A power distribution cabinet fault identification method based on big data analysis

Abstract

The application relates to the technical field of big data analysis, in particular to a power distribution cabinet fault identification method based on big data analysis. The method comprises the following steps: collecting operation physical data and external environment parameters, normalizing and deducting environment offset values to generate a purified characteristic vector; calculating divergence deviation values, triggering a weight consolidation mechanism to output updated network parameters when out of range; mapping the characteristic vector sequence and the updated network parameters to output a time sequence trajectory point set, and deriving the time sequence trajectory point set to generate a drift speed parameter and a drift acceleration parameter; inputting a hierarchical diagnosis matrix to output an evolution diagnosis label, extracting target feature identification and matching an abnormal component node when abnormal, and outputting a repair work order. The application improves the tracking and early warning capability under long-period service by constructing an organic closed loop from a data source to a physical entity.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, specifically to a method for identifying faults in power distribution cabinets based on big data analytics. Background Technology

[0002] In the construction and operation of smart grids, condition monitoring and fault identification of core electrical equipment such as distribution cabinets is an important technological application area. Existing technologies generally adopt fault diagnosis methods that combine multi-dimensional sensors with basic machine learning models. This involves monitoring changes in physical quantities (such as temperature, air pressure, partial discharge, etc.) inside the distribution cabinet. When the collected values ​​or aggregated features reach a specific threshold, the system automatically identifies and determines the health status or fault type of the equipment within a pre-trained classification network. This is currently the mainstream technology for early warning of distribution network equipment conditions.

[0003] However, existing technical solutions have significant shortcomings in terms of diagnostic accuracy, tracking continuity, and practical engineering implementation: Existing methods typically make threshold judgments directly based on raw collected data, making it difficult to effectively identify and eliminate natural deviations caused by the environment. This leads to early, weak fault precursors being easily masked by environmental changes, resulting in frequent false alarms or missed alarms. Secondly, existing condition monitoring is mostly a "snapshot" static feature identification, often relying solely on data from the current time slice for isolated judgments. It lacks the ability to dynamically track and warn of the long process from health to degradation of equipment, often only issuing alarms in the late stages when faults have already manifested, missing the best window for intervention during the latency period. In addition, existing technologies cannot adapt to the natural aging and dynamic degradation evolution of equipment under long-term service, causing early, weak fault precursors to be masked by environmental changes, and the final diagnostic results to be detached from the physical entity.

[0004] To address this, a fault identification method for power distribution cabinets based on big data analysis is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying faults in power distribution cabinets based on big data analysis, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying faults in power distribution cabinets based on big data analysis, comprising: The system collects operational physical data including cabinet temperature, chamber pressure, partial discharge amplitude, and phase current, as well as external environmental parameters including temperature, humidity, and atmospheric pressure. The external environmental parameters are input into the transfer function model to obtain environmental offset values. The operational physical data are normalized and the environmental offset values ​​are subtracted to generate a purification feature vector. The purified feature vector is compared with the preset benchmark data to calculate the divergence deviation value; when the divergence deviation value exceeds the preset deviation threshold, the weight consolidation mechanism is triggered to adjust the feature mapping network and output the updated network parameters. The feature vectors are concatenated, sorted, and purified to generate a feature vector sequence; the feature vector sequence and updated network parameters are mapped in a multidimensional way to output a time-series trajectory point set; the velocity derivative and acceleration derivative of the time-series trajectory point set are calculated to generate drift velocity parameters and drift acceleration parameters. Input the drift velocity parameters and drift acceleration parameters into the preset hierarchical diagnosis matrix and output the evolution diagnosis label; when the evolution diagnosis label is determined to enter the abnormal range, calculate the feature contribution value and extract the target feature identifier based on the feature vector sequence and updated network parameters; map the target feature identifier, match the abnormal component node, and output the maintenance work order.

[0007] Preferably, the specific calculation process of the environmental offset value includes: retrieving historical operational physical data and historical external environmental parameters, performing normalization, and outputting a dimensionless feature set; using an identification algorithm to fit the correlation response matrix based on the dimensionless feature set, and outputting a set of gain coefficients and a thermal hysteresis constant; constructing a transfer function model using the set of gain coefficients and the thermal hysteresis constant; normalizing the temperature, humidity, and atmospheric pressure variables within the external environmental parameters, and outputting dimensionless parameters; inputting the dimensionless parameters into the transfer function model, and outputting basic response parameters; performing amplitude modulation on the basic response parameters based on the set of gain coefficients and time shifting based on the thermal hysteresis constant, and outputting a same-dimensional offset vector; using the same-dimensional offset vector as the environmental offset value; extracting operational physical data and performing normalization processing, and outputting dimensionless operational parameters; performing matrix subtraction between the dimensionless operational parameters and the environmental offset value, and outputting a benchmark difference matrix; performing extreme value anomaly screening on the benchmark difference matrix, and outputting a purification feature vector.

[0008] Preferably, the specific calculation process of the divergence deviation value includes: extracting the distribution matrix of the training feature mapping network as preset benchmark data; setting a sliding observation window to collect and accumulate multiple sets of cleaned feature vectors to construct a measured distribution matrix; using a preset kernel function to calculate the spatial inner product of the measured distribution matrix and the preset benchmark data, and outputting a high-dimensional kernel matrix; calculating the empirical mapping expectation of the measured distribution matrix and the preset benchmark data based on the high-dimensional kernel matrix, and outputting the measured high-dimensional mean and the benchmark high-dimensional mean; finding the maximum mean difference between the measured high-dimensional mean and the benchmark high-dimensional mean, and outputting the divergence deviation value.

[0009] Preferably, the specific generation process of the updated network parameters includes: when the divergence deviation value exceeds a preset deviation threshold, outputting a value out-of-bounds instruction; in response to the value out-of-bounds instruction, freezing the preset feature mapping network bottom-level parameters and opening the top-level connection parameters as the network parameters to be adjusted; retrieving the Fisher information matrix retained during the historical training phase; using the Fisher information matrix to evaluate the importance of the path weights of each neuron within the network parameters to be adjusted, and outputting a path importance index; constructing a joint optimization function including cross-entropy loss and elastic regularization term; using the path importance index to control the constraint strength of the elastic regularization term, and outputting an adaptive penalty function; applying the adaptive penalty function to the purified feature vector to perform gradient descent on the network parameters to be adjusted, and outputting temporary iterative parameters; using the gradient descent algorithm to perform a numerical update operation on the temporary iterative parameters associated with path importance indices less than a preset importance value benchmark, and outputting updated network parameters.

[0010] Preferably, the specific generation process of the time-series trajectory point set includes: constructing a feature vector sequence by arranging multiple sets of purified feature vectors in series according to the order of acquisition time; injecting the feature vector sequence into the hidden layer nodes of the network driven by updating network parameters; performing latent space dimensionality upscaling mapping on the feature vector sequence using the hidden layer nodes of the network to obtain the feature dimensionality upscaling result; projecting the feature dimensionality upscaling result onto a preset high-dimensional feature state space to generate an initial trajectory point set; calculating the Mahalanobis distance distribution density of each node in the initial trajectory point set and outputting a node density scalar; filtering isolated outlier interference nodes based on the node density scalar and outputting a purified high-dimensional point set; and smoothing the purified high-dimensional point set to output the time-series trajectory point set.

[0011] Preferably, the specific calculation process of the drift velocity parameter and drift acceleration parameter includes: measuring the difference in high-dimensional spatial position between adjacent time nodes within the time series trajectory point set, and outputting a coordinate displacement difference vector; dividing the coordinate displacement difference vector by the time interval step size to obtain the vector derivative, and outputting the velocity vector parameter; extracting the first and last nodes of the time series trajectory point set to construct a global trend vector; calculating the cosine value of the angle between the global trend vector and the velocity vector parameter, and outputting a direction correction coefficient; adjusting the magnitude of the velocity vector parameter using the direction correction coefficient multiplication, and outputting the drift velocity parameter; measuring the difference change of the velocity vector parameter between adjacent time nodes, and outputting the velocity evolution difference; dividing the velocity evolution difference by the time interval step size to perform a second vector derivative calculation and outputting the drift acceleration parameter.

[0012] Preferably, the specific extraction process of the target feature identifier includes: extracting the embedded numerical boundaries of a preset hierarchical diagnosis matrix, and outputting velocity division boundaries and acceleration division boundaries; performing numerical comparison between the drift velocity parameters and the velocity division boundaries, and outputting a velocity level index; performing numerical comparison between the drift acceleration parameters and the acceleration division boundaries, and outputting an acceleration level index; inputting the velocity level index and the acceleration level index into the preset hierarchical diagnosis matrix for coordinate addressing, and outputting an evolutionary diagnosis label; determining that the evolutionary diagnosis label falls into an abnormal range, and outputting an abnormal alarm signal; responding to the abnormal alarm signal, extracting the feature vector sequence within the sliding observation window and updating the network parameters, and outputting an interpretation and deduction base; performing game-theoretic calculations on the interpretation and deduction base using the Shapley interpretation algorithm, and outputting a game distribution matrix; quantifying the marginal contribution expectations representing the cabinet temperature, air chamber pressure, partial discharge amplitude, and current dimensions of each phase according to the game distribution matrix, and outputting feature contribution values; comparing the magnitude of feature contribution values ​​of each dimension, and outputting a contribution ranking list; extracting the primary influencing dimension at the top of the contribution ranking list; and calling a preset reverse mapping rule to transform and assemble the primary influencing dimension to output the target feature identifier.

[0013] Preferably, the specific process for generating the maintenance work order includes: acquiring and parsing the structural drawings of the internal hardware equipment of the distribution cabinet; extracting the electrical connection attributes and thermal conduction attributes of circuit breaker components, disconnector components, and cable head components, and outputting multi-dimensional topological relationships; constructing a physical topology map using the multi-dimensional topology relationships; configuring the vertices of the physical topology map as hardware entities and the connections as entity interaction energy flows; inputting the target feature identifier into the physical topology map to trigger topology path optimization calculation and outputting sensitive association paths; searching for highly sensitive vertices within the physical topology map along the sensitive association paths; mapping the highly sensitive vertices to generate abnormal component nodes; extracting the embedded spatial location coordinates and hardware model of the abnormal component nodes and packaging them to output the maintenance work order.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By collecting operational physical data including cabinet temperature, chamber pressure, partial discharge amplitude, and phase current, as well as external environmental parameters including temperature, humidity, and atmospheric pressure, a mild dynamic decoupling mechanism was established. During processing, the external environmental parameters are input into the transfer function model to obtain environmental offset values. Then, at the data source, the operational physical data is normalized and the environmental offset values ​​are subtracted to generate a purification feature vector. This decoupling method follows the objective laws of equipment thermodynamics, filtering out non-fault fluctuations caused by external factors and restoring the original information reflecting the equipment's own health status. The purified data not only lays a high-confidence foundation for subsequent analysis but also effectively captures subtle early signs of degradation.

[0015] 2. By comparing the purification feature vector with the preset benchmark data, the divergence deviation value is calculated. When the divergence deviation value exceeds the preset deviation threshold, the weight consolidation mechanism can be adaptively triggered to adjust the feature mapping network and output updated network parameters. While having important state memory, it smoothly absorbs the data distribution shift caused by normal aging, so that the digital mapping network truly fits the real life cycle of the physical device.

[0016] 3. By constructing a temporal dynamics analysis perspective, the feature vector sequence generated by splicing, sorting, and purifying feature vectors is combined with the updated network parameters for multi-dimensional mapping, outputting a temporal trajectory point set. The velocity derivative and acceleration derivative of the temporal trajectory point set are calculated to generate drift velocity parameters and drift acceleration parameters. By inputting the drift velocity parameters and drift acceleration parameters into a preset hierarchical diagnostic matrix, evolutionary diagnostic labels are output, realizing a flexible transition from static judgment to dynamic trend prediction, characterizing the rate and trend of equipment state deviation, enabling maintenance personnel to calmly plan intervention strategies before faults develop.

[0017] 4. Constructing an organic closed loop from the data source to the physical entity: The front end normalizes the running physical data and subtracts environmental offset values ​​to generate a purified feature vector to solidify the data foundation; the middle end drives the generation of updated network parameters to ensure the model's adaptive cognition; when the back end determines that the evolutionary diagnostic label has entered an abnormal range, it calculates the feature contribution value and extracts the target feature identifier based on the feature vector sequence and updated network parameters, and finally maps the target feature identifier to match the abnormal component node and outputs a maintenance work order; this interlocking design completely breaks down the barriers between the digital model and the hardware itself, enabling the subtle feature evolution captured by the front end to be smoothly transmitted to the specific physical component along the algorithm chain, achieving a seamless connection between cutting-edge diagnostic algorithms and on-the-ground maintenance execution. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a power distribution cabinet fault identification method based on big data analysis, as proposed in an embodiment of this invention application; Figure 2 This is a flowchart of the environmental decoupling and data purification calculation proposed in an embodiment of this invention application; Figure 3 This is a flowchart illustrating the continuous update process of adaptive network parameters proposed in an embodiment of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figures 1-3 The present invention provides a method for identifying faults in power distribution cabinets based on big data analysis, the specific steps of which are as follows: The system collects operational physical data including cabinet temperature, chamber pressure, partial discharge amplitude, and phase current, as well as external environmental parameters including temperature, humidity, and atmospheric pressure. The external environmental parameters are input into the transfer function model to obtain environmental offset values. The operational physical data are normalized and the environmental offset values ​​are subtracted to generate a purification feature vector. The purified feature vector is compared with the preset benchmark data to calculate the divergence deviation value; when the divergence deviation value exceeds the preset deviation threshold, the weight consolidation mechanism is triggered to adjust the feature mapping network and output the updated network parameters. The feature vectors are concatenated, sorted, and purified to generate a feature vector sequence; the feature vector sequence and updated network parameters are mapped in a multidimensional way to output a time-series trajectory point set; the velocity derivative and acceleration derivative of the time-series trajectory point set are calculated to generate drift velocity parameters and drift acceleration parameters. Input the drift velocity parameters and drift acceleration parameters into the preset hierarchical diagnosis matrix and output the evolution diagnosis label; when the evolution diagnosis label is determined to enter the abnormal range, calculate the feature contribution value and extract the target feature identifier based on the feature vector sequence and updated network parameters; map the target feature identifier, match the abnormal component node, and output the maintenance work order.

[0021] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0022] Example 1 This application discloses a method for identifying faults in power distribution cabinets based on big data analysis. (See attached document.) Figure 1The specific steps proposed in this invention include: S1. Collecting operational physical data including cabinet temperature, chamber pressure, partial discharge amplitude, and phase current, as well as external environmental parameters including temperature, humidity, and atmospheric pressure; inputting the external environmental parameters into a transfer function model to obtain environmental offset values; normalizing the operational physical data and subtracting the environmental offset values ​​to generate a purification feature vector; S2. Comparing the purification feature vector with preset benchmark data to calculate the divergence deviation value; when the divergence deviation value exceeds a preset deviation threshold, triggering a weight consolidation mechanism to adjust the feature mapping network and outputting updated network parameters; S3. The feature vectors are concatenated, sorted, and purified to generate a feature vector sequence. The feature vector sequence and updated network parameters are then mapped in a multidimensional manner to output a time-series trajectory point set. The velocity derivative and acceleration derivative of the time-series trajectory point set are calculated to generate drift velocity parameters and drift acceleration parameters. In step S4, the drift velocity parameters and drift acceleration parameters are input into a preset hierarchical diagnostic matrix to output an evolutionary diagnostic label. When the evolutionary diagnostic label is determined to enter an abnormal range, the feature contribution value is calculated and the target feature identifier is extracted based on the feature vector sequence and updated network parameters. The target feature identifier is mapped to match abnormal component nodes, and a maintenance work order is output.

[0023] Furthermore, operational physical data including cabinet temperature, chamber pressure, partial discharge amplitude, and phase current are collected, along with external environmental parameters including temperature, humidity, and atmospheric pressure. These external environmental parameters are input into a transfer function model to obtain environmental offset values. The operational physical data are then normalized and the environmental offset values ​​are subtracted to generate a purification feature vector. (See also...) Figure 2 Corresponding to step S1 above; the specific implementation process includes: Historical operational physical data and historical external environmental parameters are retrieved and normalized to output a dimensionless feature set. Based on the dimensionless feature set, an identification algorithm is used to fit the correlation response matrix, outputting a set of gain coefficients and a thermal hysteresis constant. A transfer function model is constructed using the set of gain coefficients and the thermal hysteresis constant. Temperature, humidity, and atmospheric pressure variables within the external environmental parameters are normalized to output dimensionless parameters. The dimensionless parameters are input into the transfer function model to output basic response parameters. The basic response parameters are amplitude modulated based on the set of gain coefficients and time-shifted based on the thermal hysteresis constant to output a same-dimensional offset vector. The same-dimensional offset vector is used as the environmental offset value. Operational physical data is extracted and normalized to output dimensionless operational parameters. The dimensionless operational parameters are subtracted from the environmental offset value to output a benchmark difference matrix. Extreme value anomaly screening is performed on the benchmark difference matrix to output a purification feature vector.

[0024] Specifically, the calculation process for the environmental offset value is as follows: Firstly, relying on a high-precision sensor network deployed inside the distribution cabinets and in the substation environment, two key types of time-series data are collected simultaneously. The first type of input data is operational physical data, including cabinet temperature collected by PT100 sensors, chamber pressure collected by insulating gas pressure gauges, partial discharge amplitude collected by UHF sensors, and phase current collected by Hall effect current transformers. The second type of input data is external environmental parameters, including substation ambient temperature, relative humidity, and atmospheric pressure obtained from weather stations.

[0025] To eliminate dimensional differences between different physical quantities and prevent numerical overflow in matrix operations, historical operational physical data and historical external environmental parameters are first retrieved and normalized to output a dimensionless feature set. The dimensionless processing involves: for any dimension of feature parameter, subtracting the current period's measured value from the minimum value of that dimension in the historical statistical database; then dividing the difference by the span between the historical maximum and minimum values ​​of that dimension. Specifically, the historical maximum and minimum values ​​are selected during the initialization phase using stable operational period data of a preset length after equipment commissioning as the statistical sample database. A sorting algorithm is used to extract the upper and lower bounds of each physical dimension within this time window. All physical quantities with different units and amplitudes are forcibly mapped to a uniform proportional range of 0 to 1, preserving the fluctuation characteristics of the original data. External environmental parameters undergo the same operation, outputting the corresponding dimensionless parameters.

[0026] Subsequently, based on the acquired historical dimensionless feature set, the least squares system identification algorithm was used to fit the correlation response matrix, constructing a transfer function model from dimensionless parameters of the external environment to the physical parameters of operation within the cabinet. The equations of this transfer function model in the complex frequency domain are as follows: In this computational model, The matrix representing the transfer function between the environment and the internal state; For the complex frequency variable in the Laplace transform, The set of gain coefficients represents a fixed proportion of the changes in the internal state caused by steady-state changes in the external environment. Specifically, it is obtained by fitting the mean ratio of environmental variables to internal variables in the historical dimensionless feature set over a long steady-state translation period. It only applies to the dimensions of cabinet temperature and air chamber pressure that are significantly affected by slow environmental changes. For the dimensions of phase current and transient partial discharge amplitude, the corresponding gain coefficients are set to be close to 0 in the matrix. They are only normalized and then the original values ​​are retained and spliced ​​to prevent the original electrical features from being erroneously cut off. The time constant is used to reflect the dynamic response rate of heat conduction or air chamber pressure conduction in the distribution cabinet. It is measured by applying a step response test to historical operating data and measuring the time required for the output variable to reach a specific percentage (63.2% in this example) of the steady-state value. The thermal hysteresis constant represents the pure time delay by which changes in the external environment are transmitted to the cabinet, causing state fluctuations. It is calculated using a cross-correlation analysis algorithm, corresponding to the hysteresis time step corresponding to the peak value of the cross-correlation coefficient between the time series curves of the external environment and the time series curves of the parameters inside the cabinet. The correlation response matrix is ​​a specific 4×3 real transfer function matrix. Its rows correspond to four operating physical parameters: temperature, air pressure, partial discharge, and current; its columns correspond to three environmental parameters: temperature, humidity, and atmospheric pressure.

[0027] The dimensionless parameters are input into the transfer function model and calculated by discretized difference equations to output the basic response parameters. Then, the response parameters are amplitude modulated based on the set of gain coefficients and shifted on the time axis based on the thermal hysteresis constant to output a same-dimensional offset vector, which is the environmental offset value.

[0028] After obtaining the environmental offset values, the current running physical data is extracted and normalized to output dimensionless running parameters. Then, a matrix subtraction operation is performed between the dimensionless running parameters and the environmental offset values ​​to output a baseline difference matrix. This specific calculation process is an element-wise algebraic subtraction operation: for each feature dimension value in the dimensionless running parameter matrix, the corresponding feature dimension value in the environmental offset value matrix is ​​directly subtracted, and the resulting difference directly constitutes the baseline difference matrix.

[0029] In this embodiment, the fluctuation sequence of the benchmark difference matrix is ​​extracted, and a local smoothing filter is constructed using a Hamming window to perform one-dimensional convolution processing. This removes high-frequency random glitches caused by asynchronous sampling from multiple sensors, outputting a smoothed benchmark difference matrix. Specifically, the fluctuation difference sequence of each physical dimension is extracted from the benchmark difference matrix. Considering the small phase difference between the underlying analog-to-digital conversion clocks of different sensors (such as PT100 and Hall transformers) within the cabinet, the difference sequence is prone to glitches caused by sampling asynchrony. Therefore, a Hamming window of a specific length is constructed as a local smoothing filter kernel. Its weight distribution is calculated by the difference between the constant term and the cosine function, ensuring that the weight is highest at the center and decreases smoothly towards both ends. This filter kernel is then subjected to a one-dimensional discrete convolution operation with the fluctuation sequence: specifically, the original sequence points within the local window are multiplied one by one with the corresponding filter kernel weights and summed, then divided by the sum of weights to complete normalization. The absolute value of the first-order difference between adjacent differences is calculated. If it exceeds a preset glitch threshold (i.e., 1.5 times the root mean square error of historical fluctuations in that dimension), the aforementioned convolutional filtering is activated to reconstruct the abnormal data point; otherwise, the original value is retained. This adaptive window convolution calculation outputs a smoothed baseline difference matrix. In this embodiment, a smoothing window length matching the respective Nyquist sampling rate is set according to the physical feature dimension: a long window is used for temperature quantities with extremely low sampling frequencies, while an extremely short window or pulse-hold filtering mechanism is used for high-frequency partial discharge quantities of 10kHz. This process eliminates random glitches introduced by asynchronous sampling, improves the numerical stability of the matrix, and prevents erroneous triggering of subsequent extreme value screening.

[0030] It should be noted that the optimal application scenario is under normal operating conditions where the load rate of the distribution cabinet is relatively stable (load current fluctuation range within ±20%). When a step-like heavy load is detected in each phase current, resulting in significant nonlinear ohmic heat, and the residual between the current operating parameters and the basic response parameters predicted by the transfer function exceeds a preset threshold for three consecutive cycles, it will be determined that the current operating condition has exceeded the linear applicable zone. This will automatically trigger online re-identification and fitting updates of the gain coefficient set and thermal hysteresis constant to eliminate the fundamental deviation caused by the nonlinear heavy load. The preset threshold is taken as the mean of the residual distribution of the linear transfer function on historical training data plus three times the standard deviation as the boundary of the linear applicable zone.

[0031] After obtaining the difference matrix, anomaly screening is performed. The specific screening and processing process is as follows: Calculate the standard score of the current difference deviating from the historical difference mean. The preset extreme value removal threshold is selected based on the Gaussian distribution confidence interval criterion, choosing a specific multiple of the standard deviation of the historical normal residual distribution as the boundary. When the absolute value of the score in a certain dimension exceeds this threshold, it is determined as an abnormal extreme value, and linear interpolation is performed using the effective values ​​of adjacent time nodes, thereby outputting the denoised and purified feature vector.

[0032] By introducing a thermal hysteresis constant to construct a transfer function, the hysteresis effect of environmental disturbances is characterized. Combined with synchronization normalization and extreme value anomaly screening operations, the noise fluctuations caused by sudden changes in temperature and humidity are smoothed, providing a high-confidence data base for subsequent diagnosis.

[0033] Furthermore, the purified feature vector is compared with preset benchmark data to calculate the divergence deviation value; when the divergence deviation value exceeds a preset deviation threshold, a weight consolidation mechanism is triggered to adjust the feature mapping network and output updated network parameters; see reference Figure 3 Corresponding to step S2 above; the specific implementation process includes: Extract the distribution matrix of the training feature mapping network as the preset benchmark data; set a sliding observation window to collect and accumulate multiple sets of cleaned feature vectors to construct the measured distribution matrix; use a preset kernel function to calculate the spatial inner product of the measured distribution matrix and the preset benchmark data, and output a high-dimensional kernel matrix; based on the high-dimensional kernel matrix, calculate the empirical mapping expectation of the measured distribution matrix and the preset benchmark data respectively, and output the measured high-dimensional mean and the benchmark high-dimensional mean; find the maximum mean difference between the measured high-dimensional mean and the benchmark high-dimensional mean, and output the divergence deviation value.

[0034] Specifically, the calculation process for the divergence deviation value is as follows: After obtaining the cleanup feature vectors, the long-term temporal drift of the observed data distribution pattern is investigated. First, a representative historical feature distribution matrix from the training of the feature mapping network is extracted as the preset baseline data. Simultaneously, a sliding observation window is used to collect and accumulate multiple sets of time-continuous cleanup feature vectors, constructing a measured distribution matrix. The specific selection method for the sliding observation window length and step size is as follows: based on the Nyquist sampling theorem and the evolution time period of typical partial discharge or temperature rise faults in distribution cabinets, the window length is set to the number of samples that can completely encompass at least one typical early fault fluctuation cycle, and the step size is set to the single sampling interval to ensure a smooth transition of the statistical distribution characteristics.

[0035] The measured distribution matrix and the preset benchmark data are simultaneously projected into the reproducing kernel Hilbert space, outputting a high-dimensional kernel matrix. Then, the empirical mapping expectations of the measured distribution matrix and the preset benchmark data are calculated from this matrix, outputting the measured high-dimensional mean and the benchmark high-dimensional mean. Subsequently, the maximum mean difference algorithm is used to calculate the divergence deviation between the two sets of high-dimensional means. The corresponding kernel expansion calculation formula is as follows: In the above spatial mapping and difference formulas, This represents the divergence deviation value in the final calculated output; and This represents the purified feature vector sample projected from the preset benchmark data matrix; and This represents a sample vector in the currently measured distribution matrix; and These represent the total number of samples of baseline data and measured data within the observation window, respectively. This represents the Gaussian radial basis function kernel used. The kernel function contains a bandwidth parameter, the choice of which directly determines the model's sensitivity to differences in high-dimensional distributions. Specifically, it is selected using a median heuristic, which calculates the median of the Euclidean distance between all sample pairs within the joint space of the combined measured distribution matrix and the pre-defined benchmark data matrix. This joint median is then used as the bandwidth of the Gaussian kernel function, ensuring that even when the measured data and benchmark data exhibit significant distribution shifts, the bandwidth kernel still retains the characteristics of the benchmark scale, avoiding a decrease in test sensitivity caused by one-sided estimation.

[0036] By synchronously projecting the feature distribution onto the kernel Hilbert space and measuring the divergence bias by quantifying the maximum mean difference, the nonlinear shift trend hidden in high-dimensional data is captured, making the test of distribution deviation more intuitive.

[0037] When the divergence deviation exceeds a preset deviation threshold, an out-of-bounds instruction is output. In response, the preset feature map network's bottom-level parameters are frozen, while the top-level connection parameters are made available as the network parameters to be tuned. The Fisher information matrix retained from the historical training phase is retrieved. The Fisher information matrix is ​​used to evaluate the importance of each neuron's path weight within the network parameters to be tuned, and a path importance index is output. A joint optimization function including cross-entropy loss and elastic regularization is constructed. The path importance index is used to control the constraint strength of the elastic regularization term, and an adaptive penalty function is output. Gradient descent is applied to the cleaned feature vector using the adaptive penalty function to perform gradient descent on the network parameters to be tuned, and temporary iterative parameters are output. The gradient descent algorithm is used to update the temporary iterative parameters associated with path importance indices that are less than a preset importance value benchmark, and updated network parameters are output.

[0038] Specifically, the process for generating updated network parameters is as follows: When determining whether the divergence deviation value exceeds a preset deviation threshold, the specific selection method for the preset deviation threshold is as follows: Perform a rolling maximum mean difference calculation on the historical data of confirmed fault-free normal conditions with the same sliding window length. Perform probability density function statistics on the output historical deviation value array, and extract the quantile of the high percentile (e.g., the upper boundary threshold) as the preset deviation threshold. When the calculated divergence deviation value exceeds this threshold, a significant distribution shift is determined, and a value out-of-bounds instruction is triggered to initiate the subsequent parameter adjustment mechanism.

[0039] When a divergence bias value exceeds the limit, the diagnostic model needs to be updated based on new data, while simultaneously suppressing catastrophic forgetting of historical normal patterns. In response to the value exceeding the limit instruction, the underlying parameters of the preset feature mapping network are first frozen to fix the extraction logic of local basic features, and only the top-level connection parameters responsible for high-dimensional feature space fusion and classification mapping are exposed as network parameters to be tuned.

[0040] Subsequently, the Fisher information matrix saved during the initial training phase of the feature mapping network is retrieved to evaluate the non-substitutability of the path weights of each neuron within the parameters of the network to be tuned, i.e., the path importance index. The calculation model for this index is as follows: In the structure of this formula, This represents the path importance index of the k-th network parameter to be adjusted, and its physical meaning is the second-order sensitivity of the model output probability distribution to small perturbations in the weights of that specific network. This represents the set of historical network parameters locked before the distribution shift occurred; This represents a specific single parameter of the network to be adjusted; and Corresponding historical input sample set and labels; This represents the mathematical expectation operator based on the distribution patterns of historical data.

[0041] Based on the obtained path importance index, a joint optimization function is constructed, which includes a cross-entropy loss term and an elasticity regularization penalty term. Its functional form is as follows: in, This represents the overall loss value of the adaptive penalty function; The cross-entropy loss generated by performing forward inference using the new cleansed feature vectors within the current sliding window is used to force the model to fit the features of the newly evolved data; The coefficient for the elastic regularization term is selected by optimizing it on the validation set during the initialization phase using a grid search algorithm: the search interval is set to [10, 10000] with a logarithmic distribution, and the search step size is on the order of 10; under each candidate coefficient value, the joint validation error (i.e., the weighted sum of the cross-entropy loss of the new sample and the elastic regularization penalty term) is calculated using historical samples and new samples in the validation set, and the parameter value that minimizes the overall joint validation error is extracted as the penalty coefficient. The path importance index obtained in the preceding steps is used as a weight to constrain the offset range of each parameter. This represents the weight parameter corresponding to the k-th path in the historical network parameter set.

[0042] The adaptive penalty function described above is used to perform gradient descent optimization on the parameters of the network under test. The learning rate parameter in the optimization process is chosen using an exponential decay strategy based on iterative batches, with an initial value set to 10.-4 To prevent gradient explosion, temporary iterative parameters are output after multiple rounds of parameter iteration.

[0043] Finally, a threshold gating mechanism is used to filter and update the temporary iterative parameters. The specific method for selecting the preset importance value benchmark is to extract a specific low quantile value from the distribution of the path importance index of all network parameters to be adjusted as the cutoff threshold. Only temporary iterative parameters with path importance indices lower than the benchmark are numerically overwritten, and their updates are written into the network structure, thereby outputting updated network parameters with the ability to identify new evolutionary features.

[0044] In addition, to prevent sudden physical damage to the equipment from causing the degradation rate to exceed the convergence limit of incremental learning, an online update degradation mechanism is built in: when the divergence deviation values ​​of multiple (e.g., 3) consecutive sliding observation windows are detected to be rising continuously, and the corresponding model diagnostic accuracy does not recover, it is determined that the current online update has failed, and an offline retraining alarm instruction will be automatically triggered, prompting the operation and maintenance personnel to intervene and use a new full dataset containing the sudden failure data for offline retraining, thereby ensuring reliability under extreme operating conditions.

[0045] By using the Fisher information matrix to assess the importance of neuronal pathways and dynamically adjusting the constraint strength of the elastic regularization term, the existing health baseline memory is preserved while absorbing new state features, ensuring the robust evolution of the diagnostic network.

[0046] Further, the feature vectors are concatenated, sorted, and purified to generate a feature vector sequence; the feature vector sequence and updated network parameters are then mapped in a multidimensional manner to output a time-series trajectory point set; the velocity derivative and acceleration derivative are calculated on the time-series trajectory point set to generate drift velocity parameters and drift acceleration parameters; this corresponds to step S3 above; the specific implementation process includes: Multiple sets of purified feature vectors are concatenated in chronological order of collection to construct a feature vector sequence. This feature vector sequence is then injected into the hidden layer nodes of the network driven by updated network parameters. The hidden layer nodes are used to perform latent space upsizing mapping on the feature vector sequence to obtain the feature upsizing result. The feature upsizing result is projected onto a preset high-dimensional feature state space to generate an initial trajectory point set. The Mahalanobis distance distribution density of each node in the initial trajectory point set is calculated, and a node density scalar is output. Isolated outlier interference nodes are filtered based on the node density scalar, and a purified high-dimensional point set is output. The purified high-dimensional point set is then smoothed, and a time-series trajectory point set is output.

[0047] Specifically, the generation process of the time-series trajectory point set is as follows: After obtaining the updated network parameters, feature dimension transformation is performed. The feature mapping network changes from the previous backpropagation mode (independently calculating cross-entropy and penalty gradients using a single set of cleaned feature vectors as input units) to a forward inference mode. Specifically, multiple independent cleaned feature vectors are concatenated and arranged in the order of data acquisition to generate a feature vector sequence. Subsequently, this sequence is input frame by frame (i.e., a set of feature vectors for each time step) into the hidden layer nodes of the network driven by the updated network parameters using a sliding window mechanism. The hidden layer nodes independently perform latent space dimension upscaling mapping for each frame input, using matrix multiplication and nonlinear activation operations of the hidden layer weights to expand the original low-dimensional sequence, obtain the feature dimension upscaling result, and project it onto the high-dimensional feature state space to generate an initial trajectory point set arranged in time.

[0048] The high-dimensional Mahalanobis distance distribution density of each node within the initial trajectory point set is quantitatively calculated, and the distance calculation formula is as follows: In the above calculation formula, This represents the calculated Mahalanobis distance between the i-th spatial node and the center of the point set. Its physical meaning lies in eliminating the scale inconsistency and linear correlation coupling interference between the dimensions of high-dimensional features based on the data scattering shape of the point set itself. Represents the high-dimensional feature vector of discrete time nodes in the initial trajectory point set; This represents the mean centroid vector of the entire point set within the observation window in multidimensional space. This represents the result of inverting the characteristic covariance matrix of the high-dimensional point set. Before performing the matrix inversion operation, a Tikhonov small regularization perturbation (i.e., adding a very small identity matrix constant term αI, where α is 10) is forcibly applied to the main diagonal of the covariance matrix. -6 This ensures the numerical stability of Mahalanobis distance calculations under any edge conditions.

[0049] In this embodiment, the timestamps of trajectory nodes are extracted, and a time weighting coefficient model based on the exponential decay law is established. The Mahalanobis distance density scalar and the time weight are modulated by a dot product, and the timeliness of trajectory evolution is integrated to output a time-sensitive node density scalar. Specifically, after obtaining the spatial Mahalanobis distance distribution density of each node within the initial trajectory point set, a time decay mechanism is introduced to modulate the timeliness of the node density scalar. The specific implementation is as follows: the data acquisition timestamp corresponding to each node in the trajectory point set is extracted, and the current observation and analysis time is set. A time weighting coefficient model based on the negative exponent of the natural logarithm is constructed, where the absolute value of the exponent part is the product of the decay rate threshold parameter and the proportion of the node lag time in the total observation time window. The total observation time window is a preset interval, and the decay rate threshold parameter is set to a fixed constant of 0.8. This constant is based on the typical physical evolution cycle of partial discharge and abnormal temperature rise in the distribution cabinet (usually with a gradual half-life of 8 to 12 hours), ensuring that the weight of the data at the tail edge of the observation window is still retained at approximately 45%. Subsequently, a scalar multiplication operation is performed on the original Mahalanobis distance density values ​​of each node and their corresponding time weight coefficients. This modulation process increases the density weight of nodes closer to the current time. After time weight modulation, a time-sensitive node density scalar is output. The above process imparts temporal sensitivity to high-dimensional trajectories, strengthening the dominance of recently degraded data in anomaly clustering.

[0050] Based on the obtained Mahalanobis distance scalar, a density-based local clustering algorithm is used to evaluate the node density scalar and filter outout points. The filtering threshold is selected by calculating the arithmetic mean and variance of the Mahalanobis distances of all nodes in the current window, and selecting the value of the mean plus a preset constant multiple of the standard deviation as the density boundary. Isolated free nodes with distances greater than this boundary are removed from the space, thereby outputting a purified high-dimensional point set.

[0051] Finally, the purified high-dimensional point set is smoothed to remove high-frequency noise and output the time-series trajectory point set. This smoothing process utilizes a weighted moving average, specifically: the smoothed trajectory node vector at the current time step equals the product of the purified high-dimensional node vector at the current time step and the smoothing attenuation factor, plus the product of the smoothed trajectory node vector at the previous time step and the difference between the smoothing factor and the previous smoothing factor. The smoothing attenuation factor is specifically selected based on extracting the dominant frequency band of high-frequency noise using a fast Fourier transform from historical trajectory data, and then calculating it based on the reciprocal of this dominant frequency band to ensure no significant phase lag in dynamic tracking.

[0052] After the latent space is upgraded, the Mahalanobis distance distribution density is introduced as the evaluation scale. By performing gentle filtering and temporal smoothing on sparsely distributed outlier nodes, random impulse interference is effectively eliminated, and the coherent and clear high-dimensional motion trajectory of the device is outlined.

[0053] The system measures the difference in high-dimensional spatial position between adjacent time nodes within a time-series trajectory point set, outputting a coordinate displacement difference vector. It then divides this vector by the time interval step size to obtain the vector derivative, outputting the velocity vector parameter. The system extracts the first and last nodes of the time-series trajectory point set to construct a global trend vector. It calculates the cosine of the angle between the global trend vector and the velocity vector parameter, outputting a direction correction coefficient. It adjusts the magnitude of the velocity vector parameter using the direction correction coefficient multiplication, outputting the drift velocity parameter. The system measures the change in the difference between velocity vector parameters at adjacent time nodes, outputting the velocity evolution difference. Finally, it divides the velocity evolution difference by the time interval step size to calculate the second vector derivative, outputting the drift acceleration parameter.

[0054] Specifically, the calculation process for drift speed parameters and drift acceleration parameters is as follows: From the constructed time-series trajectory point set, the dynamic evolution parameters of the internal state of the distribution cabinet deteriorating in the characteristic state space are deduced. First, the drift velocity parameter is calculated. The high-dimensional spatial position difference between adjacent time nodes within the time-series trajectory point set is measured. The specific calculation process for this difference vector involves performing a vector subtraction operation between the high-dimensional coordinate vector of the current time node and the high-dimensional coordinate vector of the previous time node, outputting a coordinate displacement difference vector. Subsequently, all dimensional components within this coordinate displacement difference vector are uniformly divided by the time interval step size parameter, outputting the first-order vector derivative result, i.e., the instantaneous velocity vector parameter. The specific selection method for this time interval step size parameter is to directly adopt the fixed digital sampling period of the underlying physical sensors.

[0055] To eliminate localized random and disordered back-and-forth interference in space, the two endpoint nodes at the beginning and end of the time series trajectory point set are extracted. The coordinates of the two nodes are subtracted to construct a global trend vector representing the overall displacement direction. Subsequently, the cosine of the angle between the global trend vector and each instantaneous velocity vector parameter is calculated to output the direction correction coefficient. The calculation model for this cosine angle is as follows: in, The direction correction coefficient at time t represents the output, and its physical meaning is to assess the degree of collinearity between the current instantaneous evolution direction and the long-range decay trend. Represents the global trend vector; This represents the instantaneous velocity vector parameter at the corresponding moment. An algebraic multiplication operation is performed between this direction correction coefficient and the scalar magnitude of the instantaneous velocity vector parameter at that moment. This scaling operation adjusts the velocity magnitude, ultimately outputting the drift velocity parameter in the trend direction.

[0056] Based on the obtained smooth drift velocity parameters, the drift acceleration parameters are further calculated. This calculation also uses discretized derivative operations: the drift velocity parameter at the current time is subtracted from the drift velocity parameter at the previous adjacent time node, and the velocity evolution difference scalar is output. Then, this velocity evolution difference scalar is divided by the selected time interval step size parameter mentioned above, and the second derivative result is output. This result is the drift acceleration parameter.

[0057] By introducing the cosine of the angle between the global trend vector and the velocity vector as a direction correction coefficient, disordered high-frequency oscillations in the feature space are filtered out, making the derived velocity and acceleration more accurately reflect the dominant direction of equipment degradation.

[0058] Further, the drift velocity parameters and drift acceleration parameters are input into a preset hierarchical diagnostic matrix, and an evolutionary diagnostic label is output. When the evolutionary diagnostic label is determined to enter an abnormal range, the feature contribution value is calculated and the target feature identifier is extracted based on the feature vector sequence and updated network parameters. The target feature identifier is mapped to match abnormal component nodes, and a maintenance work order is output. This corresponds to step S4 above. The specific implementation process includes: Extract the embedded numerical boundaries of the preset hierarchical diagnostic matrix and output the velocity division boundary and acceleration division boundary; perform numerical comparison between the drift velocity parameter and the velocity division boundary to output the velocity level index; perform numerical comparison between the drift acceleration parameter and the acceleration division boundary to output the acceleration level index; input the velocity level index and acceleration level index into the preset hierarchical diagnostic matrix for coordinate addressing and output the evolution diagnostic label; determine if the evolution diagnostic label falls into the abnormal range and output an abnormal alarm signal; respond to the abnormal alarm signal, extract the feature vector sequence within the sliding observation window and update the network parameters to output the interpretation and inference base; perform game inference calculation on the interpretation and inference base using the Shapley interpretation algorithm and output the game distribution matrix; quantify the marginal contribution expectation of the cabinet temperature, air chamber pressure, partial discharge amplitude and each phase current dimension based on the game distribution matrix and output the feature contribution value; compare the size of the feature contribution values ​​of each dimension and output the contribution ranking list; extract the primary influence dimension at the top of the contribution ranking list; call the preset reverse mapping rule to transform and assemble the primary influence dimension to output the target feature identifier.

[0059] Specifically, the extraction process of target feature identifiers is as follows: First, the embedded numerical boundaries of the pre-defined hierarchical diagnostic matrix are extracted, and the speed division boundary and acceleration division boundary are output. The specific selection method for the numerical boundary parameters within the hierarchical diagnostic matrix is ​​as follows: extract the historical full life cycle defect evolution dataset with clear maintenance labels, use statistical regression analysis to determine the critical confidence points that distinguish different fault evolution stages, and use these critical confidence points as the speed division boundary and acceleration division boundary to clearly define different stages of the equipment, such as the steady-state tolerance zone, the early warning zone, and the degradation zone.

[0060] The calculated drift velocity parameters are numerically compared with the velocity classification boundaries to output a velocity level index. Specifically, the scalar value of the drift velocity parameter is checked to see which classification boundary interval it falls within (e.g., determining whether the velocity parameter exceeds the steady-state tolerance boundary), and a corresponding level code is assigned to it, forming a velocity level index. This index intuitively represents the rate level at which the equipment state deviates from the normal baseline.

[0061] Similarly, the drift acceleration parameters are numerically compared with the acceleration boundary to output an acceleration level index. By determining the numerical distribution characteristics of the acceleration parameters (e.g., positive values ​​represent the acceleration escape period, and negative values ​​represent fault mitigation), corresponding level codes are assigned to them, forming an acceleration level index. This index profoundly reflects the trend and urgency of equipment condition deterioration.

[0062] The acquired velocity level index and acceleration level index are used as two-dimensional addressing coordinates and input into a preset hierarchical diagnostic matrix for coordinate addressing. This hierarchical diagnostic matrix is ​​pre-configured with state interpretations for different combinations of velocity and acceleration levels. Through cross-mapping and precise positioning of the index coordinates, it outputs evolutionary diagnostic labels reflecting the current device health trajectory (such as "early fault acceleration evolution label" or "fault mitigation attenuation label"). Specifically, the hierarchical diagnostic matrix in this embodiment is a preset 3×3 two-dimensional mapping table. The velocity dimension is divided into three levels: "tolerance range," "warning range," and "danger range" (indices 1-3); the acceleration dimension is divided into three levels: "negative convergence," "zero drift stability," and "positive acceleration" (indices 1-3). The matrix coordinate mapping rules are as follows: coordinates (1,1) (1,2) (2,1) are mapped to "healthy and stable label"; coordinates (1,3) (2,2) (3,1) are mapped to "early fault mitigation label"; coordinates (2,3) (3,2) are mapped to "early fault accelerated evolution label"; coordinates (3,3) are mapped to "critical fault alarm label".

[0063] When the output evolutionary diagnostic label indicates that the distribution cabinet status has entered a preset abnormal range, the alarm signal is captured, and the feature vector sequence stored in the current sliding observation window and the corresponding updated network parameters are extracted simultaneously. For this state, instead of simply providing an alarm, the Shapley interpretation algorithm is used to perform a game-theoretic calculation of permutations and combinations on the feature vector sequence to determine the proportion of responsibility of each physical monitoring dimension for leading to the abnormal diagnostic conclusion. In the game-theoretic calculation, the participants are set as data features, datasets, or trained model components.

[0064] First, the marginal contribution expectations representing dimensions such as cabinet temperature, chamber pressure, partial discharge amplitude, and phase current within the feature vector sequence are exhaustively quantified. When calculating the utility value of a sub-combination, the model parameters corresponding to a specific participant combination are aggregated, and the masked feature sequence of that combination is input into the model loaded with updated network parameters for forward model derivation. For masked feature dimensions in the sequence, their historical mean sequences within the current sliding observation window are padded and aligned to maintain tensor dimension consistency and eliminate step noise. Based on this, the true utility estimate of the participant combination, i.e., the confidence level for abnormal diagnosis, is deduced and output. For the aggregated historical combination utility values, the values ​​from multiple derivations are extracted, and the lookup table is mapped and updated using calculations of the arithmetic mean, maximum extremum, or minimum extremum.

[0065] Based on the above utility lookup table, the contribution values ​​of independent features are obtained by substituting them into the preset Shapley algorithm core derivation formula: In this formula framework, The independent feature contribution value corresponding to the i-th feature dimension participant in the final output is represented by the absolute weight share of the specific feature in causing the model to deviate from the healthy state judgment; N represents the total set of all possible feature dimensions that may participate in the construction of the input; |N| is the total number of feature dimensions; S represents any feature subset extracted after removing the target dimension i; |S| is the number of dimensions contained in the subset; and These represent the true utility values ​​of the model derived from the feature mapping network in two cases: when feature i is included in the input features and when it is masked. The difference between the two values ​​obtained by subtraction directly quantifies the marginal contribution of feature i. The factorial combination fractional term acts as a weighting coefficient, ensuring that the marginal contribution is distributed unbiasedly according to the expected value across all permutations of the input order.

[0066] In this embodiment, the signal-to-noise ratio (SNR) sequence of each physical sensor within the historical observation window is extracted. A confidence mapping model based on the arctangent function is constructed to calculate the dimension confidence weights. These weights are then multiplied by the initial feature contribution values ​​to output the optimized feature contribution values. Specifically, after obtaining the initial feature contribution values ​​for each dimension, the SNR evaluation parameters of the corresponding physical sensors within the sliding observation window are extracted. Based on these parameters, a data confidence mapping model is constructed. Its numerical function is defined as follows: the confidence weight is equal to the arctangent of the difference between the current SNR and the preset standard SNR. The calculated result is divided by the constant pi, and the quotient is increased by a constant 0.5. The preset standard SNR is set to 20 dB according to the recommended communication and anti-interference lower limit for measurement equipment in IEC 61850 smart substations. When the actual sensor SNR is significantly lower than 20 dB (e.g., due to extreme electromagnetic crosstalk), the confidence weights will be compressed to near 0 through this arctangent model to prevent misleading Shapley marginal contribution assessments due to high noise. When the SNR is higher than 20 dB, the weights smoothly converge to 1. Subsequently, the confidence weights of each dimension obtained from the mapping are multiplied by the original feature contribution values ​​output by the Shapley interpretation algorithm, thereby outputting the reconstructed and optimized feature contribution values. This process effectively mitigates spurious high contribution weights caused by sensor polarization or short-term electromagnetic crosstalk.

[0067] The above-mentioned expected extrapolation calculations are performed for each data acquisition dimension. The absolute values ​​of the independent feature contribution values ​​output by each dimension are compared, and a contribution ranking list is generated by sorting the values ​​in descending order. Then, the primary influencing dimension parameter at the top of the list is extracted. Finally, the character reverse mapping rule (i.e., a string conversion dictionary from physical parameters to system codes) preset in the database is retrieved, and the extracted primary influencing dimensions are converted and concatenated to output a target feature identifier with a clear direction. The character reverse mapping rule is a key-value pair dictionary that hard-codes physical monitoring parameters to standard system fault types (e.g., mapping the "PT100 cabinet internal temperature" parameter to the "T_FAULT_thermal fault" identifier, and mapping the "UHF partial discharge amplitude" parameter to the "PD_FAULT_insulation fault" identifier).

[0068] By introducing the Shapley algorithm to perform permutation and combination game deduction, the marginal contribution expectation of independent dimensions such as temperature and partial discharge is objectively quantified, the source of the primary influencing dimension is traced, and the early warning signal is given solid algorithmic interpretability support.

[0069] The system acquires and parses the internal hardware structure drawings of the distribution cabinet; extracts the electrical connection and thermal conduction attributes of circuit breaker components, disconnector components, and cable head components, and outputs multi-dimensional topological relationships; constructs a physical topology map using these relationships; configures the vertices of the physical topology map as hardware entities and the connections as entity interaction energy flows; inputs the target feature identifier into the physical topology map to trigger topology path optimization calculation and outputs sensitive association paths; searches for highly sensitive vertices within the physical topology map along these sensitive association paths; maps these highly sensitive vertices to generate abnormal component nodes; extracts the embedded spatial coordinates and hardware model of the abnormal component nodes and packages them into a maintenance work order.

[0070] Specifically, the process for generating a repair work order is as follows: After obtaining the target feature identifier, the final physical entity mapping stage begins. A 3D spatial equipment drawing structural model matching the internal hardware model of the distribution cabinet is retrieved and parsed from the substation digital twin or asset management database. Semantic extraction algorithms are used to separate the node attributes of various equipment entities, including circuit breaker components, disconnector components, and cable head components. Simultaneously, electrical connection attributes (covering busbar contact relationships and conductor connections) and thermal conduction attributes (covering air convection and solid metal heat conduction paths) between these nodes are extracted, and the integrated data is output as a multi-dimensional topology.

[0071] The multidimensional topological relationships are instantiated into a physical topology graph for storage and computation. The graph is defined as follows: the vertex structure of the graph is configured as the specific hardware entities inside the distribution cabinet, and the connection structure between the vertices is configured as the energy flow of entity interaction. The specific selection method for the weights of each connection is as follows: the weights of the electrical network topology are assigned based on the contact resistance reference value or spatial distributed capacitance value extracted from the equipment manufacturer's technical specifications; the weights of the thermodynamic network topology are assigned based on the spatial physical geometric spacing and the thermal conductivity parameter of the connecting medium, using the inverse proportional conversion of Fourier's law of thermal conductivity to obtain the thermal impedance value.

[0072] Subsequently, the target feature identifiers generated previously are input into the physical topology graph to trigger topology path optimization calculation. Based on the type of physical parameters contained in the target feature identifiers (e.g., electrical or thermal), the matching edge weight layer network is automatically switched. The optimization calculation uses Dijkstra's shortest path algorithm as the basis for diffusion reverse tracing. Starting from the global endpoint in the matched graph network layer, it performs reverse deduction and node iterative scanning along the branch connection with the minimum set impedance weight until the associated domain pointed to by the identifier is covered, thus outputting the sensitive associated path.

[0073] Following the output sensitive correlation path, the connection convergence degree and energy dissipation weight of each entity on the path are compared to search for highly sensitive vertices within the graph where convergence and divergence indicators are at their peak. After confirming the coordinates of this vertex, it is mapped to generate an abnormal component node with clear physical semantics. Finally, the embedded 3D coordinate numerical parameters of this abnormal component node and the factory hardware model parameters bound to it are extracted from the database. The abnormal component node name, 3D coordinate values, hardware model parameters, and evolutionary diagnostic tag content generated in the previous steps are uniformly encapsulated into a standardized dispatch template data stream to generate and output a maintenance work order for direct use by the maintenance team.

[0074] By integrating electrical connection and thermal conduction properties to construct a topology map, the algorithm is guided to perform path optimization in the energy flow of entity interaction, so that the evolution of the front-end algorithm is transformed into maintenance guidance that is accurate to specific components.

[0075] The specific implementation details of the offline foundation construction and pre-training stage of the preset feature mapping network in this embodiment are as follows: Regarding the architecture construction of network models: In this embodiment, the preset feature mapping network adopts a deep concatenated architecture that integrates a one-dimensional convolutional neural network and a multilayer perceptron. Its bottom-level feature extraction module contains at least two one-dimensional convolutional layers and pooling layers. The temporal receptive field length of the one-dimensional convolutional kernel is strictly bound to the sampling frequency and feature period of the power distribution cabinet's physical sensors. Specifically: for high-frequency partial discharge monitoring data with a sampling rate of 10kHz, the convolutional kernel length is set to cover the time width of a typical partial discharge pulse (e.g., 20 to 50 sampling steps); for slowly changing PT100 temperature data, dilated convolutions with increased strides are used to cover at least a 5-minute heat conduction time window. The aforementioned multi-source physical data are then mapped to 16 to 32 feature channels to capture the local temporal correlation of the purification feature vector sequence in the time dimension. Its top-level mapping module contains at least three fully connected hidden layers. Each hidden layer neuron uses the ReLU nonlinear activation function, and a Softmax output layer is connected after the top-level connection parameters to output the probability distribution corresponding to each physical fault classification.

[0076] Massive monitoring data of the same model of distribution cabinet throughout its entire lifecycle were retrieved. Based on expert maintenance logs, physical data was manually extracted, including the healthy and stable operation phase, as well as typical fault evolution stages with clear physical characteristics such as the initial stage of partial discharge, poor contact and overheating, and gas chamber leakage. After performing the same decoupling, purification, and normalization processing as described above on the above data, samples were extracted by time sliding window, and each sample was assigned a corresponding state diagnosis classification label (i.e., the parameter y in the aforementioned Fisher information formula), thereby constructing a supervised learning dataset with high confidence.

[0077] The constructed dataset is randomly divided into training, validation, and test sets in a preset ratio of 7:2:1. The training set is used to calculate the gradient direction of the network parameters and perform weight iteration during the pre-training phase. The validation set does not directly participate in weight updates but is used to evaluate the model's generalization error after each training round to assist in adjusting the learning rate, determining early stopping mechanisms, and establishing the optimal value of the elastic penalty coefficient in the joint optimization function. The test set is used for final simulated real-world network environment testing before the model is fully deployed, ensuring that the model's basic accuracy meets the target.

[0078] Training set samples are extracted and injected in batches into the uninitialized network. Forward propagation network derivation is performed, and the predicted state probabilities are output. The standard cross-entropy function is used as the core loss function to quantify the distribution difference between the output predicted probabilities and the real human labels. Subsequently, the partial derivatives of the loss function with respect to each weight matrix and bias vector in the network are calculated using the backpropagation algorithm, and the parameter space is gradually corrected along the direction of gradient descent.

[0079] During the pre-training phase, an adaptive moment estimation optimizer is used to perform parameter updates. The batch size for a single training session is configured to be 64 or 128; the initial learning rate is set to 10. -4 The system is configured to automatically reduce the learning rate by a decay factor of 0.1 if the validation set loss does not decrease for five consecutive epochs; the global maximum number of training epochs is set to 500. When the loss function of the pre-trained network on the validation set converges and the accuracy no longer improves, the network parameters are locked. At this point, the network distribution matrix is ​​extracted as the aforementioned "preset baseline data," and the corresponding Fisher information matrix is ​​calculated and persistently saved, thus providing a solid foundation for subsequent long-term online weight consolidation and evolution warning.

[0080] This invention provides a method for identifying power distribution cabinet faults based on big data analytics. By introducing a transfer function model at the data source to subtract environmental offsets, it eliminates underlying interference caused by climate fluctuations and restores the true physical operating status of the equipment. Furthermore, this method not only dynamically updates the network using a weight consolidation mechanism when encountering data distribution drift, but also projects temporal features into a high-dimensional space, using velocity and acceleration derivatives to characterize the fault evolution trajectory. Finally, by combining feature contribution evaluation and physical entity topology mapping, a smooth transition from black-box algorithm warning to tracing specific abnormal components is achieved. This global design effectively enhances the tracking and warning capabilities under long-term service, enabling operation and maintenance management to understand the hidden degradation process of equipment.

[0081] Example 2 This embodiment describes the specific implementation process of a power distribution cabinet fault identification method based on big data analysis applied to power distribution cabinet fault identification in an industrial park.

[0082] The system uses built-in temperature sensors, barometers, high-frequency partial discharge monitors, and multi-phase current transformers to collect real-time physical data for up to 30 months. Simultaneously, it utilizes an external weather station to obtain temperature, humidity, and atmospheric pressure parameters around the substation. This monitoring comprehensively recorded the continuous operational status of a main power supply high-voltage switchgear.

[0083] Because extreme cold weather leads to extremely low external environmental baselines, historical physical data and external environmental parameters are first normalized. Then, a least-squares algorithm is used to fit the correlation response matrix, constructing a transfer function model to extract the thermal hysteresis constant and gain coefficient. The current dimensionless environmental parameters are input into this model to deduce the environmental offset value. This value is then subtracted from the current operating parameters, and the model is filtered using a Gaussian distribution with a threshold of three standard deviations to remove outliers. Finally, a pure feature vector, unaffected by the severe interference of the cold and dry climate, is output.

[0084] A sliding observation window is set to continuously accumulate multiple sets of purification feature vectors, constructing a measured distribution matrix. This matrix, along with preset baseline data from the initial stage of equipment operation, is projected into a high-dimensional space containing the Gaussian radial basis function kernel. The feature mean values ​​of the two sets of data in the high-dimensional space are extracted, and the maximum mean difference is calculated to obtain the divergence deviation value. When this value exceeds the preset probability density upper limit threshold for five consecutive monitoring cycles, it is determined that the equipment operating status has significantly deviated.

[0085] To adapt to the evolving characteristics of new cold operating conditions without forgetting the initial normal state, the low-level feature extraction connections of the mapping network are frozen, and the Fisher information matrix left over from historical training is retrieved to evaluate the importance of the paths of the top-level parameters to be tuned. A joint optimization function containing cross-entropy loss and an adaptive penalty term is constructed, with the penalty coefficient finely set to 2000. Coverage updates are performed only on approximately 10% of the parameters with extremely low importance, thereby generating updated network parameters.

[0086] The time-ordered feature sequences are input into the hidden layers of the network for dimensionality-upgrading mapping, forming an initial trajectory point set in a multi-dimensional state space. The Mahalanobis distance distribution density of each node relative to the centroid is calculated, and detached interference nodes with distances greater than the mean plus twice the standard deviation are filtered out. A weighted moving average algorithm is used to output a smooth temporal trajectory point set. The drift velocity is obtained by measuring the spatial displacement of adjacent nodes and dividing by the time step; further, the drift acceleration parameter is obtained by calculating the velocity evolution difference and dividing by the time step.

[0087] When the dynamic parameters fall into the deterioration range, the average marginal contribution of each individual joining the combination is calculated based on the Shapley game principle to ensure a fair allocation of responsibility. Calculations confirm that the air chamber pressure dimension has the most significant independent feature contribution in this state migration due to pressure fluctuations exceeding 10 kPa. Therefore, the air pressure parameter is extracted as the primary influencing dimension and a target feature identifier is generated.

[0088] The 3D spatial structure drawing of the power supply cabinet was analyzed to extract the physical topology map of the circuit breaker, disconnector, and insulating gas chamber, and the connections were configured with airflow impedance weights. The identifiers were input into the map for reverse impedance optimization and tracing, precisely locating the sealing valve behind the insulating gas chamber, which exhibited a slight deformation (estimated offset of 5 mm). The spatial coordinates and Chinese model specifications of the valve were automatically packaged to generate the first highly sensitive maintenance work order, accurately guiding maintenance workers to the site to perform the seal replacement work.

[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying faults in power distribution cabinets based on big data analysis, characterized in that, include: The system collects operational physical data including cabinet temperature, chamber pressure, partial discharge amplitude, and phase current, as well as external environmental parameters including temperature, humidity, and atmospheric pressure. The external environmental parameters are input into the transfer function model to obtain environmental offset values. The operational physical data are normalized and the environmental offset values ​​are subtracted to generate a purification feature vector. The purified feature vector is compared with the preset benchmark data to calculate the divergence deviation value; when the divergence deviation value exceeds the preset deviation threshold, the weight consolidation mechanism is triggered to adjust the feature mapping network and output the updated network parameters. The feature vectors are concatenated, sorted, and purified to generate a sequence of feature vectors. The feature vector sequence and updated network parameters are mapped in a multidimensional way to output a time-series trajectory point set; the velocity derivative and acceleration derivative of the time-series trajectory point set are calculated to generate drift velocity parameters and drift acceleration parameters. Input the drift velocity parameters and drift acceleration parameters into the preset hierarchical diagnosis matrix and output the evolution diagnosis label; when the evolution diagnosis label is determined to enter the abnormal range, calculate the feature contribution value and extract the target feature identifier based on the feature vector sequence and updated network parameters; map the target feature identifier, match the abnormal component node, and output the maintenance work order.

2. The method for identifying power distribution cabinet faults based on big data analysis according to claim 1, characterized in that, The specific calculation process of the environmental offset value includes: retrieving historical operational physical data and historical external environmental parameters, performing normalization, and outputting a dimensionless feature set; using an identification algorithm to fit the correlation response matrix based on the dimensionless feature set, and outputting a set of gain coefficients and a thermal hysteresis constant; constructing a transfer function model using the set of gain coefficients and the thermal hysteresis constant; normalizing the temperature, humidity, and atmospheric pressure variables within the external environmental parameters, and outputting dimensionless parameters; inputting the dimensionless parameters into the transfer function model, and outputting basic response parameters; performing amplitude modulation on the basic response parameters based on the set of gain coefficients and time shifting based on the thermal hysteresis constant, and outputting a same-dimensional offset vector; using the same-dimensional offset vector as the environmental offset value; extracting operational physical data and performing normalization processing, and outputting dimensionless operational parameters; performing matrix subtraction between the dimensionless operational parameters and the environmental offset value, and outputting a benchmark difference matrix; performing extreme value anomaly screening on the benchmark difference matrix, and outputting a purification feature vector.

3. The method for identifying power distribution cabinet faults based on big data analysis according to claim 1, characterized in that, The specific calculation process of the divergence deviation value includes: extracting the distribution matrix of the training feature mapping network as the preset benchmark data; setting a sliding observation window to collect and accumulate multiple sets of cleaned feature vectors to construct the measured distribution matrix; using a preset kernel function to calculate the spatial inner product of the measured distribution matrix and the preset benchmark data, and outputting a high-dimensional kernel matrix; calculating the empirical mapping expectation of the measured distribution matrix and the preset benchmark data based on the high-dimensional kernel matrix, and outputting the measured high-dimensional mean and the benchmark high-dimensional mean; finding the maximum mean difference between the measured high-dimensional mean and the benchmark high-dimensional mean, and outputting the divergence deviation value.

4. The method for identifying power distribution cabinet faults based on big data analysis according to claim 1, characterized in that, The specific process for generating updated network parameters includes: when the divergence deviation exceeds a preset deviation threshold, an out-of-bounds instruction is output; in response to the out-of-bounds instruction, the preset feature mapping network bottom-level parameters are frozen, and the top-level connection parameters are opened as network parameters to be adjusted; the Fisher information matrix retained during the historical training phase is retrieved; the Fisher information matrix is ​​used to evaluate the importance of the path weights of each neuron within the network parameters to be adjusted, and a path importance index is output; a joint optimization function including cross-entropy loss and elastic regularization term is constructed; the path importance index is used to control the constraint strength of the elastic regularization term, and an adaptive penalty function is output; the adaptive penalty function is applied to the cleaned feature vector to perform gradient descent on the network parameters to be adjusted, and temporary iterative parameters are output; the gradient descent algorithm is used to perform numerical update operations on the temporary iterative parameters associated with path importance indices that are less than a preset importance value benchmark, and updated network parameters are output.

5. The method for identifying power distribution cabinet faults based on big data analysis according to claim 1, characterized in that, The specific generation process of the time-series trajectory point set includes: constructing a feature vector sequence by arranging multiple sets of purified feature vectors in series according to the order of acquisition time; injecting the feature vector sequence into the hidden layer nodes of the network driven by updating network parameters; performing latent space dimensionality upscaling mapping on the feature vector sequence using the hidden layer nodes of the network to obtain the feature dimensionality upscaling result; projecting the feature dimensionality upscaling result onto a preset high-dimensional feature state space to generate an initial trajectory point set; calculating the Mahalanobis distance distribution density of each node in the initial trajectory point set and outputting a node density scalar; filtering isolated outlier interference nodes based on the node density scalar and outputting a purified high-dimensional point set; and smoothing the purified high-dimensional point set to output the time-series trajectory point set.

6. The method for identifying power distribution cabinet faults based on big data analysis according to claim 1, characterized in that, The specific calculation process for the drift velocity parameters and drift acceleration parameters includes: measuring the difference in high-dimensional spatial positions of adjacent time nodes within the time series trajectory point set, and outputting a coordinate displacement difference vector; dividing the coordinate displacement difference vector by the time interval step size to obtain the vector derivative, and outputting the velocity vector parameters; extracting the first and last nodes of the time series trajectory point set to construct a global trend vector; calculating the cosine value of the angle between the global trend vector and the velocity vector parameters, and outputting the direction correction coefficient; adjusting the magnitude of the velocity vector parameters using the direction correction coefficient multiplication, and outputting the drift velocity parameters; measuring the change in the difference of the velocity vector parameters between adjacent time nodes, and outputting the velocity evolution difference; dividing the velocity evolution difference by the time interval step size to perform a second vector derivative calculation and outputting the drift acceleration parameters.

7. The method for identifying power distribution cabinet faults based on big data analysis according to claim 1, characterized in that, The specific extraction process of the target feature identifier includes: extracting the embedded numerical boundaries of the preset hierarchical diagnosis matrix, and outputting the velocity division boundary and acceleration division boundary; performing numerical comparison between the drift velocity parameter and the velocity division boundary, and outputting the velocity level index; performing numerical comparison between the drift acceleration parameter and the acceleration division boundary, and outputting the acceleration level index; inputting the velocity level index and the acceleration level index into the preset hierarchical diagnosis matrix for coordinate addressing, and outputting the evolutionary diagnosis label; determining that the evolutionary diagnosis label falls into the abnormal range, and outputting an abnormal alarm signal; responding to the abnormal alarm signal, extracting the feature vector sequence within the sliding observation window and updating the network parameters, and outputting the interpretation and deduction base; performing game-theoretic calculation on the interpretation and deduction base using the Shapley interpretation algorithm, and outputting the game distribution matrix; quantifying the marginal contribution expectation of the dimensions representing the cabinet temperature, air chamber pressure, partial discharge amplitude, and current of each phase according to the game distribution matrix, and outputting the feature contribution value; comparing the size of the feature contribution values ​​of each dimension, and outputting the contribution ranking list; extracting the primary influence dimension at the top of the contribution ranking list; and calling the preset reverse mapping rule to transform and assemble the primary influence dimension to output the target feature identifier.

8. The method for identifying power distribution cabinet faults based on big data analysis according to claim 1, characterized in that, The specific process for generating the maintenance work order includes: acquiring and parsing the internal hardware equipment drawings of the distribution cabinet; extracting the electrical connection and thermal conduction attributes of circuit breaker components, disconnector components, and cable head components, and outputting multi-dimensional topological relationships; constructing a physical topology map using the multi-dimensional topology relationships; configuring the vertices of the physical topology map as hardware entities and the connections as entity interaction energy flows; inputting the target feature identifier into the physical topology map to trigger topology path optimization calculation and outputting sensitive association paths; searching for highly sensitive vertices within the physical topology map along the sensitive association paths; mapping the highly sensitive vertices to generate abnormal component nodes; extracting the embedded spatial location coordinates and hardware model of the abnormal component nodes and packaging them to output the maintenance work order.

Images