A distribution network fault identification method and system based on transient morphological features

Abstract

The application discloses a distribution network fault identification method and system based on transient state feature, and belongs to the technical field of power system automation and intelligent operation and maintenance. The method obtains distribution network transient recording data and extracts multi-dimensional features, and constructs a state description vector mapping the underlying physical discharge state. The state description vector is combined with the multi-dimensional features to construct a mechanism collaborative feature subset, and redundant features in a random discrete state are removed in combination with distribution network grounding operation parameters. The remaining subset is projected into a transient latent variable space, and a state membership degree vector of the subset and a preset physical state anchor point is calculated. The state membership degree vector is converted into a control mask to inject a deep neural network, and a joint loss function is used to constrain the network hidden layer feature mapping result to fit the corresponding physical state anchor point, so that an initial classification is output. Finally, the feature marginal contribution degree of an inference link is calculated, and a final diagnosis result is output. The application realizes deep integration of physical mechanism and artificial intelligence, and solves the problem of lack of physical basis of a black box AI model.

Description

Technical Field

[0001] This invention belongs to the field of power system automation and intelligent operation and maintenance technology, specifically relating to a method and system for identifying distribution network faults based on transient morphological characteristics. Background Technology

[0002] With the continuous advancement of the construction of new power systems, distribution networks are gradually transforming into active AC / DC hybrid power grids, and their network topologies are becoming increasingly complex. The safe and stable operation of distribution networks places higher demands on rapid fault detection and accurate diagnosis. However, the causes of distribution network faults are complex and diverse, especially hidden faults caused by tree-line conflicts, bird nest-like connections, and connections involving foreign plastic objects. These faults are characterized by short transient processes and significant high-resistance characteristics, making them highly susceptible to being missed or misdiagnosed by distribution network automation systems.

[0003] Currently, traditional fault diagnosis methods for power distribution networks mainly rely on steady-state power frequency signals (such as overcurrent and zero-sequence current) for threshold discrimination. This type of method is effective in identifying metallic short-circuit faults, but when faced with complex transient faults, it is difficult to extract high-discrimination feature parameters because it cannot effectively characterize the underlying physical features such as transient high-frequency disturbances and intermittent discharge.

[0004] In recent years, with the widespread adoption of smart power distribution terminals and the massive accumulation of transient waveform data, existing technologies have begun to employ deep learning to achieve end-to-end classification and identification of fault transient waveforms. However, purely data-driven diagnostic solutions have significant shortcomings in engineering implementation: First, the model does not incorporate multi-physics coupling and prior constraints of operating state. The feature processing stage blindly receives all multi-dimensional data, which easily misjudges random environmental interference and unstable high-order harmonics with severe attenuation as key features, resulting in feature redundancy, severe data drift under different grounding systems, and poor model robustness.

[0005] Secondly, the "black box" nature of deep learning contradicts the high reliability requirements of power systems. The evolution of intermediate-layer features in existing network architectures lacks physical constraints, and the output relies solely on mathematical probability mappings, failing to provide the distribution network automation dispatching end with a chain of evidence confirming the consistency of the underlying discharge physical form. This disconnect between the reasoning process and physical laws leads to a lack of logical confidence in the results received by the control system, easily causing protection malfunctions and hindering the support of high-order automation decision-making loops in the distribution network.

[0006] In summary, existing distribution network diagnostic systems urgently need a fault identification method that can deeply integrate the underlying physical discharge laws with the deep network reasoning process, and realize dynamic feature optimization and reasoning trajectory mechanism tracing in the back-end control link. Summary of the Invention

[0007] This invention aims to solve the technical problems of existing pure data-driven power distribution network fault diagnosis systems, which are susceptible to interference from redundant random features due to the lack of underlying physical mechanism constraints, and the inability to interpret the black-box reasoning process of the model, thus failing to output a physical evidence chain consistent with the underlying discharge law of the power grid.

[0008] This invention discloses a method and system for distribution network fault identification based on transient morphological features. The method acquires transient waveform data of the distribution network and extracts multi-dimensional features to construct a morphological description vector mapping the underlying physical discharge state. This vector is then combined with the multi-dimensional features to construct a mechanism-coordinated feature subset. Redundant features exhibiting random discrete states are eliminated by combining distribution network grounding operation parameters. The retained subset is projected into the transient latent variable space, and its state dependency vector with a preset physical state anchor point is calculated. This state dependency vector is transformed into a control mask and injected into a deep neural network. A joint loss function constrains the hidden layer feature mapping results of the network to fit the corresponding physical state anchor point, outputting an initial classification. Finally, the feature marginal contribution of the inference link is calculated, and the consistency with the physical state anchor point is verified to output the final diagnostic result. This invention achieves a deep integration of physical mechanisms and artificial intelligence, solving the problem of black-box AI models lacking physical basis.

[0009] The present invention adopts the following technical solution.

[0010] In a first aspect, the present invention provides a method for identifying distribution network faults based on transient morphological characteristics, comprising: Acquire transient waveform data of the distribution network, extract the basic multidimensional features of the transient waveform data, and construct a transient discharge morphology description vector based on the transient waveform data; The transient discharge morphology description vector is combined with the basic multidimensional features to construct a mechanism collaborative feature subset. The reliability is evaluated by combining the grounding operation parameters of the distribution network, and redundant feature subsets that are in a random discrete state are eliminated. The retained mechanism-cooperative feature subset is projected into the transient latent variable space, and the state membership vector between the current mapped coordinates and multiple preset physical state anchor points is calculated. The state is transformed from a membership vector into a control mask and injected into a deep neural network. Based on a joint loss function that includes a penalty term for deviation from the physical state, the intermediate layer feature representation of the deep neural network is constrained to be consistent with the corresponding physical state anchor point, and an initial diagnostic classification is output. Calculate the marginal contribution of each mechanism collaborative feature subset during network inference; for the initial diagnostic classification, extract the mechanism collaborative feature subset with the highest marginal contribution, and verify whether the physical state anchor point corresponding to the mechanism collaborative feature subset is consistent with the initial diagnostic classification; if consistent, output the final diagnostic result.

[0011] Preferably, the transient waveform data of the distribution network is acquired, the basic multidimensional features of the transient waveform data are extracted, and a transient discharge morphology description vector is constructed based on the transient waveform data, specifically including: The transient waveform data includes sequences of zero-sequence current, three-phase current, three-phase voltage, and electric field changes; the basic multidimensional features include time-domain features, frequency-domain features, and time-frequency-domain features. Extract the waveform range from the transient waveform data that exceeds a preset threshold to recover below the preset threshold, and mark it as a pulse single peak; extract the time span of the pulse single peak to generate a morphological component characterizing the instantaneous conduction duration; The frequency of changes in local extreme points of the transient waveform data per unit time is statistically analyzed to generate morphological components that characterize the density of local high-frequency disturbances. The time interval between adjacent pulse peaks is extracted to generate a morphological component characterizing the periodic stability of intermittent discharge. Calculate the difference between the three-phase voltage and zero-sequence current within a preset period before and after the fault transient, and generate a morphological component characterizing the self-recovery capability of the insulation. The transient discharge morphology description vector is constructed by combining and encoding each of the morphological components according to the correlation of discharge behavior.

[0012] Preferably, the transient discharge morphology description vector is combined with the basic multidimensional features to construct a mechanism-coordinated feature subset. This subset is then used for reliability assessment in conjunction with the grounding operation parameters of the distribution network. Redundant feature subsets exhibiting random discrete states are eliminated. Specifically, this includes: Based on the coupling relationship between current, voltage and electric field during transient discharge, the morphological components in the transient discharge morphology description vector are combined with the basic multidimensional features that have physical correlation to generate multiple mechanism synergistic feature subsets. Obtain the current grounding operation parameters of the distribution network, and determine the characteristic physical constraint boundary of the corresponding grounding system based on the grounding operation parameters; Calculate the distribution difference index of each of the aforementioned mechanism collaborative feature subsets when distinguishing different fault types, and the information entropy index characterizing the degree of feature fluctuation dispersion; By integrating the physical constraint boundary of the features, the distribution difference index, and the information entropy index, the mechanism collaboration credibility score of each of the mechanism collaboration feature subsets is calculated; The subset of mechanism-cooperation features whose reliability score is lower than a preset evaluation threshold is determined as a redundant subset of features in a random discrete state.

[0013] Preferably, the retained subset of mechanistic collaborative features is projected onto the transient latent variable space, and the state dependency vector between the current mapped coordinates and multiple preset physical state anchor points is calculated, specifically including: A manifold mapping method based on neighborhood structure constraints is adopted to map the retained mechanism collaborative feature subset to the transient latent variable space and obtain the corresponding current mapping coordinates; In the transient latent variable space, multiple physical state anchor points are set according to the fault prior mechanism; the physical state anchor points include at least a first anchor point characterizing the insulation fast recovery constraint state, a second anchor point characterizing the unstable lap connection conduction state, and a third anchor point characterizing the micro-gap dynamic disturbance state. Calculate the cooperative distribution distance based on the mean and covariance distributions of the current mapped coordinates and each physical state anchor point; Based on the normalized result of the cooperative distribution distance, and combined with the covariance distribution characteristics contained in each of the physical state anchor points, the state membership vector of each of the physical state anchor points corresponding to the mechanism cooperative feature subset is calculated.

[0014] Preferably, the state from the membership vector is transformed into a control mask and injected into the deep neural network. Based on a joint loss function including a penalty term for deviation from the physical state, the intermediate layer feature representation of the deep neural network is constrained to be consistent with the corresponding physical state anchor point, and an initial diagnostic classification is output, specifically including: Convert the state from the membership vector into a control mask corresponding to the feature dimension; The control mask is injected into the deep neural network, and the feature extraction weights of the deep neural network are adjusted using the control mask. Calculate the state difference value between the intermediate layer feature representation and the corresponding physical state anchor point, and use the state difference value as the physical state deviation penalty term; Calculate the classification prediction error of the deep neural network, and combine the classification prediction error with the physical state deviation penalty term to construct the joint loss function; The model parameters of the deep neural network are updated based on the joint loss function to constrain the intermediate layer feature representation to converge to the physical consistency range defined by the corresponding physical state anchor point. The initial diagnostic classification is output through a deep neural network with updated parameters.

[0015] Preferably, the marginal contribution of each mechanism collaborative feature subset during network inference is calculated; for the initial diagnostic classification, the mechanism collaborative feature subset with the highest marginal contribution is extracted, and the physical state anchor point corresponding to the mechanism collaborative feature subset is verified to be consistent with the initial diagnostic classification; if consistent, the final diagnostic result is output, specifically including: Obtain the intermediate layer feature representations when the deep neural network outputs the initial diagnostic classification; The feature attribution algorithm is invoked, and the marginal contribution of each of the mechanism-coordinated feature subsets to the generation of the initial diagnostic classification is calculated in combination with the intermediate layer feature representation. Extract the mechanism collaborative feature subset with the highest marginal contribution; obtain the diagnostic state anchor point corresponding to the initial diagnostic classification; determine the attribution state anchor point corresponding to the extracted mechanism collaborative feature subset in the transient latent variable space; Compare whether the attribution state anchor point and the diagnostic state anchor point are the same physical state anchor point; if so, output the initial diagnostic classification as the final diagnostic result.

[0016] Preferably, the feature attribution algorithm is invoked, and combined with the intermediate layer feature representation, the marginal contribution of each of the mechanism-coordinated feature subsets to the generation of the initial diagnostic classification is calculated, specifically including: A feature attribution algorithm based on a game-theoretic alliance mechanism is used to calculate the marginal probability difference of each of the aforementioned mechanism collaborative feature subsets to the initial diagnostic classification, and the result after permutation and combination weighted integration is used as the marginal contribution. Based on the extracted mechanism-coordinated feature subset with the largest marginal contribution value, the key waveform intervals in the original transient waveform data are determined; The final diagnostic results and the key waveform intervals are encapsulated into structured fault evidence chain data and sent to the distribution network automation dispatch and control terminal.

[0017] Secondly, the present invention provides a distribution network fault identification system based on transient morphological characteristics, which operates the distribution network fault identification method based on transient morphological characteristics as described above, including: The transient data acquisition and morphology construction module is used to acquire transient waveform data of the distribution network, extract the basic multidimensional features of the transient waveform data, and construct a transient discharge morphology description vector based on the transient waveform data. The feature synergy combination and credibility assessment module is used to combine the transient discharge morphology description vector with the basic multidimensional features to construct a mechanism synergy feature subset, and combine it with the grounding operation parameters of the distribution network to conduct credibility assessment and eliminate redundant feature subsets that are in a random discrete state. The latent variable space mapping and state anchoring module is used to project the retained mechanism collaborative feature subset onto the transient latent variable space and calculate the state membership vector between the current mapped coordinates and multiple preset physical state anchor points. The physical mask injection and constrained reasoning module is used to transform the state membership vector into a control mask for injection into the deep neural network, and based on a joint loss function including a physical state deviation penalty term, constrain the intermediate layer feature representation of the deep neural network to be consistent with the corresponding physical state anchor point, and output the initial diagnostic classification. The feature attribution and physical consistency verification module is used to calculate the marginal contribution of each mechanism collaborative feature subset during network inference; for the initial diagnostic classification, the mechanism collaborative feature subset with the highest marginal contribution is extracted, and the physical state anchor point corresponding to the mechanism collaborative feature subset is verified to be consistent with the initial diagnostic classification; if consistent, the final diagnostic result is output.

[0018] Thirdly, the present invention provides a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method as described.

[0019] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method.

[0020] The beneficial effects of this invention are compared with those of the prior art: 1. This invention overcomes the shortcomings of traditional pure data-driven dimensionality reduction, which tends to retain high variance noise, by constructing a transient discharge morphology description vector containing multi-dimensional morphological components and generating a mechanism-coordinated feature subset based on physical coupling relationships. Furthermore, by combining the physical constraint boundaries set by the distribution network grounding operation parameters for credibility assessment, redundant features that violate the limits of the current operating conditions can be effectively filtered out in the early stages of feature selection. This improves the model's feature adaptability and robustness under complex environmental background noise and power grid grounding mode switching conditions.

[0021] 2. This invention maps transient features to a latent variable space with preset physical state anchors and transforms the calculated state membership degree into a control mask injected into the hidden feature layer of the deep neural network. Simultaneously, a joint loss function including a physical state deviation penalty term is constructed during backpropagation. This mechanism introduces prior physical boundary guidance into the data optimization process of the deep learning model, limiting the deviation of the intermediate layer feature evolution trajectory from objective discharge laws, effectively reducing the probability of misjudgment due to fitting spurious correlation features, and improving the engineering reliability of fault identification results.

[0022] 3. This invention utilizes a feature attribution algorithm to inversely calculate the marginal contribution of each mechanism-coordinated feature subset, and achieves physical mechanism verification of the network inference trajectory by bidirectionally comparing whether the attribution state anchor points of high-contribution features are consistent with the diagnostic state anchor points of the initial diagnosis. This verification mechanism enables the diagnostic system to output a structured fault evidence chain including key waveform intervals, thereby providing the distribution network automation dispatching terminal with a decision-making basis supported by transparent mechanisms, enhancing the dispatching confidence and proactive defense efficiency of the intelligent operation and maintenance system. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of a distribution network fault identification method based on transient morphological characteristics provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0025] Example 1: like Figure 1 As shown, this invention provides a method for identifying distribution network faults based on transient morphological characteristics, specifically including: Step 1: Obtain transient waveform data of the distribution network, extract the basic multidimensional features of the transient waveform data, and construct a transient discharge morphology description vector based on the transient waveform data.

[0026] The specific implementation process of step 1 is as follows: Step 1.1: Acquire transient waveform data of the distribution network The status of distribution network line nodes is monitored in real time. When a sudden change in the electric field strength between the line and ground is detected, such as when the rate of change of the electric field exceeds a preset threshold, transient waveform recording is automatically triggered. A specific time window containing the steady-state state before the fault and the transient process after the fault is extracted as analysis data. This specific time window is, for example, from one cycle before the fault to three cycles after the fault. The acquired transient waveform data is a synchronously sampled time series, specifically including: zero-sequence current sequence, three-phase current sequence, three-phase voltage sequence, and electric field change sequence.

[0027] Step 1.2: Extract the basic multidimensional features of the transient waveform data. The acquired transient waveform data undergoes signal preprocessing, such as denoising and filtering to remove bias. Subsequently, conventional signal analysis algorithms are used to extract fundamental multidimensional features to provide multidimensional numerical references. Specifically, this includes: Time-domain characteristics: Extract the effective value, wavefront steepness, peak absolute value, and decay time constant of the transient zero-sequence current; Frequency domain characteristics: Fast Fourier Transform is used to extract the integral energy, fundamental amplitude, and proportion of each harmonic in the high-frequency band; Time-frequency domain features: Discrete wavelet transform or short-time Fourier transform are used to extract multi-scale energy distribution features and wavelet energy entropy.

[0028] Step 1.3: Construct the transient discharge morphology description vector Traditional frequency domain or time-frequency domain statistical features are insufficient to characterize the dynamic mechanism evolution relationship between conduction, disturbance, and insulation recovery during foreign object discharge. Therefore, based on the time-series waveforms in transient waveform recordings, waveform morphology statistical analysis methods are used to quantify the underlying physical discharge behavior, constructing a transient discharge morphology description vector containing multiple morphological components. This step overcomes the deficiency of traditional black-box model feature inputs lacking physical meaning, giving the machine the ability to read the underlying discharge dynamic evolution law in the feature space. The specific quantization and encoding process is as follows: (1) Generating a morphological component characterizing the duration of instantaneous conduction: For the zero-sequence current or fault phase current waveform, an effective discharge preset threshold is set. The effective discharge preset threshold is adaptively determined based on a preset multiple of the root mean square value of the steady-state background noise. Traverse the waveform data, extract the complete waveform interval from when the waveform exceeds the preset threshold until it decays back to below the preset threshold, and mark this interval as a pulse peak. Calculate the time span of the pulse peak, for example, how many milliseconds it lasts. This morphological component is used to characterize the extremely short duration of instantaneous conduction after foreign objects (such as wooden branches or bird nests) overlap and are then destroyed by arc erosion.

[0029] (2) Generating a morphological component characterizing the density of local high-frequency disturbances: Using a sliding window with a preset step size, the first derivative of the transient waveform data is calculated, and the number of times the sign of the first derivative flips within a unit time window is counted, i.e., the frequency of change of local extreme points is identified. The higher the frequency, the denser the waveform spikes. This morphological component is used to characterize the physical phenomenon of micro-vibration and dense discharge spikes generated by metal wires or plastic foreign objects under the action of inductive force.

[0030] (3) Generating a morphological component characterizing the periodic stability of intermittent discharge: If multiple conductions occur during the transient process, extract the time interval between the starting points of two adjacent pulse peaks. Calculate the mean and variance of this time interval. If the interval is usually greater than one cycle and the variance is small, the value of this morphological component increases, which reflects the intermittent repetitive discharge characteristics caused by the swaying of branches in the wind or the inability of the arc to completely destroy the overlapping relationship in the tree line contradiction.

[0031] (4) Generate the morphological component characterizing the self-recovery capability of insulation: Calculate the effective values ​​of the three-phase voltage and zero-sequence current during the steady-state period before the fault (e.g., the first cycle before the fault) and the later transient period after the fault (e.g., the third cycle after the fault), and obtain the difference between the two. When the difference value approaches zero, it indicates that the insulation state has been rapidly restored. This component is mainly used to characterize the physical mechanism of the rapid recovery of high-resistance insulation state after the surge arrester operates to discharge overvoltage.

[0032] Step 1.4: Combination Encoding The extracted morphological components of the four dimensions are normalized and then cascaded and encoded into a structured vector according to the timing and causal relationship of the fault discharge, i.e., conduction, disturbance, intermittent, and recovery. For example, this vector is denoted as vector. This refers to the transient discharge morphology description vector. , , and These represent the instantaneous conduction duration morphological component, the local high-frequency disturbance density morphological component, the intermittent discharge periodic stability morphological component, and the insulation self-recovery capability morphological component, respectively. This vector will directly serve as the core physical state characterization data for subsequent characterization of the underlying physical constraints.

[0033] Step 2: Combine the transient discharge morphology description vector with the basic multidimensional features to construct a mechanism collaborative feature subset, and combine it with the grounding operation parameters of the distribution network to conduct a credibility assessment, and eliminate redundant feature subsets that are in a random discrete state.

[0034] Step 2 aims to overcome the shortcomings of traditional machine learning that treats features as isolated variables by filtering noise through multi-physics coupling with the real power grid operation boundary. The specific implementation process is as follows: Step 2.1: Combining and generating collaborative feature subsets of the mechanism Instead of evaluating the effectiveness of a single feature in isolation, the system constructs a joint feature vector based on the physical causal and coupling relationships between current, voltage, and electric field during transient discharge. This is achieved by structurally combining the morphological components of the transient discharge morphology description vector output in step 1 with fundamental multidimensional features that have physical relationships. Specifically, the system concatenates and splices physically related morphological components with fundamental multidimensional features to generate multiple mechanism-coordinated feature subsets. ,in , N The total number of subsets generated by the combination.

[0035] For example, the high-frequency integrated energy in the basic multidimensional features is extracted and combined with the morphological component in the morphological description vector that represents the local high-frequency perturbation density to construct a metal perturbation-type cooperative feature subset that maps the vibration discharge phenomenon of metal or plastic foreign objects; the effective value of zero-sequence current in the basic multidimensional features is extracted and combined with the morphological component that represents the periodic stability of intermittent discharge to construct a tree-line type cooperative feature subset that maps tree-line contradictions.

[0036] Step 2.2: Determine the characteristic physical constraint boundaries of the corresponding grounding system The absolute values ​​of electrical quantities in a distribution network vary greatly under different grounding methods, necessitating the introduction of prior constraints on operating conditions. This involves obtaining the current grounding operating parameters of the distribution network and determining the characteristic physical constraint boundaries of the corresponding grounding system based on these parameters.

[0037] Specifically, the system automatically acquires the current grounding operation parameters of the distribution network, such as identifying whether the current line is grounded using an arc suppression coil or a low-resistance grounding method. Based on these grounding operation parameters, a preset rule base is invoked to determine characteristic physical constraint boundaries, and a physical boundary violation judgment function is defined. When identified as a low-current grounding system, a physical upper limit boundary for the effective value of the transient zero-sequence current is set, based on the current rated operating parameters of the grounding system; when the feature subset When the relevant electrical quantities exceed the physical boundary, set ,otherwise .

[0038] To address the diagnostic model failure caused by sudden changes in zero-sequence electrical quantities due to switching grounding methods in distribution networks, this invention introduces distribution network grounding operation parameters as strong physical boundary constraints. By applying a large physical boundary violation penalty factor, redundant features that violate the electrical limits of the current operating system are forcibly eliminated, enabling the model to possess resistance to environmental interference and adaptability to operating conditions.

[0039] Step 2.3: Calculate the distribution difference index and information entropy index To quantify the mathematical reliability of each feature subset, the following calculations are performed using a historical fault sample database: Calculate the distribution difference index of each of the aforementioned mechanism co-feature subsets when distinguishing different fault types. The variance is quantified by the ratio of between-class variance to within-class variance, as shown in the following formula:

[0040] In the formula, This represents the total number of fault types. For the first The number of fault samples, For this type of sample in the feature subset The mean center below, The population mean of all samples. This represents the feature value of a specific sample. The larger this indicator is, the more effective it is in distinguishing different fault types.

[0041] Calculate the information entropy index that characterizes the dispersion of characteristic fluctuations. This is used to quantify the fluctuation of this feature subset under similar fault conditions, and the formula is as follows:

[0042] In the formula, the continuous numerical space of the eigenvalue fluctuation is equally spaced as follows: Each interval The number of intervals into which the eigenvalue space is divided. For the eigenvalue to fall within the th The probability of each interval. The larger this index is, the more likely the feature is to exhibit an approximately uniform random distribution and be highly susceptible to environmental noise interference.

[0043] Step 2.4: Computer-based collaborative credibility score By integrating the physical constraint boundary of the features, the distribution difference index, and the information entropy index, the mechanism collaboration credibility score of each of the mechanism collaboration feature subsets is calculated. .

[0044] The specific fusion calculation formula is as follows:

[0045] In the formula, the distribution difference index is used as a positive gain term, with a weight of . The information entropy index is used as a negative penalty term, with a weight of . If the characteristic distribution goes out of bounds, a physical deviation penalty constant is applied. (in This allows the physical boundary violation item to play a dominant suppressive role in the credibility score. Among them, and Based on a historical fault sample set with expert labels, the faults are determined through grid search combined with cross-validation, and satisfy the following conditions: The preferred numerical range is , .

[0046] Step 2.5: Remove redundant feature subsets that are in a random discrete state. Set preset evaluation threshold Compare the mechanistic synergy credibility scores of each mechanistic synergy feature subset. With the preset evaluation threshold The mechanism's collaborative credibility score is lower than a preset evaluation threshold. The mechanism-based collaborative feature subset is determined to be a redundant feature subset in a random discrete state, exhibiting high information entropy and low class clustering. For example, combinations containing unstable high-order harmonics and high random fluctuations, or combinations that clearly violate the current limits of the current in the current system, are completely removed from the feature input links of subsequent deep neural networks.

[0047] By constructing a subset of mechanistic collaborative features based on physical coupling relationships, and combining bidirectional evaluation of distribution differences and information entropy, it is ensured that the selected features not only have high mathematical discriminative power, but also strong synergy in the underlying discharge logic, thus purifying the input quality of the model from the source.

[0048] Step 3: Project the retained mechanism-cooperative feature subset onto the transient latent variable space, and calculate the state dependency vector between the current mapped coordinates and multiple preset physical state anchor points.

[0049] The purpose of this step is to transform the purely mathematical feature space into a state space absolutely constrained by the underlying physical laws, thereby providing a physical benchmark for subsequent model inference. The specific implementation process is as follows: Step 3.1: Manifold Mapping and Coordinate Acquisition Extract the subset of mechanistic collaborative features retained in step 2, and let its high-dimensional feature matrix be... A manifold mapping method based on neighborhood structure constraints is adopted to project the manifold onto a low-dimensional transient latent variable space to obtain the corresponding current mapped coordinates. Z The transient latent variable space refers to a low-dimensional manifold space specifically used to characterize the underlying core physical invariants of transient signals after redundant background noise in the high-dimensional observation space has been filtered out by a nonlinear dimensionality reduction algorithm.

[0050] Specifically, uniform manifold approximation and projection (UMAP) or isometric mapping algorithms can be used. These manifold mapping algorithms can preserve the local nonlinear topological relationships between feature dimensions within the mechanistic co-feature subset during dimensionality reduction, preventing the loss of physical coupling information of features during spatial transformation.

[0051] Step 3.2: Preset physical state anchor points In the transient latent variable space, instead of unsupervised blind clustering (such as K-Means), multiple physical state anchor points with clear physical meanings are pre-defined based on the prior mechanisms of real distribution network faults. Specifically, the system extracts the statistical mean of the coordinates mapped to specific typical fault mechanisms in the transient latent variable space from a massive historical sample of standard faults with expert diagnostic labels, using these coordinates as the center coordinates of the anchor points. And extract its spatial evolution and distribution patterns as the covariance matrix. The physical state anchor points include at least: The first anchor point (characterizing the rapid recovery constraint state of insulation): corresponds to the physical process of rapid insulation recovery of the system after lightning flashover or instantaneous burn-through by foreign objects. Due to the extremely short transient singularity of this type of discharge process and the absence of subsequent disturbances, its spatial distribution exhibits an extremely high degree of central aggregation; The second anchor point (characterizing the unstable overlapping conduction state): corresponds to the intermittent and periodic discharge process caused by wind deflection such as tree line friction. Since this type of discharge characteristic exhibits continuous temporal evolution and periodic repetition with environmental wind speed and mechanical friction, its spatial distribution shows a banded manifold characteristic extending along a specific behavioral evolution trajectory.

[0052] The third anchor point (characterizing the dynamic disturbance state of the micro-gap): corresponds to the dense arc ablation process caused by the broken wire being suspended or the metal wire shaking in the wind. Because the electromagnetic stress release of the suspended micro-gap in three-dimensional space has extremely strong high-frequency randomness and disorder, its spatial distribution exhibits a high-frequency random divergence spherical cluster characteristic.

[0053] Step 3.3: Calculate the cooperative distribution distance Calculate the current mapped coordinates Z Cooperative distribution distance between each of the aforementioned physical state anchor points .

[0054] Specifically, to fully consider the spatial distribution of each physical state, Mahalanobis distance is used as the metric for cooperative distribution distance. The specific calculation formula is as follows:

[0055] In the formula, Let the center coordinates of the k-th physical state anchor point be... It is the inverse matrix of the covariance distribution characteristics contained in the k-th physical state anchor point.

[0056] By introducing the covariance distribution characteristics of physical states through Mahalanobis distance, not only is the absolute spatial distance between the current data and the anchor point considered, but it also fits the unique geometric manifold distribution of different discharge phenomena. This solves the problem of discharge state confusion caused by the spatial isotropy of traditional Euclidean distance, and improves the accuracy of transient characteristic state attribution determination.

[0057] Step 3.4: Calculate the state membership vector Based on the cooperative distribution distance Calculate the state membership vector of each physical state anchor point corresponding to the mechanism collaborative feature subset.

[0058] By employing an exponential decay function based on a Gaussian kernel or a Softmax normalization mechanism, the cooperative distribution distance is transformed into a probabilistic membership component. The calculation formula is:

[0059] In the formula, K The total number of physical state anchor points is set, such as , Temperature hyperparameters are used to adjust distance sensitivity. The final output is the state membership vector. Each value in this vector represents the degree of fit between the current transient feature and the corresponding underlying physical evolution state.

[0060] By forcibly injecting physical state anchors with clear physical boundaries into the dimensionality-reduced latent space, the abstract high-dimensional features are aligned with the known discharge mechanism of the underlying power grid. This breaks through the black-box technical bottleneck of unreadable hidden layer features in traditional neural networks and provides an explanatory benchmark with common sense in physics for subsequent constraint reasoning.

[0061] Step 4: Transform the state membership vector into a control mask and inject it into the deep neural network. Based on the joint loss function that includes a physical state deviation penalty term, constrain the intermediate layer feature representation of the deep neural network to be consistent with the corresponding physical state anchor point, and output the initial diagnostic classification.

[0062] This step aims to change the traditional black-box optimization model of deep learning models, which relies solely on data. By injecting a mask into the physical state vector and combining it with the penalty of the loss function, it forces the evolution trajectory of network features based on prior physical knowledge. The specific implementation process is as follows: Step 4.1: Convert the state from the attribute vector into a control mask corresponding to the feature dimension. Receive the state membership vector output from step 3 By using a pre-defined mask generation network, such as a fully connected perceptron with a sigmoid activation function, the low-dimensional membership vector is projected and expanded to generate a control mask that matches the dimension of the feature map of the intermediate layer of the deep neural network. The control mask is not generated by freely learning internal network parameters, but is directly driven by the physical state membership vector. The calculation formula can be expressed as:

[0063] In the formula, and For the mapping weights and biases of the mask generation network, This is the activation function. The mask is essentially a feature attention matrix, the magnitude of which reflects the expected activation of specific high-dimensional feature channels by the underlying physical state. The mapping weights... With bias It is pre-set based on the prior knowledge base of power grid mechanisms. It is used to characterize the degree of correlation between each physical state and the feature channel. The stronger the correlation, the larger the absolute value of its corresponding weight. Used to adjust the base activation level of the mask generation network. Preferred The value range of each internal element is set to [-3.0, 3.0]. The value range of each internal element is set to [-1.0, 1.0].

[0064] Step 4.2: Inject the control mask into the deep neural network and adjust the feature extraction weights. During the forward propagation of a deep neural network model, such as a one-dimensional convolutional neural network (1D-CNN) or a Transformer architecture network, the generated control mask is... Inject hidden feature extraction layers. Employ an element-wise multiplication broadcast mechanism and dynamically adjust the feature extraction weights of the deep neural network using the control mask:

[0065] In the formula, The intermediate layer feature representation extracted by the network refers specifically to the high-dimensional mathematical tensor or feature map set extracted by the deep neural network before the final classification decision layer to characterize the deep nonlinear mapping relationship of the input data. In this embodiment, it not only contains the mathematical features of the original waveform data, but also carries the high-dimensional topological evolution information after being forcibly guided by the physical mask. This indicates element-wise multiplication; This is the feature representation guided by the physical state. This step forces the network feature response to focus on the feature channel corresponding to the physical state with high dependency. For example, if the state dependency vector indicates that the current situation is most likely a micro-gap dynamic perturbation, the mask will amplify the weight of the high-frequency spike feature channel and suppress the weight of the low-frequency steady-state channel.

[0066] By transforming the state of the underlying physical state from a membership vector into an attention control mask and injecting it into the hidden layer, the attention allocation of the model is forced to conform to the underlying discharge mechanism, thereby improving the network's sensitivity to key fault features and its anti-interference ability.

[0067] Step 4.3: Calculate the state difference value and generate the physical state deviation penalty term. During the network training phase, in order to prevent intermediate layer feature representations Deviating from objective physical laws, based on the mechanism labels of historical fault samples, each fault category is pre-associated to the corresponding target physical state anchor point, the feature representation of this intermediate layer is extracted, and a linear mapping function is constructed through a pre-set feature projection layer. Reconstruct it back into the transient latent variable space and calculate its relationship with the corresponding target physical state anchor point. The state difference values ​​between them:

[0068] The linear mapping function can be implemented by a single-layer fully connected network. This state difference value quantifies the topological distance between the network's internal mathematical characteristic manifold and the absolute physical reference, and is directly used as a penalty term for physical state deviation. By imposing continuous physical constraints on the evolution process of intermediate layer features, deviations of the feature manifold within the network from the actual discharge mechanism are avoided.

[0069] Step 4.4: Combine classification prediction errors and construct a joint loss function. Calculate the classification prediction error between the output classification of the deep neural network and the actual fault label. Specifically, standard cross-entropy loss can be used. Then, the classification prediction error is combined with the physical state deviation penalty term to construct a joint loss function. :

[0070] In the formula, This is the physical constraint penalty coefficient, used to adjust the balance between pure data-driven fitting and physical prior constraints.

[0071] Step 4.5: Update model parameters based on the joint loss function and output the initial diagnostic classification. Based on the joint loss function The backpropagation algorithm is executed to update the model parameters of the deep neural network.

[0072] During parameter updates, if the network attempts to activate features with high mathematical fit but violate physical laws, such as relying on environmental noise to reduce... The physical state deviation penalty term This will rise sharply, generating a huge inverse gradient. This constrains the intermediate layer feature representation to converge to a range of physical consistency defined by the corresponding physical state anchor point.

[0073] In the actual reasoning application stage, a deep neural network with updated parameters and physical state constraints is used to perform fully connected decoding on the masked features. The probability distribution of each fault type is output through the Softmax layer, and the one with the highest probability is extracted as the initial diagnostic classification.

[0074] By employing a physical consistency loss constraint mechanism, the flaw of models drawing erroneous conclusions by exploiting spurious correlations such as environmental noise is addressed. This mechanism is equivalent to adding a physical rule checker to the neural network, forcing the hidden layer feature manifold to adhere to expert-calibrated physical boundaries. This not only improves the model's generalization ability under complex signal-to-noise ratios but also enhances the transparency of the black-box model, meeting the reliability requirements of automated scheduling and control systems.

[0075] Step 5: Calculate the marginal contribution of each mechanism collaborative feature subset during network inference; for the initial diagnostic classification, extract the mechanism collaborative feature subset with the highest marginal contribution, and verify whether the physical state anchor point corresponding to the mechanism collaborative feature subset is consistent with the initial diagnostic classification; if consistent, output the final diagnostic result.

[0076] This step aims to verify, through a marginal contribution inversion mechanism, whether the internal decision-making basis of the deep neural network conforms to the underlying physical logic of the power grid, thus forming a closed-loop mapping. The specific implementation process is as follows: Step 5.1: Obtain the intermediate layer feature representation Real-time acquisition of intermediate layer feature representations of the deep neural network when outputting the initial diagnostic classification. The intermediate layer feature representation contains the full high-dimensional manifold information that is activated when the sample data flows through the hidden layers of the network.

[0077] Step 5.2: Feature Attribution and Marginal Contribution Calculation Invoke the feature attribution algorithm, and combine it with the intermediate layer feature representation. Calculate the marginal contribution of each of the aforementioned mechanism co-feature subsets to the generation of the initial diagnostic classification.

[0078] Specifically, a feature attribution algorithm based on a game-theoretic alliance mechanism is adopted. The complete set consisting of all retained subsets of collaborative features is denoted as the game-theoretic alliance set. (Assume the total number of subsets is) To quantify specific mechanism-related collaborative feature subsets. i The contribution of the initial diagnostic classification is directly calculated using the following formula based on spatial probability difference as its marginal contribution. :

[0079] In the formula, For subsets of collaborative features that do not contain specific mechanisms Any combination of subsets, From the complete collection N Remove the first i One, and all remaining feature subsets; The number of elements in the combination is represented by the symbol "!", which is the mathematical factorial operator. The fractional terms are used to assign the combination weights based on the order in which features are introduced. The marginal contribution calculus function corresponds to the intermediate layer of the deep neural network, which only accepts a combination of inputs. When the feature subset is selected, the network outputs the classification probability value of the initial diagnostic classification label. Then a feature subset is introduced accordingly. The network then outputs the classification probability value for this category label. The difference between the two is... That is, the feature subset The probability change.

[0080] Step 5.3: Extract the highest contribution subset and determine the anchor point Iterate through the marginal contributions of all feature subsets, extract the mechanism-cooperative feature subset with the largest marginal contribution value, and denote it as... Subsequently, bidirectional anchor point positioning was performed: Diagnostic state anchor point: Based on the initial diagnostic classification label output at present, such as "tree-line contradiction fault", the preset mechanism knowledge base is called to obtain the target physical state anchor point corresponding to the fault type in the latent variable space, and it is marked as the diagnostic state anchor point.

[0081] Determine the attribution state anchor point: Extract the mechanism collaborative feature subset of the maximum marginal contribution. Obtain its distance metric mechanism in step 3 and calculate the feature subset. For each physical state anchor point, extract the physical state anchor point with the highest state dependency (highest probability) and mark it as the attribution state anchor point.

[0082] Step 5.4: Consistency Comparison and Result Output Compare whether the attribution state anchor point and the diagnostic state anchor point are the same physical state anchor point: If both are at the same physical state anchor point, for example, if the initial diagnostic classification is "tree-line contradiction," and its diagnostic state anchor point is the second anchor point; and the feature subset that contributes the most to driving this classification is also mapped to the second anchor point representing the unstable interconnected conduction state, then the reasoning logic of the deep neural network is determined to fully conform to the underlying physical mechanism of electricity, and the verification passes. The initial diagnostic classification is then output as the final diagnostic result.

[0083] If the two are inconsistent, it is determined that the model has made a mathematical misjudgment based on spurious correlation, triggering the blocking mechanism, refusing to output a diagnostic conclusion and reporting a confidence warning.

[0084] By introducing a feature attribution algorithm, the abstract network inference process is deconstructed into the marginal contribution of each mechanism's collaborative feature subset, and bidirectional consistency verification is performed through diagnostic state anchors and attribution state anchors. This mechanism does not merely provide a post-hoc interpretation of the classification results of the deep neural network, but further verifies whether the core features driving the generation of the classification results originate from underlying physical state anchors consistent with the fault type, thereby achieving physical mechanism auditing of the network inference logic.

[0085] When the attribution state anchor point and the diagnostic state anchor point are consistent, it indicates that the deep neural network has completed fault identification based on the correct discharge mechanism. When they are inconsistent, it is determined that the model has an erroneous reasoning tendency based on environmental noise or spurious correlation features, and a blocking mechanism is triggered to refuse to output the diagnostic result. Through this dual-anchor consistency verification mechanism, an additional physical logic verification closed loop is established outside the output layer of the deep learning model. This ensures that the network reasoning results not only have mathematical classification correctness but also consistency with the underlying power mechanism, thereby effectively reducing the risk of misjudgment in complex power distribution environments and improving the physical credibility and engineering usability of the fault identification results.

[0086] Step 5.5: Extraction of key waveform intervals and distribution of evidence chain Upon successful verification and output of the final diagnosis result, based on the extracted... The original timestamps of the time recording are extracted, such as the start time and time span of the pulse peak. Based on the original timestamps, the corresponding key waveform intervals, such as specific discharge cycle segments, are extracted from the transient waveform data.

[0087] Finally, the final diagnostic results, the key waveform ranges, and the feature subsets driving the decision are combined and encapsulated into structured fault evidence chain data, and adaptively sent to the distribution network automation dispatch and control terminal through the distribution network communication interface to directly drive the relay protection device or automatic switch to make switching decisions.

[0088] This process not only outputs the final diagnostic results, but also uses feature marginal contributions to reverse locate and extract key waveform intervals from the original data, encapsulating them into structured fault evidence chain data and directly sending it to the automated dispatch and control terminal. This standardized data stream can be directly understood and executed by the distribution network control platform, realizing a fault identification and protection closed loop without manual intervention, greatly improving the proactive defense and intelligent operation and maintenance level of the distribution network.

[0089] Example 2: This invention provides a distribution network fault identification system based on transient morphological characteristics, which operates the distribution network fault identification method based on transient morphological characteristics as described above, including: The transient data acquisition and morphology construction module is used to acquire transient waveform data of the distribution network, extract the basic multidimensional features of the transient waveform data, and construct a transient discharge morphology description vector based on the transient waveform data. The feature synergy combination and credibility assessment module is used to combine the transient discharge morphology description vector with the basic multidimensional features to construct a mechanism synergy feature subset, and combine it with the grounding operation parameters of the distribution network to conduct credibility assessment and eliminate redundant feature subsets that are in a random discrete state. The latent variable space mapping and state anchoring module is used to project the retained mechanism collaborative feature subset onto the transient latent variable space and calculate the state membership vector between the current mapped coordinates and multiple preset physical state anchor points. The physical mask injection and constrained reasoning module is used to transform the state membership vector into a control mask for injection into the deep neural network, and based on a joint loss function including a physical state deviation penalty term, constrain the intermediate layer feature representation of the deep neural network to be consistent with the corresponding physical state anchor point, and output the initial diagnostic classification. The feature attribution and physical consistency verification module is used to calculate the marginal contribution of each mechanism collaborative feature subset during network inference; for the initial diagnostic classification, the mechanism collaborative feature subset with the highest marginal contribution is extracted, and the physical state anchor point corresponding to the mechanism collaborative feature subset is verified to be consistent with the initial diagnostic classification; if consistent, the final diagnostic result is output.

[0090] Example 3: This invention provides a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method as described.

[0091] Example 4: The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0092] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0093] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0094] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0095] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for identifying distribution network faults based on transient morphological characteristics, characterized in that, include: Acquire transient waveform data of the distribution network, extract the basic multidimensional features of the transient waveform data, and construct a transient discharge morphology description vector based on the transient waveform data; The transient discharge morphology description vector is combined with the basic multidimensional features to construct a mechanism collaborative feature subset. The reliability is evaluated by combining the grounding operation parameters of the distribution network, and redundant feature subsets that are in a random discrete state are eliminated. The retained mechanism-cooperative feature subset is projected into the transient latent variable space, and the state membership vector between the current mapped coordinates and multiple preset physical state anchor points is calculated. The state is transformed from a membership vector into a control mask and injected into a deep neural network. Based on a joint loss function that includes a penalty term for deviation from the physical state, the intermediate layer feature representation of the deep neural network is constrained to be consistent with the corresponding physical state anchor point, and an initial diagnostic classification is output. Calculate the marginal contribution of each mechanism collaborative feature subset during network inference; for the initial diagnostic classification, extract the mechanism collaborative feature subset with the highest marginal contribution, and verify whether the physical state anchor point corresponding to the mechanism collaborative feature subset is consistent with the initial diagnostic classification; if consistent, output the final diagnostic result.

2. The distribution network fault identification method based on transient morphological characteristics according to claim 1, characterized in that, Acquire transient waveform data of the distribution network, extract the basic multidimensional features of the transient waveform data, and construct a transient discharge morphology description vector based on the transient waveform data, specifically including: The transient waveform data includes sequences of zero-sequence current, three-phase current, three-phase voltage, and electric field changes; the basic multidimensional features include time-domain features, frequency-domain features, and time-frequency-domain features. Extract the waveform range from the transient waveform data that exceeds a preset threshold to recover below the preset threshold, and mark it as a pulse single peak; extract the time span of the pulse single peak to generate a morphological component characterizing the instantaneous conduction duration; The frequency of changes in local extreme points of the transient waveform data per unit time is statistically analyzed to generate morphological components that characterize the density of local high-frequency disturbances. The time interval between adjacent pulse peaks is extracted to generate a morphological component characterizing the periodic stability of intermittent discharge. Calculate the difference between the three-phase voltage and zero-sequence current within a preset period before and after the fault transient, and generate morphological components characterizing the self-recovery capability of the insulation. The morphological components are combined and encoded according to their correlation with discharge behavior to construct the transient discharge morphological description vector.

3. The distribution network fault identification method based on transient morphological characteristics according to claim 2, characterized in that, The transient discharge morphology description vector is combined with the basic multidimensional features to construct a mechanism-coordinated feature subset. This subset is then used for reliability assessment in conjunction with the grounding operation parameters of the distribution network. Redundant feature subsets exhibiting random discrete states are eliminated. Specifically, this includes: Based on the coupling relationship between current, voltage and electric field during transient discharge, the morphological components in the transient discharge morphology description vector are combined with the basic multidimensional features that have physical correlation to generate multiple mechanism synergistic feature subsets. Obtain the current grounding operation parameters of the distribution network, and determine the characteristic physical constraint boundary of the corresponding grounding system based on the grounding operation parameters; Calculate the distribution difference index of each of the aforementioned mechanism collaborative feature subsets when distinguishing different fault types, and the information entropy index characterizing the degree of feature fluctuation dispersion; By integrating the physical constraint boundary of the features, the distribution difference index, and the information entropy index, the mechanism collaboration credibility score of each of the mechanism collaboration feature subsets is calculated; The subset of mechanism-cooperation features whose reliability score is lower than a preset evaluation threshold is determined to be a redundant subset of features in a random discrete state.

4. The distribution network fault identification method based on transient morphological characteristics according to claim 3, characterized in that, The retained subset of mechanistic collaborative features is projected into the transient latent variable space, and the state dependency vector between the current mapped coordinates and multiple preset physical state anchor points is calculated, specifically including: A manifold mapping method based on neighborhood structure constraints is adopted to map the retained mechanism collaborative feature subset to the transient latent variable space and obtain the corresponding current mapping coordinates; In the transient latent variable space, multiple physical state anchor points are set according to the fault prior mechanism; the physical state anchor points include at least a first anchor point characterizing the insulation fast recovery constraint state, a second anchor point characterizing the unstable lap connection conduction state, and a third anchor point characterizing the micro-gap dynamic disturbance state. Calculate the cooperative distribution distance based on the mean and covariance distributions of the current mapped coordinates and each physical state anchor point; Based on the normalized result of the cooperative distribution distance, and combined with the covariance distribution characteristics contained in each of the physical state anchor points, the state membership vector of each of the physical state anchor points corresponding to the mechanism cooperative feature subset is calculated.

5. The distribution network fault identification method based on transient morphological characteristics according to claim 4, characterized in that, The state is transformed from a membership vector into a control mask and injected into a deep neural network. Based on a joint loss function that includes a penalty term for deviation from the physical state, the intermediate layer feature representations of the deep neural network are constrained to be consistent with the corresponding physical state anchor points, and an initial diagnostic classification is output, specifically including: Convert the state from the membership vector into a control mask corresponding to the feature dimension; The control mask is injected into the deep neural network, and the feature extraction weights of the deep neural network are adjusted using the control mask. Calculate the state difference value between the intermediate layer feature representation and the corresponding physical state anchor point, and use the state difference value as the physical state deviation penalty term; Calculate the classification prediction error of the deep neural network, and combine the classification prediction error with the physical state deviation penalty term to construct the joint loss function; The model parameters of the deep neural network are updated based on the joint loss function to constrain the intermediate layer feature representation to converge to the physical consistency range defined by the corresponding physical state anchor point. The initial diagnostic classification is output through a deep neural network with updated parameters.

6. The distribution network fault identification method based on transient morphological characteristics according to claim 5, characterized in that, Calculate the marginal contribution of each mechanism collaborative feature subset during network inference; for the initial diagnostic classification, extract the mechanism collaborative feature subset with the highest marginal contribution, and verify whether the physical state anchor point corresponding to the mechanism collaborative feature subset is consistent with the initial diagnostic classification; if consistent, output the final diagnostic result, specifically including: Obtain the intermediate layer feature representations when the deep neural network outputs the initial diagnostic classification; The feature attribution algorithm is invoked, and the marginal contribution of each of the mechanism-coordinated feature subsets to the generation of the initial diagnostic classification is calculated in combination with the intermediate layer feature representation. Extract the mechanism collaborative feature subset with the highest marginal contribution; obtain the diagnostic state anchor point corresponding to the initial diagnostic classification; determine the attribution state anchor point corresponding to the extracted mechanism collaborative feature subset in the transient latent variable space; Compare whether the attribution state anchor point and the diagnostic state anchor point are the same physical state anchor point; if so, output the initial diagnostic classification as the final diagnostic result.

7. The distribution network fault identification method based on transient morphological characteristics according to claim 6, characterized in that, The feature attribution algorithm is invoked, and combined with the intermediate layer feature representation, to calculate the marginal contribution of each of the aforementioned mechanism-related collaborative feature subsets to the generation of the initial diagnostic classification. Specifically, this includes: A feature attribution algorithm based on a game-theoretic alliance mechanism is used to calculate the marginal probability difference of each of the aforementioned mechanism collaborative feature subsets to the initial diagnostic classification, and the result after permutation and combination weighted integration is used as the marginal contribution. Based on the extracted mechanistic collaborative feature subset with the largest marginal contribution value, the key waveform intervals in the original transient waveform data are determined; The final diagnostic results and the key waveform intervals are encapsulated into structured fault evidence chain data and sent to the distribution network automation dispatch and control terminal.

8. A distribution network fault identification system based on transient morphological characteristics, operating the distribution network fault identification method based on transient morphological characteristics as described in any one of claims 1-7, characterized in that, include: The transient data acquisition and morphology construction module is used to acquire transient waveform data of the distribution network, extract the basic multidimensional features of the transient waveform data, and construct a transient discharge morphology description vector based on the transient waveform data. The feature synergy combination and credibility assessment module is used to combine the transient discharge morphology description vector with the basic multidimensional features to construct a mechanism synergy feature subset, and combine it with the grounding operation parameters of the distribution network to conduct credibility assessment and eliminate redundant feature subsets that are in a random discrete state. The latent variable space mapping and state anchoring module is used to project the retained mechanism collaborative feature subset onto the transient latent variable space and calculate the state membership vector between the current mapped coordinates and multiple preset physical state anchor points. The physical mask injection and constrained reasoning module is used to transform the state membership vector into a control mask for injection into the deep neural network, and based on a joint loss function including a physical state deviation penalty term, constrain the intermediate layer feature representation of the deep neural network to be consistent with the corresponding physical state anchor point, and output the initial diagnostic classification. The feature attribution and physical consistency verification module is used to calculate the marginal contribution of each mechanism collaborative feature subset during network inference; for the initial diagnostic classification, the mechanism collaborative feature subset with the highest marginal contribution is extracted, and the physical state anchor point corresponding to the mechanism collaborative feature subset is verified to be consistent with the initial diagnostic classification; if consistent, the final diagnostic result is output.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

Images