampling point index (t=1,2,…,T), c For feature channel index, , These are the dimensionless balance coefficients for voltage and current, respectively. To prevent extremely small constants with a denominator of zero.

[0059] Calculate the multiphysics trend coupling factor based on one or more multiphysics characteristics, including short-circuit circulating current, winding loss, local temperature rise, and port current distortion. :

[0060]

[0061] in, Indicates the first The first time window One characteristic parameter, This represents the baseline value of the corresponding characteristic parameter under normal operating conditions. Indicates the first The weights of each feature parameter This indicates the number of feature parameters involved in the fusion. To prevent extremely small constants with a denominator of zero, , They represent the first k The feature parameter at the th in the th i Time, Number i -1 represents the normalized deviation (or relative deviation rate) relative to the normal state.

[0062] The multiphysics trend coupling factor Weights in Determined using the sensitivity-stability joint weighting method:

[0063] in, This is the fault sensitivity coefficient. The stability coefficient; For the first k The original weights of each feature parameter before normalization; To and The definitions are the same, for the first j The original weights of each feature parameter j To sum and iterate through the indices, j =1,2,… m ; for The mean of the first characteristic parameter under fault or deterioration conditions; This represents the mean of the corresponding characteristic parameters under normal operating conditions. This represents the standard deviation of the corresponding characteristic parameter under normal operating conditions. For the first The coefficient of variation of each feature parameter in the same state sample; To prevent extremely small constants with a denominator of zero.

[0064] Zero-order sensitivity factor Coupling factor with multiphysics trend By fusing the results, a zero-order sensitive multiphysics coupling degradation index is obtained. :

[0065] in, The fusion coefficient is... This is an interactive enhancement term used to improve the sensitivity of degradation characterization when zero-order asymmetric features and multi-physics degradation trends are enhanced simultaneously.

[0066] The zero-order sensitive multiphysics coupling degradation index obtained from the above calculations The specific auxiliary correction implementation rules are as follows: based on the time-varying transition resistance The initial state label of each time window sample is determined by the interval in which it is located, and the zero-order sensitive multiphysics coupling degradation index in samples at different degradation stages is used as the basis for further determination. The statistical distribution determines the threshold for determining adjacent states. When the coupling degradation index... When the mapped state matches the initial state, the initial state label remains unchanged; when they do not match, a review is performed considering the changing trends of short-circuit circulating current, winding losses, local temperature rise, port current distortion, and the zero-sequence sensitivity factor. If the coupling degradation index... If multiple consecutive time windows exceed the adjacent state determination threshold, and at least two auxiliary features show a continuous change trend consistent with the direction of insulation degradation, the initial state is corrected to the adjacent more severe degradation stage by one level. If the coupling degradation index does not continuously exceed the adjacent state determination threshold, and the auxiliary features do not show a continuous degradation trend, the original state label is maintained. A single correction does not cross two or more state levels to reduce the risk of miscorrection caused by transient disturbances and random noise.

[0067] Finally, the four states are encoded as follows: normal operation is encoded as 0, very early latent state is encoded as 1, accelerated degradation state is encoded as 2, and complete breakdown state is encoded as 3, resulting in a mixed state dataset with state labels and coupling degradation index:

[0068] in, Let i be the i-th high-dimensional multi-channel electrical feature matrix. Here, N represents the corresponding state encoding label, and N is the total number of samples.

[0069] Step 5: Construct an early intelligent warning model (referred to as the model or network model) based on hybrid deep learning; wherein, the first feature extraction network is used to extract local linkage features in the spatial domain and perform high-frequency noise reduction, the second feature extraction network is used to track long-distance temporal features in the transient and steady-state evolution process, and a hyperparameter optimization algorithm is introduced to adaptively optimize the model.

[0070] First, the high-dimensional multi-channel electrical feature matrix is ​​input into the first feature extraction network. A one-dimensional convolution kernel performs convolution operations along the time direction on the local waveform distortion, transient changes, and high-frequency disturbances of each electrical channel to extract local temporal features in the process of inter-turn insulation degradation of the distribution transformer. After the convolutional layer, an activation function and a pooling layer are set. The pooling layer compresses and reduces the dimensionality of the convolutional features to weaken the influence of random noise and redundant features on subsequent classification, thereby obtaining a locally denoised feature sequence.

[0071] Then, the feature sequence after local noise reduction is input into the second feature extraction network. The forward evolution information and backward correlation information of electrical parameters in the deterioration process are extracted through the forward and backward time propagation directions, respectively, so as to obtain long-distance time-dependent features that can characterize the entire process of the inter-turn insulation of the distribution transformer from normal, high resistance latency, medium and low resistance deterioration to low resistance breakdown.

[0072] Finally, the temporal features output by the second feature extraction network are input into the fully connected classification layer and the softmax output layer to obtain the probability values ​​of the sample to be diagnosed belonging to four states: normal operation, very early latency, accelerated degradation, and complete breakdown. P The state category with the highest probability value is then used as the model output.

[0073]

[0074]

[0075] in, This represents the probability of normal operation. This represents the probability of a very early latent state. This indicates the probability of accelerating the deterioration state. This represents the probability of a complete breakdown state.

[0076] like Figure 6 As shown, in a preferred embodiment, the model is a BOA-CNN-BiLSTM model, the first feature extraction network is a one-dimensional convolutional neural network (1D-CNN), the second feature extraction network is a bidirectional long short-term memory network, and the hyperparameter optimization algorithm is the Bayesian optimization algorithm (BOA). The Bayesian optimization algorithm is used to adaptively optimize the hyperparameters of the early intelligent warning model. The hyperparameters include, but are not limited to, the number of convolutional kernels, kernel size, pooling window size, number of hidden units in the one-dimensional convolutional neural network, number of nodes in the BiLSTM, learning rate, batch size, and regularization coefficient.

[0077] The Bayesian optimization algorithm uses the validation set classification accuracy, loss function value, or macro-average F1 value as the objective function. It establishes a surrogate model between the hyperparameter combination and the objective function through a Gaussian process, and selects the next set of hyperparameters to be evaluated based on the acquisition function. After multiple iterations, the optimal hyperparameter combination is obtained, thereby constructing the optimal early intelligent warning model.

[0078] It should be noted that the Bayesian optimization algorithm is only used for offline optimization of hyperparameters during the model training phase. After training is completed, the forward inference and state classification of the early intelligent warning model are performed only by a one-dimensional convolutional neural network and a bidirectional long short-term memory network. The Bayesian optimization algorithm does not participate in the online inference process.

[0079] Step 6: Divide the mixed-state dataset into a training set, a validation set, and a test set according to a preset ratio. Use the training set to train the network and perform parameter tuning through BOA. Save the trained optimal network model.

[0080] The mixed-state dataset obtained in step 4 is divided into training set, validation set and test set according to a preset ratio; the 1D-CNN-BiLSTM network is trained with parameters using the training set, the model performance under different hyperparameter combinations is evaluated using the validation set, and the optimal network structure and training parameters are determined by Bayesian optimization algorithm; when the model reaches the preset convergence condition or the maximum number of iterations, training is stopped and the optimal network model is saved.

[0081] During training, the cross-entropy loss function is used to measure the difference between the model's output state category and the true state label, and the network weights are updated through the backpropagation algorithm. At the same time, the test set is independently validated to obtain the model's identification results for normal operation, very early latency, accelerated degradation and complete breakdown states, so as to evaluate the model's generalization performance and early warning capability.

[0082] Table 1 Hyperparameter space and optimal parameter combinations

[0083] As shown in Table 1, the number of hidden neurons in BiLSTM is optimized to 154 for 16-channel input features, indicating that the model requires a large temporal memory capacity to extract long-distance evolution features from multi-channel electrical quantities.

[0084] Table 2 presents the recognition results under different channel combinations. As shown in Table 2, the overall recognition accuracy of the model is low when only single-phase or two-phase current is used. Introducing zero-sequence features significantly improves the recognition accuracy, indicating that zero-sequence components can effectively enhance the model's sensitivity to local asymmetric degradation of the winding. The 16-channel full-quantity features also exhibit high recognition performance, demonstrating that multi-channel information fusion is beneficial for improving the model's generalization ability to complex operating conditions.

[0085] Table 2 Recognition performance under different feature channel combinations

[0086] Table 3 shows the recognition results of different models under the same sample conditions. In this specific embodiment, the BOA-CNN-BiLSTM model has a higher overall accuracy and macro-average F1 score compared to support vector machines, random forests, one-way LSTM, CNN-BiLSTM, and random search CNN-BiLSTM models, indicating that the joint modeling method of 1D-CNN, BiLSTM, and BOA can improve the ability to identify early deterioration states.

[0087] Table 3 Comparison of results from different algorithms

[0088] Step 7: As Figure 7 As shown, the actual multi-channel electrical quantities of the distribution transformer to be diagnosed are collected in real time. After performing the same data preprocessing as in step 3, the data is input into the optimal network model for forward reasoning and mapped to the corresponding evolution stage, thereby realizing online monitoring and early intelligent warning of insulation degradation of the distribution transformer.

[0089] During the operation of the distribution transformer to be diagnosed, the three-phase voltage and three-phase current signals of its high and low voltage sides are collected in real time, and the zero-sequence voltage and zero-sequence current of the high and low voltage sides are obtained in real time according to the zero-sequence component calculation method in step 1. Subsequently, the high-dimensional multi-channel electrical feature matrix to be diagnosed is constructed according to the channel arrangement, time window length and normalization rule in step 3, and it is input into the optimal network model saved in step 6 for forward inference.

[0090] The optimal network model outputs the probability values ​​of the distribution transformer under diagnosis in each evolution stage. When the output result is in normal operation, no warning is triggered; when the output result is in an extremely early latent state, a level one warning signal is issued, indicating the risk of early deterioration of inter-turn insulation; when the output result is in an accelerated deterioration state, a level two warning signal is issued, prompting maintenance personnel to arrange key inspections or planned maintenance; when the output result is in a complete breakdown state, a level three alarm signal is issued, indicating that the equipment has a serious risk of inter-turn short circuit fault and recommending shutdown protection or emergency maintenance measures.

[0091] To reduce the risk of false alarms caused by occasional noise, the online early warning process also includes a continuous time window criterion: when the model output results of multiple consecutive time windows (e.g., 3 to 5) all point to the same abnormal state, or the probability of the abnormal state exceeds a preset threshold (e.g., 0.7 to 0.9), the corresponding level of early warning signal is confirmed to be triggered; otherwise, the monitoring state continues.

[0092] Example 3: An electronic device. This electronic device includes a processor and a memory. The memory and processor can be interconnected via a bus system and / or other forms of connection. The processor may include any one or more processors such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP). The memory may include volatile memory, such as random access memory (RAM). The memory may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD). The memory stores executable program code, which the processor executes to implement the steps of the model training method described in Example 1, or the steps of the early warning method described in Example 2.

[0093] Example 4: A computer-readable storage medium. For example, the computer-readable storage medium may be a read-only memory (ROM), random access memory (RAM), compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc. This computer-readable storage medium is used to store computer-readable instructions, which, when executed by a computer, can implement the steps of the model training method described in Example 1, or the steps of the early warning method described in Example 2.

[0094] This invention enables online monitoring and early intelligent warning of the entire process of inter-turn insulation degradation without the need for additional complex invasive sensors, utilizing conventional voltage and current measurement data and zero-sequence characteristics of the high and low voltage sides of the distribution transformer.

Claims

1. A training method for an early warning model of inter-turn short-circuit faults in distribution transformers, characterized in that, include: Acquire real sample data, which includes the zero-sequence component of multi-channel electrical quantity calculation under the actual operation scenario of distribution transformer and experimental fault data; Using the real sample data as a physical calibration reference, supplementary simulation data is generated by introducing a transformer simulation model based on a time-varying transition resistance model with a Sigmoid function. The real sample data and the simulation supplementary data are fused together to construct a high-dimensional multi-channel electrical feature matrix; According to the dynamic evolution trajectory of the time-varying transition resistance model, the high-dimensional multi-channel electrical feature matrix is ​​divided into four evolution states: normal operation, very early latency, accelerated degradation, and complete breakdown. Based on the coupling degradation index of zero-order sensitive multiphysics field, the stage division results are assisted in correction and state coding and labeling to construct a hybrid state dataset. An early intelligent warning model based on a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a Bayesian optimization algorithm is constructed. The early intelligent warning model is trained and optimized using the mixed state dataset to obtain the optimal network model.

2. The training method for an early warning model of inter-turn short-circuit faults in distribution transformers according to claim 1, characterized in that, The time-varying transition resistance model is defined as follows: In the formula, Indicates time-varying transition resistance. and These represent the high resistance value when the insulation is normal and the minimum resistance value when it is completely broken down, respectively. The shape parameter characterizing the degradation rate and Greater than 0, The time offset center for accelerated aging of insulation, t The time-varying transition resistance model is used as a dynamic coupling variable and substituted into the control equation of the distribution transformer winding to generate simulated high and low voltage three-phase voltage and current signals under each weak degradation stage, and the corresponding simulated zero-sequence component is extracted simultaneously as simulation supplementary data.

3. The training method for an early warning model of inter-turn short-circuit faults in distribution transformers according to claim 1, characterized in that, The channels of the high-dimensional multi-channel electrical feature matrix include high-voltage side three-phase voltage, high-voltage side three-phase current, low-voltage side three-phase voltage, low-voltage side three-phase current, high-voltage side zero-sequence voltage, high-voltage side zero-sequence current, low-voltage side zero-sequence voltage, and low-voltage side zero-sequence current. When constructing the high-dimensional multi-channel electrical feature matrix, Gaussian white noise with a preset signal-to-noise ratio is superimposed on the simulation supplementary data, and the high-dimensional multi-channel electrical feature matrix is ​​expanded by a time-series translation enhancement method to form feature matrix sub-samples with the same fault state but different start times.

4. The training method for an early warning model of inter-turn short-circuit faults in distribution transformers according to claim 2, characterized in that, The criteria for dividing the stages according to the dynamic evolution trajectory of the time-varying transition resistance model are as follows: when When, it is determined to be in normal operating condition; when At that time, it was determined to be in a very early latent state; when When, it is judged to be in an accelerated deterioration state; when When the time is right, it is determined to be a complete breakdown state.

5. The training method for an early warning model of inter-turn short-circuit faults in distribution transformers according to claim 1, characterized in that, The rules for auxiliary correction and state coding labeling include: The initial state labels of the high-dimensional multi-channel electrical feature matrices corresponding to each time window are determined based on the interval of the time-varying transition resistance model, and the adjacent state judgment thresholds are determined based on the statistical distribution of the coupling degradation index of the zero-order sensitive multi-physics field in the high-dimensional multi-channel electrical feature matrices corresponding to different degradation stages. When the state mapped by the coupling degradation index is consistent with the initial state, the initial state label remains unchanged; when they are inconsistent, the verification is performed in conjunction with auxiliary features, which include the changing trends of short-circuit circulating current, winding loss, local temperature rise, port current distortion, and zero-sequence sensitivity factor. If the coupling degradation index exceeds the adjacent state judgment threshold for multiple consecutive time windows, and at least two of the auxiliary features show a continuous change trend consistent with the direction of insulation degradation, then the initial state will be corrected to the adjacent more severe degradation state by one level. If the coupling degradation index does not continuously exceed the adjacent state determination threshold, and the auxiliary feature does not show a continuous degradation trend, then the original state label is maintained. A single calibration should not span two or more status levels.

6. The training method for an early warning model of inter-turn short-circuit faults in distribution transformers according to claim 1, characterized in that, The coupling degradation index is calculated as follows: Zero-sequence energy is calculated by extracting zero-sequence channels from the high-dimensional multi-channel electrical feature matrix, and the total port energy is calculated using the three-phase voltage and current channels on the high and low voltage sides. The zero-sequence sensitivity factor is determined based on the ratio of zero-sequence energy to the total port energy. The multiphysics trend coupling factor is calculated based on one or more characteristic parameters among short-circuit circulating current, winding loss, local temperature rise, and port current distortion. The zero-order sensitivity factor is fused with the multiphysics trend coupling factor to obtain the coupling degradation index.

7. The training method for an early warning model of inter-turn short-circuit faults in distribution transformers according to claim 1, characterized in that, In the aforementioned early intelligent warning model, a one-dimensional convolutional neural network is used to perform convolution operations on the local waveform distortion and high-frequency disturbances of each electrical channel and to reduce dimensionality and noise; a bidirectional long short-term memory network is used to extract long-distance temporal dependence features through the forward and reverse temporal propagation directions. The Bayesian optimization algorithm uses the validation set classification accuracy, loss function value, or macro-average F1 value as the objective function, and adaptively optimizes the hyperparameters of the early intelligent warning model through a Gaussian process.

8. A method for early warning of inter-turn short-circuit faults in distribution transformers, characterized in that, include: Real-time acquisition of multi-channel electrical quantities on the high and low voltage sides of the distribution transformer under diagnosis and calculation of zero-sequence components; A high-dimensional multi-channel electrical feature matrix to be diagnosed is constructed based on the collected electrical quantities and zero-sequence components; The high-dimensional multi-channel electrical feature matrix to be diagnosed is input into the optimal network model trained by the method described in any one of claims 1 to 7 for forward inference, and mapped and classified to the corresponding evolution state, so as to realize online monitoring and early intelligent warning of insulation degradation of distribution transformer.

9. The early warning method for inter-turn short-circuit faults in distribution transformers according to claim 8, characterized in that, Early intelligent warning includes: no warning is triggered when the output result of the optimal network model is in normal operation; a level 1 warning signal is issued when the output result is in an extremely early latent state; a level 2 warning signal is issued when the output result is in an accelerated degradation state; a level 3 warning signal is issued when the output result is in a complete breakdown state; when the output results of the optimal network model in multiple consecutive time windows all point to the same abnormal state, or when the probability of the abnormal state exceeds a preset threshold, the corresponding level of warning signal is confirmed to be triggered; otherwise, the monitoring state continues.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the model training method as described in any one of claims 1 to 7, or to implement the steps of the early warning method as described in any one of claims 8 to 9.

Images