Distribution cable branch box state monitoring method based on state identification

Abstract

The invention discloses a distribution cable branch box state monitoring method based on state recognition, and particularly relates to the technical field of power monitoring, and the method comprises the steps: collecting multi-dimensional operation data of a plurality of branch boxes in a continuous time period, and constructing a state evolution sequence; calculating a state offset score and an adjacent equipment state consistency score to judge whether an evaluation process is triggered or not; after triggering, constructing a state influence probability map, identifying an abnormal influence path, extracting a state disturbance diffusion index and a cooperative behavior deviation index, inputting into a pre-trained risk identification model, generating a risk level interval and a cause probability distribution vector, executing an influence regulation and control measure, and updating a state identification logic; according to the method, dynamic perception of the state of the branch box is realized by constructing a state evolution sequence, combined judgment of individual and group behaviors is realized by combining a state offset score and an adjacent equipment state consistency score, and a risk level interval and cause probability distribution vector are used for driving regulation and control strategies and map updating. And the monitoring precision and the self-adaptive capability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power monitoring, and more specifically, to a method for monitoring the state of a distribution cable branch box based on state recognition. Background Art

[0002] As a key node in the medium-voltage distribution network, the distribution cable branch box undertakes important functions of power distribution and transfer in urban distribution networks, industrial parks, rail transit, and large infrastructure power supply systems. Due to its wide geographical distribution, complex structural connections, and often being in outdoor or semi-enclosed environments, the branch box is extremely vulnerable to the combined effects of factors such as temperature and humidity fluctuations, cable aging, abnormal current carrying, environmental interference, and poor equipment contact during long-term operation, resulting in the gradual accumulation of hidden faults such as partial discharge, insulation degradation, and loose electrical connections, and ultimately triggering system-level faults or cascading tripping events.

[0003] Traditional methods for monitoring the state of branch boxes mainly rely on periodic manual inspections, infrared temperature measurement, and discrete data acquisition devices to obtain the state. These methods have the following obvious deficiencies: Response lag and coverage blind spots: Manual inspections are difficult to achieve continuous monitoring of all equipment, and state changes can only be recorded after an event occurs, unable to capture abnormal behaviors during the latent period; Single-point threshold monitoring model: Existing online monitoring systems mostly adopt an alarm strategy based on fixed thresholds to detect overlimits of single physical quantities such as current, voltage, temperature, humidity, or partial discharge, but it is difficult to adapt to the natural fluctuation characteristics of different equipment under different seasons and load conditions, resulting in false alarms or missed alarms; Lack of perspective on behavior evolution and causal reasoning ability: Current technologies usually judge the collected data based on the "single-point state at the current moment", ignoring the trend information of equipment state evolution over time, and failing to establish a state conduction mechanism between devices, unable to identify the propagation and regional coordination of abnormal states.

[0004] In recent years, with the development of the power Internet of Things and edge computing technologies, more and more branch boxes have been deployed with multi-dimensional sensing terminals, which can collect various state data including current, voltage, temperature and humidity, and partial discharge in real time, forming a large amount of device operation time-series data. This provides a basis for modeling and intelligent recognition based on the change trajectory of device states. However, how to effectively extract representative behavior features from these high-dimensional time-series data, identify the evolution trend of abnormal states, and establish a causal model across devices in combination with a distributed structure is still one of the technical difficulties in the current industry. Therefore, the present invention proposes a method for monitoring the state of a distribution cable branch box based on state recognition in order to solve the above problems. Summary of the Invention

[0005] To achieve the above object, the present invention provides the following technical solutions: A method for monitoring the state of a distribution cable branch box based on state recognition, comprising the following steps: Collect the current, voltage, temperature, humidity and partial discharge data of multiple branch boxes within a continuous time period, and combine with historical data to obtain the state evolution sequence of each branch box; Based on each state evolution sequence, calculate the state deviation score and the adjacent device state consistency score. If any score exceeds the corresponding preset threshold, trigger the state evaluation process; if neither exceeds, maintain real-time update and monitoring; After triggering the evaluation, construct a state influence probability map between devices based on the historical state event sequence, identify potential abnormal influence paths, and extract two types of indices from them: the state perturbation diffusion index, which is used to measure the intensity of the current state anomaly spreading to other devices through the causal path; the collaborative behavior deviation index, which is used to measure the difference degree of the time series characteristics between the current device behavior and the behavior of its neighborhood group; Input the state perturbation diffusion index and the collaborative behavior deviation index into a pre-trained risk recognition model to generate the risk level interval and the cause probability distribution vector of the current state, which are used to characterize the risk scenario type to which the current state belongs and the corresponding abnormal cause; Based on the generated risk level interval and the cause probability distribution vector, execute corresponding influence control measures according to the preset strategy, and iteratively update the causal path map in the state recognition logic according to the cause distribution.

[0006] In a preferred embodiment, the state evolution sequence includes the current eigenvalue, the change slope and the trajectory deviation degree, which are used to express the evolution trend of the device. The current eigenvalue is the original physical quantity data collected at the current time point, including the instantaneous values of current, voltage, temperature, humidity and partial discharge. The change slope is the change rate obtained by differentiating the current eigenvalues of two adjacent time points and dividing by the time interval, which is used to reflect the dynamic growth or decay trend of the physical quantity. The trajectory deviation degree is the calculation result of the distance between the current eigenvalue and the mean or median value of the device in the same historical time period, which is used to evaluate whether the current state of the device deviates from its long-term behavior trajectory. The change slope and the trajectory deviation degree are both numerically normalized in a standardized manner to ensure comparability.

[0007] In a preferred embodiment, the calculation of the trajectory deviation degree is implemented based on any one of Euclidean distance, cosine similarity or Mahalanobis distance.

[0008] In a preferred embodiment, the state deviation score is obtained by combining and calculating the current eigenvalue, change slope, and trajectory deviation in the state evolution sequence to obtain a specific value, which is used to evaluate the overall deviation degree of the current state relative to the historical behavior model. The current eigenvalue and change slope respectively represent the instantaneous state and the trend change, and the trajectory deviation represents the historical consistency in the time series. When calculating, the three are first normalized respectively to form a standard feature vector, and then the standard feature vector is matched and calculated with the set of historical normal state sample feature vectors of the device in the corresponding season and corresponding time period. The calculation method is to use the Mahalanobis distance function to evaluate the covariance matrix of the current standard feature vector and the historical feature distribution, and output the statistical distance of the current state deviating from the distribution center as the state deviation score. This score is a single value, which is used to represent the degree of deviation of the device state from its historical normal distribution. The higher the state deviation score value, the greater the difference between the current state and the historical stable state, indicating that there may be potential anomalies or trend evolution risks. The state deviation score is generated only based on the state evolution sequence of the device itself, which is targeted and traceable.

[0009] In a preferred embodiment, the adjacent device state consistency score is obtained by comparing and calculating the state evolution sequence of the current device with the state evolution sequences of multiple other branch boxes adjacent to it in terms of structure or physical distribution to obtain a specific value, which is used to measure the behavioral consistency between the current device state evolution trend and the neighboring devices. The state evolution sequence of each device includes the current eigenvalue, change slope, and trajectory deviation. The score calculation includes two steps: The first step is to select a set of adjacent devices that are directly connected in terms of spatial location or power supply path or whose physical distance is less than the set spatial threshold; The second step is to calculate the dynamic time warping distance of the state evolution sequences of the current device and each adjacent device in the same time period, and take the average distance value of all adjacent devices as the preliminary consistency measurement result. Then, take the reciprocal of this result and perform normalization processing to obtain the adjacent device state consistency score. The higher the value, the higher the consistency between the current device state and the neighboring device states. The lower the value, the more discrete the current device state evolution trend is relative to the adjacent group. This scoring mechanism can identify possible deviation behaviors in the local area and provide a horizontal comparison basis for whether to trigger the state evaluation process, ensuring that the identification process is analyzed not only based on individual characteristics but also combined with group structure behavior characteristics.

[0010] In a preferred embodiment, after triggering the evaluation, a state influence probability map between devices is constructed based on the historical state event sequence. The construction process includes the following steps: First, collect the state evolution sequences and their corresponding risk level intervals experienced by multiple branch boxes in different past time periods, and organize them into a set of historical state event sequences; Perform frequency analysis on the occurrence order of events in the set through statistical learning methods, construct a directed relationship structure based on the chronological relationship of state evolution times between different devices, and form an initial event association network; Adopt a conditional probability calculation method to evaluate the joint probability that adjacent devices have abnormal state deviation scores after a certain number of moments on the premise that the current device state deviation score exceeds the threshold, and establish a cross-device state causal association relationship; According to the association strength, i.e., the joint probability, screen the edge weights to form a sparse graph structure that only retains high-probability paths exceeding the preset probability standard value, which is the state influence probability map between devices. This map uses the state propagation probability as the edge weight and the historical event time series as the path basis, representing the set of paths through which a specific state evolution trajectory may spread from one device to other devices.

[0011] In a preferred embodiment, on the basis of constructing the state influence probability map between devices, the process of identifying potential abnormal influence paths includes the following operations: Taking the current device as the starting node, traverse all directed edges in the state influence probability map with this device as the source, screen the target nodes with edge weights higher than the preset propagation threshold, and form the first-layer affected node set; For each first-layer node, repeat the above screening operation, expand the influence link layer by layer outward until the traversal depth reaches the preset propagation step length or the propagation strength is lower than the dynamic termination condition, and form the entire high-probability state influence path set starting from the current device; During the identification process, each path records the device sequence and the corresponding propagation probability in its propagation link, sorts them by the path cumulative probability, and selects the top several paths as the output result of the final abnormal influence path. This process ensures that starting from a single-point state anomaly, the most likely multi-hop propagation link can be reasonably identified, providing a causal chain basis for extracting the state perturbation diffusion index later, and at the same time ensuring the convergence and determination accuracy of the influence path identification in a multi-device scenario with a complex structure.

[0012] In a preferred embodiment, after identifying the potential abnormal influence paths, the process of extracting the state perturbation diffusion index and the collaborative behavior deviation index from them includes the following steps: For the state perturbation diffusion index, an exponential expression model is constructed based on the propagation probability, the number of path hops, and the state perturbation intensity of each path in the abnormal influence path. The state perturbation diffusion index is defined as an exponential decay function of the perturbation intensity multiplied by the total path length, that is, the perturbation diffusion index is equal to the perturbation intensity multiplied by the path probability weighting coefficient, and then multiplied by the negative exponential function of the number of hops. The calculation formula is: D = S × (∑p_i × w_i) × exp(-λ × h), where D is the state perturbation diffusion index, S is the current state deviation score, p_i is the propagation probability of path i, w_i is the normalized weight of the state transfer factor in the path, h is the number of path hops, and λ is the path attenuation coefficient, and its value is set according to the experience of the device hierarchical structure; For the collaborative behavior deviation index, the state evolution sequences of multiple neighboring devices are regarded as distribution sequences, and a deviation measurement model based on the entropy weight distribution of time series features is constructed. The dynamic weighted information entropy between the current device state evolution sequence and the first-layer neighboring device sequence is calculated in three dimensions: the current eigenvalue, the change slope, and the trajectory deviation degree. The collaborative behavior deviation index is defined as: C = ∑(E_j × δ_j), where C is the collaborative behavior deviation index, E_j is the relative information entropy of the jth feature dimension, and δ_j is the mean square error between the current device and the neighboring device in this dimension. The sequences are calculated after normalizing each dimension using the historical mean. This method uses a composite modeling strategy of migrating "distribution discreteness and change amplitude" to express both behavior fluctuations and retain the correlation of the evolution characteristics between sequences. Both types of indices only depend on the obtained state evolution sequences, state deviation scores, and abnormal influence path parameters, and have high endogeneity, evolvability, and generalization ability.

[0013] In a preferred embodiment, the pre-trained risk identification model is a fuzzy logic controller, which receives the state perturbation diffusion index and the collaborative behavior deviation index as input variables, and uses a fuzzy inference mechanism based on the membership function to construct the mapping relationship from input to output, for generating the risk level interval and the cause probability distribution vector of the current state; Among them, the state perturbation diffusion index and the collaborative behavior deviation index are respectively defined as the input dimensions, and are divided into three hierarchical intervals of low, medium, and high through a three-segment membership function. Each input variable corresponds to three fuzzy linguistic terms, and nine input state rule combinations are formed by combining the input spaces. The fuzzy logic unit uses an inference engine based on a fuzzy rule table to map each combined state to an output of five risk levels. Among them, the risk level interval with the largest fuzzy membership value is used as the final judgment result. At the same time, according to the position of the fuzzy aggregation center point of the output layer, combined with the statistical frequency of the event cause labels in the training samples for each rule, a cause probability distribution vector is constructed, indicating the attribution probability of each type of abnormal cause in this input state. The risk level interval is used to describe the urgency of state evolution, and the cause probability distribution vector is used to identify which historical cause mechanisms the current state may be attributed to, so as to assist in the subsequent implementation of impact control measures and the update of the causal path map.

[0014] Technical effects and advantages of the present invention: By collecting the current, voltage, temperature, humidity, and partial discharge data of multiple branch boxes within a continuous time period, and combining historical data to construct the state evolution sequence of each device, the present invention realizes the time-continuous description of the key operating states. Compared with the traditional method mainly based on single-point static detection, the present invention introduces the time dimension of state evolution, enabling the system to capture the weak change signals gradually accumulated during the operation of the device, significantly improving the recognition sensitivity and depth of early abnormal trends, and enhancing the timeliness and warning ability of the distribution system operation monitoring.

[0015] The present invention not only calculates the state deviation score based on the state evolution sequence of a single device, but also synchronously evaluates the state consistency between other devices that are structurally or physically adjacent to it. By using the state consistency score of adjacent devices as a horizontal comparison index, a behavior consistency comparison framework between multiple devices is constructed. This mechanism can effectively identify whether a single device shows a deviation from the evolution trend in the group, avoiding false triggering caused by individual misjudgment or abnormal diffusion, thereby realizing the collaborative judgment of individual trends and group behaviors in the monitoring strategy, and significantly improving the accuracy and robustness of the monitoring results.

[0016] After identifying the state perturbation diffusion index and the collaborative behavior deviation index, the present invention inputs them into the pre-trained model to generate the risk level interval and the cause probability distribution vector of the current state, and then automatically executes the impact control measures according to the results, and iteratively optimizes the path structure established in the state recognition logic. This process forms a feedback closed-loop among state recognition, risk judgment, control execution, and model update, not only enabling the state monitoring to have an intelligent response ability, but also enabling the system to continuously evolve and optimize during actual operation, enhancing the adaptability and sustainability in dealing with various complex state evolution scenarios. Description of the Drawings

[0017] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of the distribution cable branch box status monitoring method based on status identification in the present invention. DETAILED DESCRIPTION

[0018] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] Reference Figure 1 The following examples were obtained: Example 1: A method for monitoring the state of a distribution cable branch box based on state recognition, comprising the following steps: Collect the current, voltage, temperature, humidity and partial discharge data of multiple branch boxes in a continuous time period, and combine the historical data to obtain the state evolution sequence of each branch box; this step is the basic data preparation process for the present invention to realize the state identification and monitoring function. By collecting real-time multi-dimensional state data of multiple branch boxes in a continuous operation cycle, covering key physical indicators such as current, voltage, temperature, humidity and partial discharge, the operating characteristics of the equipment and its microscopic electrical behavior can be fully reflected. Combined with the historical monitoring data of each branch box, time series alignment and data structure standardization are performed on a unified time scale to obtain a continuous multi-dimensional state evolution sequence, which is used to express the evolution trend of the equipment state over time in multiple physical dimensions. This state evolution sequence plays a core role in data source and feature expression in subsequent feature extraction, trend modeling and dynamic evaluation, ensuring the continuity and deducibility of state judgment.

[0020] Based on each state evolution sequence, calculate the state deviation score and the adjacent device state consistency score. If any score exceeds the corresponding preset threshold, trigger the state assessment process; if neither exceeds, maintain real-time update and monitoring. This step is to preliminarily determine whether the current state of the device is abnormal and decide whether to initiate a higher-level risk identification process. By mathematically modeling the state evolution sequence of each device, calculate the state deviation score reflecting the degree of deviation from the normal operation trajectory, and the adjacent device state consistency score measuring the similarity degree between its state and the evolution trend of adjacent devices. The two types of scores respectively start from the two dimensions of "individual behavior abnormality" and "group consistency disruption" to comprehensively judge whether there is a potential state abnormality. When any score exceeds the corresponding threshold, it indicates that the device has an abnormal evolution trend or is behaviorally disconnected from neighboring devices. At this time, trigger the assessment process; otherwise, it indicates that the current state is in a stable operation range, and the original monitoring strategy can be maintained. This mechanism has dynamics and pertinence, effectively balancing monitoring resources and computing efficiency.

[0021] After triggering the assessment, construct a state influence probability map between devices based on the historical state event sequence, identify potential abnormal influence paths, and extract two types of indices from them: the state perturbation diffusion index, which is used to measure the intensity of the current state abnormality spreading to other devices through causal paths; the collaborative behavior deviation index, which is used to measure the degree of difference in temporal characteristics between the behavior of the current device and the behavior of its neighboring group. This step is oriented towards the systematic level of state abnormality identification. After triggering the assessment, use the temporal relationship and joint probability of historical state events to construct a state propagation influence relationship map between devices, that is, the state influence probability map between devices. This map not only contains the paths through which the current device may affect other devices, but also represents the diffusion possibility of different paths with probability values. On this basis, further identify the abnormal influence paths with higher state propagation risks. Subsequently, extract two key assessment indices: the state perturbation diffusion index reflects the potential of the current state to cause a chain reaction to other devices and is a measure of the threat degree to system stability; the collaborative behavior deviation index measures whether the behavior characteristics of the current device are significantly deviated from its neighboring group, which helps to identify potential mismatch or individual mutation behaviors. These two indices provide a quantitative basis for subsequent risk level determination.

[0022] The state perturbation diffusion index and the collaborative behavior deviation index are input into the pre-trained risk identification model to generate the risk level interval and the cause probability distribution vector of the current state, which are used to characterize the risk scenario type to which the current state belongs and the corresponding abnormal causes; this step introduces an intelligent decision-making mechanism to output the state recognition result in a structured form. The pre-trained risk identification model is constructed based on a large number of previous state data and actual fault events, and has the ability to establish a mapping relationship between the input index and the risk level and causes. When the state perturbation diffusion index and the collaborative behavior deviation index are used as inputs, the model calculates the corresponding risk level interval through a fuzzy inference mechanism and further derives the cause probability distribution vector. The risk level interval can be used to determine which category of normal, minor, moderate, severe or emergency state the current state risk belongs to, while the cause probability distribution vector reflects the weight distribution of various risk causes such as electrical failure, environmental impact or structural aging that may correspond to the current state. This result constitutes the core judgment output of intelligent monitoring.

[0023] Based on the generated risk level interval and the cause probability distribution vector, corresponding impact regulation measures are executed according to the preset strategy, and the causal path map in the state recognition logic is iteratively updated according to the cause distribution. This step realizes the closed-loop linkage between the risk identification result, the monitoring strategy and the operation and maintenance behavior. According to the different levels of the risk level interval and the main cause types reflected by the cause probability distribution vector, the preset strategy will point to different regulation behaviors respectively, such as increasing the data sampling frequency, preferentially dispatching maintenance work orders, activating adjacent devices to strengthen monitoring, etc. In addition, the cause types in the current recognition result can also be used to feedback and correct the causal path map used by the state recognition model, that is, to adjust the weights of the influence chain, so as to improve the accuracy of the model in recognizing future similar states and realize the adaptive evolution of the model. This update mechanism ensures the continuous optimization of the monitoring logic and forms the ability of continuous learning and risk adaptation.

[0024] In the present invention, to realize the dynamic recognition and evolutionary monitoring of the operating state of the distribution cable branch box, it is first necessary to construct a state evolution sequence that can truly reflect the operating trend of the device. The state evolution sequence refers to a data sequence formed by arranging multiple physical state characteristics of the branch box in chronological order within a continuous time period, which is used to express the change trend and behavior trajectory of the device in multi-dimensional indicators. The state evolution sequence includes three types of core time series feature data, namely the current feature value, the change slope and the trajectory deviation degree. These three types of features are uniformly constructed into a multi-dimensional time series feature structure, which are jointly used to express the evolution trend of the device.

[0025] Among them, the current eigenvalue refers to the original physical quantity value collected for each monitoring index at the current monitoring time point. The monitoring indexes involved in the present invention include but are not limited to current, voltage, temperature, humidity, and partial discharge intensity, all of which are physical quantities that can reflect the electrical and environmental operating states of the equipment in real time. Taking temperature as an example, the current eigenvalue is the temperature reading measured by the sensor at a specific moment, and the unit is usually degrees Celsius.

[0026] The change slope is used to quantify the growth or decline rate of a physical quantity of the device at two adjacent time points. Specifically, for each type of current eigenvalue, the change slope is calculated as follows: subtract the eigenvalue of the previous time point from the eigenvalue of the current time point, and divide by the sampling interval between the two time points to obtain the change rate of the physical quantity per unit time. For example, if the temperature of a branch box is 48°C at time t1 and 45°C at time t0, and the time interval between the two times is 10 minutes, then the temperature change slope is (48 - 45) / 10 = 0.3°C / min. This feature reflects the dynamic evolution trend of the state and can be used to identify risk features such as rapid temperature rise and sudden voltage drop.

[0027] The trajectory deviation degree is used to represent the deviation degree of the current state from the behavior pattern of the device in the historical normal state. The calculation process of the trajectory deviation degree needs to construct a "benchmark behavior model" based on the state data collected in the historical same-type time period. The so-called "same-type time period" can include historical records in dimensions such as the same season, similar environmental load, and similar operating duration. When comparing the current eigenvalue with the historical sample set, one of the following three methods can be selected for modeling the deviation degree calculation: Euclidean distance: Calculate the straight-line distance between the current feature vector and the historical mean vector in the multi-dimensional space; Cosine similarity: Reflect the consistency of the evolution direction through the cosine value of the angle between vectors. The smaller the angle (i.e., the larger the cosine value), the smaller the deviation; Mahalanobis distance: A multi-variable distance metric that considers the covariance relationship between features, suitable for high-dimensional feature spaces with strong correlations, and can avoid the problem of error magnification caused by different distributions between features.

[0028] For example, during a low-load night period in winter, the normal current fluctuation range of a certain branch box is [70, 75] A, and the voltage is stable at [380, 385] V. If the current at the current moment is 85 A and the voltage is 390 V, the Euclidean distance can be used to calculate the distance between the current value and the historical vector formed by the above mean center, obtaining a numerical deviation degree. To improve the comparability of various indicators and avoid the influence of dimension and scale on the analysis results, both the change slope and the trajectory deviation degree are standardized, that is, the original values of different characteristic dimensions are subjected to mean zero-centering and standard deviation normalization transformation to make their distributions have a unified scale. The standardized results are used for subsequent scoring and model input to ensure the consistency and numerical stability of the state recognition model when processing multiple physical quantities. These three types of characteristics jointly form the state evolution sequence of the device. The sequence can be continuously updated in a time-window sliding manner and provides an input basis for each key step such as subsequent state deviation scoring, adjacent device consistency scoring, abnormal path recognition, and index generation.

[0029] State deviation scoring is an important quantitative parameter in the present invention for quantifying whether the individual operating state of a device deviates from its historical normal behavior trajectory. This scoring starts from the time dimension of a single device and is analyzed and calculated based on the state evolution sequence that has been constructed. The state evolution sequence already contains three key sub-items: the current characteristic value, the change slope, and the trajectory deviation degree. Each sub-item respectively represents the current instantaneous state, the change trend, and the long-term behavior consistency of the device.

[0030] Before calculating the score, it is first necessary to standardize these three characteristic dimensions. Standardization refers to converting the original physical quantity values (such as current, voltage, temperature, etc.) into dimensionless relative values, making the characteristics of different units and scales comparable within the same evaluation space. The standardization method uses mean centering and standard deviation normalization, that is, subtracting the historical mean from the original value and then dividing by the standard deviation to form a standard characteristic vector. Subsequently, the standardized characteristic vector at the current moment is matched and calculated with the characteristic distribution of the device in the historical normal state. The selection method for the set of historical normal state sample characteristic vectors is as follows: in the historical records with time similarity and load condition matching with the current operating environment, several standard sample data representing the "normal state" are selected. For example, the intervals with small fluctuations and stable scores in the device state records in the same season and the same operating time period are selected as the sample set. This set can represent the typical state behavior distribution of the device under normal circumstances.

[0031] In the matching calculation, the Mahalanobis distance function is used for evaluation. The Mahalanobis distance is a multivariate distance metric that takes into account the covariance relationship between various dimensional features and is applicable to the calculation of the deviation degree when there is a correlation between multiple features. The core idea of this function is to measure the distance between the standard feature vector and the mean vector of the historical sample set, and at the same time, weight and adjust the features of different dimensions according to the covariance matrix. The specific calculation formula is: ; where \(x\) is the current feature standard vector, \(\mu\) is the mean vector of the historical sample set, \(\Sigma\) is the covariance matrix of this sample set, represents its inverse matrix, \(T\) represents the transpose operation, and this statistical distance \(D\) is the current device state deviation score. The scoring result is a single value, indicating the degree to which the current device state deviates from its historical stable distribution in the multi-dimensional feature space. The larger the scoring value, the more the current state tends to be abnormal or unstable, and the more likely it is in an atypical operating state or the early stage of fault evolution.

[0032] For example, at a certain moment, the standard feature vector of a certain branch box is \([1.2, 0.8, 1.5]\), while the mean vector of its historical normal behavior is \([0, 0, 0]\), and the covariance matrix is the identity matrix. Then the calculated result of the Mahalanobis distance is approximately equal to 2.1, indicating that the current state has deviated significantly and may be near the warning state boundary. It should be noted that this scoring is completely based on the calculation of the device's own state evolution sequence and does not depend on the behavior of adjacent devices. Therefore, it has individual independence, traceability, and high sensitivity. In the subsequent state evaluation process, the state deviation score and the adjacent device state consistency score together constitute the trigger condition, ensuring that the state recognition considers both individual deviations and introduces a group behavior comparison mechanism to improve the intelligent recognition ability of the monitoring system.

[0033] The adjacent device state consistency score is an important indicator used to measure whether the state evolution trend of the current branch box device is synchronized or consistent with its neighboring devices during its operation. As a horizontal reference mechanism in the state recognition system, this scoring is mainly used to identify whether a device shows "behavioral anomalies" or "feature outliers" among group devices, and can effectively assist the state deviation score in multi-dimensional collaborative judgment, thereby improving the accuracy and robustness of the overall recognition.

[0034] The calculation of this scoring is based on the state evolution sequences of the current device and its adjacent branch boxes. The state evolution sequence has been defined above and consists of three types of features: the current feature value, the change slope, and the trajectory deviation degree, which are used to express the operation state change trajectory and evolution trend of the device within a certain period of time. The scoring calculation process includes the following two steps: Step 1: Determination of the adjacent device set. First, it is necessary to determine the adjacent objects of the current device. The establishment methods of the adjacent relationship include two types of criteria: Spatial location adjacency: It refers to other branch boxes that are less than a certain set spatial threshold away from the current device in the physical geographical space. For example, if the spatial threshold is set to 100 meters, other devices within a radius of 100 meters from the current device will be recognized as adjacent devices; Power supply path adjacency: It refers to other devices that have a direct cable connection relationship with the current device or belong to the same power supply link segment, forming an adjacency relationship in the electrical structure. By judging through any one of the above conditions or the combination of both, the set of adjacent devices of the current device can be formed for subsequent state comparison calculations.

[0035] Step 2: Calculation of state consistency. After determining the set of adjacent devices, it is necessary to evaluate whether the state evolution trends between the current device and each adjacent device are consistent. For this purpose, the state evolution sequences of each device within the same time window are selected for comparison, and the dynamic time warping distance (DTW) is used for difference calculation. DTW is a method for measuring the similarity between two time series under the condition of non-strict alignment of the time scale, which is applicable to the actual situation where there are sampling offsets or different rhythms between devices.

[0036] The specific approach is as follows: Take the state evolution sequence of the current device as the reference sequence, and calculate the DTW distance between it and the state sequences of adjacent devices one by one to obtain multiple distance values. Take the average of all distance values to get the average dynamic time warping distance between the current device and its adjacent group. This average value is used to represent the degree of consistency between the current device and the neighboring devices in terms of the multi-dimensional feature evolution trend. To make the scoring value consistent with the representation logic of "the more consistent, the higher the value", it is necessary to take the reciprocal of the above average distance value and then perform normalization processing to convert it into a scoring value between 0 and 1, which is the final state consistency score of adjacent devices. The closer the value is to 1, it represents a high degree of consistency in the state evolution between the current device and the neighboring devices; the lower the value, it represents that there are obvious differences in the evolution trajectory of the current device, and there may be local mutations, early anomalies or passive interferences, etc.

[0037] Application example: For example, the state evolution sequences composed of three types of features, namely current, temperature and partial discharge, collected by the current branch box A and the adjacent devices B, C, and D in the recent 15 minutes. The matching distances with B, C, and D are calculated as 1.2, 1.0, and 1.5 respectively through DTW. The average distance is 1.23. The reciprocal is 0.813. After normalization processing, the consistency score is about 0.85, indicating that the current state trend is generally synchronized with the neighborhood without obvious anomalies. The state consistency score of adjacent devices provides a horizontal comparison basis for whether to trigger the state evaluation process, and provides an important supplementary judgment when the state deviation score cannot independently determine anomalies; at the same time, this score strengthens the recognition ability of the present invention for sudden isolated events or regional anomalies in the actual engineering environment by introducing the idea of group behavior collaborative analysis, and improves the comprehensive intelligent level of the state recognition mechanism.

[0038] In the state recognition method of the present invention, in order to identify the potential propagation paths of abnormal states in the distribution cable branch box network, it is necessary to construct a state influence probability map between devices after the state evaluation is triggered. This map is used to express the historical probability structure of the propagation of specific state anomalies between devices and is one of the core bases for subsequent extraction of state perturbation diffusion indices, auxiliary risk determination, and generation of dynamic control strategies. The construction process of the state influence probability map includes the following public steps: Step 1: Construction of the historical state event sequence set. First, it is necessary to collect the state evolution data experienced by multiple branch box devices during past operation. The structure of the state evolution sequence has been defined above and includes three types of time features: the current eigenvalue, the change slope, and the trajectory deviation degree. On this basis, the evolution sequence within each time period is paired with the risk level interval corresponding to the system monitoring results of that time period to form a structured historical state event sample. The risk level interval is a classification output representing the risk degree of the device state in a specific time period, usually divided into five levels: normal, slightly abnormal, moderately abnormal, severely abnormal, and emergency state. These data are sorted along the time axis through an archiving method to form a sequence of "state-risk" pairs arranged in chronological order, constituting the historical state event sequence set. This set is used for statistical learning and causal path analysis.

[0039] Step 2: Construction of the initial event association network. Based on the historical state event sequence set, a statistical learning method is used to analyze the order of occurrence of state anomalies between devices. Specifically, after the state deviation score of a certain device increases (i.e., enters the abnormal range), it is statistically determined whether the adjacent device also has a similar deviation in a short period of time. If the frequency of occurrence of such joint events is relatively high in history, it indicates that there may be a causal relationship between the two. In a specific implementation, a directed edge is established to point the "earlier abnormal" device to the "later abnormal" device, forming an initial event association network based on the chronological order. In this network structure, nodes represent devices, and directed edges represent the chronological relationship of state influence. The existence of edges is only based on frequency statistics.

[0040] Step 3: Establishment of state causal association relationships. In order to further extract the true causal influence from statistical linkages, a conditional probability calculation mechanism is introduced based on the initial network. Specifically, the conditional probability of the following joint event is calculated: If the state deviation score of the current device A exceeds the threshold at time t (entering the abnormal state), then what is the probability that the state deviation score of the adjacent device B also exceeds the threshold at time t + Δt? This conditional probability can be obtained by counting the occurrence frequency of device B being abnormal under the premise condition (device A is abnormal) in all historical event sequences. This joint probability is used to quantify the possibility of abnormal propagation across devices. The core of this step is to map and supplement the previous chronological order into a probability-weighted structure, thereby realizing the transformation modeling from "correlation" to "causality".

[0041] Step 4: Screen edge weights to construct a sparse graph structure. After obtaining the joint probabilities between all devices, assign the corresponding conditional probability to each directed edge as the edge weight, which represents the probability value of the state propagating from the source node to the target node.

[0042] To improve the recognition efficiency and model sparsity, only retain the edges with joint probabilities exceeding a certain preset probability standard value (such as 0.3 or 0.5), and delete the remaining edges. The resulting directed graph is the state influence probability map with a sparse structure. This map satisfies the following technical attributes: the nodes are device numbers; the edges are the paths of abnormal state propagation; the edge weights are the state propagation probabilities (ranging from 0 to 1); the directed structure represents the state propagation direction; only retain the paths with actual propagation evidence in history.

[0043] Application example: For example, in the historical operation data, the state deviation score of device A has been abnormal five times in a row, and in four of them, the physical adjacent device B also has an abnormal deviation score within 15 minutes. Then the conditional probability is 4 / 5 = 0.8, which is much higher than the preset standard value of 0.5. The device A→B is regarded as the state propagation path, and the edge weight is 0.8. Finally, the state influence probability map between devices is formed after the evaluation is triggered and before being input into the risk recognition model. It is the core bridge connecting the abnormal recognition logic and the behavior path deduction. Its structure not only serves the generation of the state perturbation diffusion index, but also can be used to regulate the risk conduction prediction and resource allocation priority ranking in the control strategy.

[0044] Based on the construction of the state influence probability map between devices, to further achieve a quantitative judgment of the abnormal state propagation trend, it is necessary to perform the identification operation of potential abnormal influence paths on the devices currently in an abnormal state. This identification process aims to start from the devices whose current state deviation scores have triggered the evaluation conditions, and combine the historical experience structure in the state influence probability map to identify the downstream device links that may be affected by them, and quantify the propagation possibility of each path. This operation provides a path basis for the generation of the state perturbation diffusion index and is used for subsequent control decisions such as risk linkage response, local early warning zoning, and resource allocation. The identification process includes the following operation steps: Step 1: Initialize the starting point and screen the first-layer nodes. Use the device currently in an abnormal state as the starting node, and this device triggers the evaluation process by the state deviation score or the consistent state score of adjacent devices. Extract all the sets of directed edges emitted from this starting node from the state influence probability map. The directed edges represent the paths that may affect other devices starting from the current device state abnormality in history, and the edge weights are the propagation probabilities.

[0045] Filter target nodes from this set whose edge weights (i.e., propagation probabilities) are higher than the preset propagation threshold. These target nodes form the first-layer set of affected nodes of this device. The propagation threshold is a preset control parameter of the system, which is used to filter low-impact paths and ensure that the path recognition results have representative practical risks. For example, if the propagation threshold is set to 0.4, then only when the propagation probability of an edge ≥ 0.4, this path is included in the analysis path.

[0046] Step 2: Multilayer path recursive expansion. For the first-layer affected nodes, repeat the above operation: take it as the current node and continue to find target nodes with edge weights greater than the propagation threshold among its outgoing edges to form the second-layer affected nodes. Expand layer by layer outward in this way to form a path tree structure. To prevent infinite expansion or the path depth exceeding the analysis capability range, the following two types of termination conditions are introduced: Preset propagation step: Limit the maximum number of hops of the path (such as not exceeding 3 layers) to control the path structure within a reasonable complexity; Dynamic propagation intensity limit: If the propagation probabilities of all outgoing edges of a node are lower than a dynamically adjusted threshold, it is considered that the propagation trend weakens and the path expansion is automatically terminated. Through the layer-by-layer expansion mechanism, a set of all paths that meet the propagation conditions with the current device as the source can be constructed, representing its potential abnormal influence range in the probability graph.

[0047] Step 3: Path recording and sorting output. During the above expansion process, the system records the device sequence and the corresponding propagation probability in each path. To facilitate the evaluation of path effectiveness, calculate the cumulative probability of each path, that is, the product of the propagation probabilities of each edge in the path, which is used to represent the connectivity propagation possibility of the entire path. For example, if the propagation probabilities of the three edges in the path A→B→C are 0.8, 0.7, and 0.6 respectively, then the cumulative propagation probability of this path is 0.8×0.7×0.6 = 0.336.

[0048] Sort all paths in descending order according to the cumulative probability, and select the first several paths (such as the first 5 paths or the cumulative probability covering the first 80%) as the output result of the final abnormal influence paths. This result not only reflects the influence range of the current abnormal device but also clarifies the structural chain of its high-risk influence paths.

[0049] Example illustration: Assume that device A triggers the status evaluation. A→B (edge weight 0.8), A→C (edge weight 0.2), B→D (edge weight 0.6), C→D (edge weight 0.5). Let the propagation threshold be 0.3 and the maximum step size be 2. Then: In the first step, B is screened out; in the second step, it expands from B to D (because the edge weight 0.6 > 0.3); the cumulative probability of the path A→B→D is 0.8×0.6 = 0.48 and can be included in the final output; the path A→C is excluded (0.2 < 0.3). This recognition process realizes the recognition of multi-hop propagation links derived from a single-point anomaly based on the historical statistical probability structure, enabling the system to still complete the convergence path search in complex scenarios of multiple devices and having strong logical interpretability and risk orientation. At the same time, this path structure provides key factor support such as the number of paths, depth, and propagation weight for the subsequent calculation of the "status perturbation diffusion index", occupying an important position in the overall status recognition mechanism.

[0050] After completing the recognition of the abnormal influence path of the current device, it is necessary to further extract two core quantitative indicators for subsequent risk recognition modeling from the recognized path results, namely, the status perturbation diffusion index and the collaborative behavior deviation index. These two types of indices, as the fusion expressions of the device in the dimensions of multi-dimensional status analysis and group behavior analysis, are important basic variables for constructing the risk recognition system of the present invention. The calculation processes of the two are as follows: I. Extraction method of the state perturbation diffusion index: The state perturbation diffusion index is used to measure the possibility and intensity of the current device state anomaly spreading to other devices in the graph structure, and is a trend evaluation quantity reflecting the overall impact range of the current anomaly on the system. The construction of this index is based on the identified set of anomaly influence paths. Each path contains key attributes such as the device sequence, path propagation probability, and path hop count. At the same time, referring to the state deviation score of the current device, an index expression model is comprehensively constructed. The definition is as follows: The calculation formula is: D = S×(∑p_i×w_i)×exp(-λ×h), where D is the state perturbation diffusion index, which is the final output result, S is the current state deviation score, indicating the current state anomaly degree, which has been obtained based on the Mahalanobis distance through the previous method, p_i is the propagation probability of path i, originating from the edge weight in the state influence probability map, w_i is the normalized weight of the state transfer factor in the path, indicating the relative importance of the propagation ability in each path, which can be set proportionally according to the path propagation intensity and structural complexity, assigned by experts, or set by other alternative methods, h is the path hop count, that is, the total number of device nodes in the path, reflecting the length of the anomaly propagation chain, and λ is the path attenuation coefficient, which is an empirical value determined according to the network topology complexity or hierarchical structure. The larger the value, the more the hop count attenuation effect is emphasized. In this formula, the weighted sum of the propagation probability and the transfer weight represents the propagation ability in the entire path set. After multiplying by the state deviation score, it expresses the initial anomaly intensity multiplied by the propagation trend. The negative exponential decay function of the hop count is used to reflect that the longer the path length, the more its comprehensive influence decays, so as to control the exaggerated expression of the anomaly influence on the remote nodes.

[0051] II. Extraction method of collaborative behavior deviation index: The collaborative behavior deviation index is used to measure whether the behavior pattern of the current device in the state evolution sequence significantly deviates from that of the devices in its neighborhood group, emphasizing the importance of the consistency of individual behaviors in the group structure. This index models the state evolution sequences of multiple adjacent devices and constructs a time-series feature entropy weight distribution model to measure the behavior differences. The collaborative behavior deviation index is defined as: C = ∑(E_j × δ_j), where C is the collaborative behavior deviation index, E_j is the relative information entropy of the j-th feature dimension, reflecting the distribution discreteness of the behavior pattern in this dimension, and δ_j is the mean square deviation between the current device and the neighborhood devices in this dimension, representing the degree of outlier of the behavior trend in terms of magnitude. The feature dimension j includes three items: the current feature value, the change slope, and the trajectory deviation degree, all of which are the core components of the state evolution sequence defined above; the three dimensions are calculated for the sequence after being normalized by the historical mean. This method, through the composite modeling strategy of migrating "distribution discreteness and change amplitude", not only expresses the behavior fluctuations but also retains the correlation of the evolution characteristics between sequences. Both types of indexes only depend on the obtained state evolution sequence, state deviation score, and abnormal influence path parameters, and have high endogeneity, evolvability, and generalization ability. The information entropy is calculated using the Shannon entropy method. If the state changes of the adjacent devices are relatively concentrated in the current feature value dimension, the entropy value is low, indicating strong group consistency; if the behavior distribution is relatively wide, the entropy value is high, indicating the existence of unstable or noisy behaviors.

[0052] In the state monitoring method of the present invention, after identifying the state perturbation diffusion index and the collaborative behavior deviation index, in order to further determine the risk level of the current device state and its potential causes, a pre-trained fuzzy logic controller is used as the risk identification model. This model has the ability to drive multi-level fuzzy rule reasoning with uncertain inputs, and can combine the mapping relationship between the state evolution characteristics and risk labels in historical samples to output an interpretable risk interval and cause type, providing a key basis for the implementation of state regulation measures.

[0053] In this fuzzy logic controller model, the state perturbation diffusion index and the collaborative behavior deviation index are used as input variables, respectively representing the risk trends of the current device in the direction of abnormal influence propagation and the direction of group behavior deviation. Both of these indexes have been calculated through variables such as path propagation probability, hop count, information entropy, and mean square deviation in the manner disclosed above, and have evolvability and a normalized structure, facilitating their use as fuzzy logic inputs. For each input variable, a set of three-segment membership functions is constructed respectively, that is, its value range is divided into three intervals: "low", "medium", and "high". The three-segment membership function can adopt the following form: Low: corresponding to the interval less than a certain threshold (such as 0.3), using a Z-type function; Medium: with the peak at the center of the interval (such as 0.5), using a trapezoidal function or a Gaussian function; High: corresponding to the interval greater than a certain threshold (such as 0.7), using an S-type function.

[0054] Each interval is mapped to a fuzzy linguistic term, such as "weak diffusion", "medium diffusion", "strong diffusion" or "small deviation", "medium deviation", "large deviation". In this way, the input space is combined by three items of each of the two variables, forming 3×3 = 9 combinations of input state rules, that is, all possible fuzzy condition cross - cases.

[0055] An inference engine based on a rule table is built inside the fuzzy logic unit. Each rule has the following form: If the "state perturbation diffusion index" is high and the "synergistic behavior deviation index" is medium, then the "risk level" is severely abnormal and the "cause of abnormality" is category A. In the model training stage, according to the actual risk levels and event labels corresponding to the evolution characteristics of each state in the historical dataset, the above - mentioned rule set is constructed through statistical learning. Each rule is associated with a set of training frequency weights for subsequent inference. The inference process adopts the Mamdani - type fuzzy inference mechanism, including: calculating the membership degrees of the current input index values under three membership functions respectively; activating the corresponding rules according to the membership degree combination to obtain multiple fuzzy outputs; after aggregating the fuzzy output set, using the weighted average method to obtain the final output; The output includes two parts: Risk level interval: According to the five - segment division, such as "normal", "slightly abnormal", "moderately abnormal", "severely abnormal", "urgent", the level corresponding to the fuzzy linguistic term with the largest membership degree is the main output; Cause probability distribution vector: By statistically counting the frequencies of historical sample labels corresponding to each activated rule and combining with the membership degree weighted summation, a normalized vector is formed, representing the relative probabilities of various causes of abnormality under the current input combination.

[0056] The risk level interval is used to describe the urgency of the current state. If the output is "severely abnormal" or "urgent", then the high - priority impact control measures in the preset strategy are triggered, such as forcibly switching the power supply or dispatching a standby branch; the cause probability distribution vector is a probability vector with a length of n, where n is the number of preset abnormal scenario categories in the system (such as temperature anomaly, humidity mutation, load fluctuation, discharge increase, etc.). Each dimension value pk ∈ [0,1], representing the probability that the current state is attributed to this cause of abnormality.

[0057] Implementation example description: The input state perturbation diffusion index D = 0.65, and the collaborative behavior deviation index C = 0.42; D is 0.6 and 0.4 on the "medium" and "high" membership functions respectively, and C is 0.9 on the "medium" membership function; Activation rules such as: "medium - medium", "high - medium", corresponding outputs are "moderate anomaly" and "severe anomaly"; After the aggregated output, the fuzzy center falls within the "severe anomaly" interval; Cause vector [temperature fluctuation: 0.1, humidity anomaly: 0.3, voltage perturbation: 0.5, discharge enhancement: 0.1].

[0058] The output cause probability distribution vector can be used as a feedback signal to participate in the adaptive update mechanism of the causal path graph, that is, if the same cause label is repeatedly associated with a certain state causal edge in the path graph, the statistical weight of this path will be enhanced, thereby improving the self-learning and generalization ability of the state propagation model.

[0059] The implementation of impact control measures and the iterative update mechanism of the causal path graph are as follows: In the present invention, after the risk identification of the current state is completed, two types of key information will be output: the risk level interval and the cause probability distribution vector. The former describes the urgency of the current device state and is used to quantify which level of the five levels of normal, mild anomaly, moderate anomaly, severe anomaly, or emergency the current state belongs to; the latter expresses various possible abnormal causes and their probability distributions behind this state, providing a basis for abnormal origin tracing and decision-making.

[0060] Execute impact control measures according to the preset strategy: Impact control measures refer to the control behaviors automatically executed by the system according to the preset strategy when it is identified that a certain device state has a high risk level or is attributed to a certain type of high-risk abnormal cause. These strategies are predefined and a mapping relationship is established corresponding to the risk level interval and the cause probability distribution vector.

[0061] Impact control measures may include: Strategy A: Increase the data sampling frequency, which is used for the situation where the risk level is in the "mild anomaly" state to increase the monitoring granularity; Strategy B: Automatic alarm publishing, which is used for levels of "moderate anomaly" and above, pushing real-time alarms with risk levels and cause labels to the operation and maintenance platform; Strategy C: Temporary load switching, which is used for levels above "severe anomaly", automatically performing local distribution load adjustment to avoid further cascading risks; Strategy D: Physical isolation operation, which is used for the "emergency" level, performing safety isolation measures such as power-off protection to prevent the state from further spreading; Strategy E: Push operation and maintenance instructions, combined with the abnormal mechanism with the highest cause probability. For example, if the probability of "cable insulation deterioration" is 0.78, then a special inspection task for the insulation state will be issued. This strategy library and the output results of the fuzzy logic controller form a "rule - action" mapping set. After each identification is completed, the control execution process will be immediately entered to achieve adaptive intervention control.

[0062] Iteratively update the causal path graph based on the cause distribution: The causal path graph in the present invention is a structure formed on the basis of the device - to - device state propagation model constructed above. This graph contains directed edges for state propagation between device nodes, where the edge weights represent propagation probabilities and the paths represent possible state diffusion trajectories. To enhance the timeliness and adaptability of the model, it is necessary to iteratively update the graph based on the identified cause probability distribution vector. This update process includes the following mechanisms: Strengthening the attribution of cause labels: Correlate the currently identified main cause (i.e., the one with the highest probability) with all the edges associated with the hit propagation paths in the graph path. If an edge is repeatedly determined to be highly relevant to this cause, then the "cause label strength" of this edge is weighted and enhanced; Dynamically correcting the propagation probability: If a path is continuously identified as an abnormal propagation path and the probability corresponding to the same cause category is relatively high in each output, then the propagation probability of this path is increased to reflect the "data - driven correction" ability of the actual propagation trend; New path mining mechanism: If there is a jump - type propagation path in the currently identified path that has not appeared in the graph structure before, but the corresponding edge factors are reasonable (for example, similar abnormalities occur in adjacent devices after a short - time large - amplitude disturbance), then add this edge to the graph to form an expansion mechanism for "potential causal paths". This update process operates in a closed - loop manner of "cause feedback → graph reconstruction → weight redistribution" to ensure that the causal graph can continuously learn and evolve with state changes, improving the adaptive ability and long - term accuracy of the overall state recognition system.

[0063] Specific implementation example: For example, the current device number is A, the state disturbance diffusion index D = 0.71, and the collaborative behavior deviation index C = 0.58; the risk identification result shows that the risk level range is "seriously abnormal", and the proportion of the "local cable discharge deterioration" label in the cause probability distribution vector is 0.63; execute strategy C: immediately switch the branch box A to the standby branch temporarily to avoid current concentration load; at the same time, increase the weight of the edge factors marked as related to "local discharge deterioration" in the A→B→C three - hop propagation path; if device B also confirms the same cause in the next round of evaluation, automatically increase the propagation probability of this path by 10% to form a path reinforcement learning mechanism. Through the above - mentioned mechanisms, the present invention realizes a closed - loop intelligent identification and control system of risk identification - control decision - model update, which not only has the ability of immediate response, but also supports the self - evolution of the model driven by data, significantly improving the response accuracy and operation and maintenance efficiency for complex abnormal states in a dynamic power distribution environment.

[0064] All the above formulas are dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data and performing software simulation to get a formula closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0065] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0066] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0067] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0068] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A method for monitoring the state of a distribution cable branch box based on state recognition, characterized in that It includes the following steps: Collect the current, voltage, temperature, humidity and partial discharge data of multiple branch boxes within a continuous time period, and combine the historical data to obtain the state evolution sequence of each branch box; Based on each state evolution sequence, calculate the state deviation score and the adjacent device state consistency score. If any score exceeds the corresponding preset threshold, trigger the state evaluation process; if neither exceeds, maintain real-time update and monitoring; After triggering the evaluation, construct a state influence probability map between devices according to the historical state event sequence, identify potential abnormal influence paths, and extract two types of indices from them: the state perturbation diffusion index, which is used to measure the intensity of the current state anomaly spreading to other devices through the causal path; the collaborative behavior deviation index, which is used to measure the difference degree in the time series characteristics between the current device behavior and the behavior of its neighborhood group; Input the state perturbation diffusion index and the collaborative behavior deviation index into the pre-trained risk identification model to generate the risk level interval and the cause probability distribution vector of the current state, which are used to characterize the risk scenario type to which the current state belongs and the corresponding abnormal causes; Based on the generated risk level interval and the cause probability distribution vector, execute the corresponding influence control measures according to the preset strategy, and iteratively update the causal path map in the state recognition logic according to the cause distribution.

2. The state monitoring method for a distribution cable branch box based on state recognition according to claim 1, characterized in that The state evolution sequence includes the current eigenvalue, the change slope and the trajectory deviation degree, which are used to express the evolution trend of the device. The current eigenvalue is the original physical quantity data collected at the current time point, including the instantaneous values of current, voltage, temperature, humidity and partial discharge. The change slope is the change rate obtained by taking the difference between the current eigenvalues of two adjacent time points and dividing by the time interval, which is used to reflect the dynamic growth or decay trend of the physical quantity. The trajectory deviation degree is the calculation result of the distance between the current eigenvalue and the mean or median value of the device in the historical same type time period, which is used to evaluate whether the current state of the device deviates from its long-term behavior trajectory. The change slope and the trajectory deviation degree are both numerically normalized in a standardized manner to ensure comparability.

3. The state monitoring method for a distribution cable branch box based on state recognition according to claim 2, wherein The calculation of the trajectory deviation degree is implemented based on any one of the Euclidean distance, cosine similarity or Mahalanobis distance.

4. The state monitoring method for a distribution cable branch box based on state recognition according to claim 3, characterized in that, The state deviation score is obtained by combining and calculating the current eigenvalue, the change slope and the trajectory deviation degree in the state evolution sequence to obtain a specific value, which is used to evaluate the overall deviation degree of the current state relative to the historical behavior model. When calculating the three, they are first normalized respectively to form a standard feature vector, and then the standard feature vector is matched and calculated with the set of historical normal state sample feature vectors of the device in the corresponding season and corresponding time period. The calculation method is to use the Mahalanobis distance function to evaluate the covariance matrix of the current standard feature vector and the historical feature distribution, and output the statistical distance of the current state deviating from the center of the distribution as the state deviation score.

5. The method for monitoring the state of a distribution cable branch box based on state recognition according to claim 4, wherein, The adjacent device status consistency score is obtained by comparing and calculating the status evolution sequences of the current device with those of multiple other branch boxes adjacent to it in terms of structure or physical distribution, and is used to measure the behavioral consistency between the current device's status evolution trend and that of neighboring devices. The status evolution sequence of each device includes the current eigenvalue, change slope, and trajectory deviation. The score calculation includes two steps: The first step is to select a set of adjacent devices that are directly connected in terms of spatial location or power supply path or whose physical distance is less than the set spatial threshold; The second step is to calculate the dynamic time warping distance between the status evolution sequences of the current device and each adjacent device in the same time period, and take the average distance value of all adjacent devices as the preliminary consistency measurement result. Then, take the reciprocal of this result and perform normalization processing to obtain the adjacent device status consistency score.

6. The state monitoring method for a distribution cable branch box based on state recognition according to claim 5, characterized in that, After triggering the evaluation, construct a probability map of the state influence between devices based on the historical state event sequence. The construction process includes the following steps: First, collect the status evolution sequences experienced by multiple branch boxes in different past time periods and their corresponding risk level intervals, and organize them into a set of historical state event sequences; Perform frequency analysis on the order of occurrence of events in this set through statistical learning methods, and construct a directed relationship structure based on the temporal precedence relationship between the status evolutions of different devices to form an initial event association network; Adopt the conditional probability calculation method to evaluate the joint probability that the status deviation score of an adjacent device becomes abnormal after a certain number of moments on the premise that the status deviation score of the current device exceeds the threshold, and establish a cross-device state causal association relationship; According to the association strength, that is, the joint probability, screen the edge weights to form a sparse graph structure that only retains high-probability paths exceeding the preset probability standard value, which is the probability map of the state influence between devices.

7. The state monitoring method for a distribution cable branch box based on state recognition according to claim 6, characterized in that, Based on the construction of the probability map of the state influence between devices, the process of identifying potential abnormal influence paths includes the following operations: Taking the current device as the starting node, traverse all the directed edges in the probability map of the state influence that originate from this device, and screen the target nodes with edge weights higher than the preset propagation threshold to form the first-layer set of affected nodes; For each first-layer node, repeat the above screening operation, and expand the influence link layer by layer outward until the traversal depth reaches the preset propagation step length or the propagation intensity is lower than the dynamic termination condition, forming the set of all high-probability state influence paths starting from the current device; During the identification process, each path records the device sequence and the corresponding propagation probability in its propagation link, and sorts them by the cumulative probability of the path. Select the top several paths as the output result of the final abnormal influence path.

8. The state monitoring method for a distribution cable branch box based on state recognition according to claim 7, characterized in that After identifying the potential abnormal influence paths, the process of extracting the state perturbation diffusion index and the collaborative behavior deviation index from them includes the following steps: For the state perturbation diffusion index, an exponential expression model is constructed based on the propagation probability, the number of path hops, and the state perturbation intensity of each path in the abnormal influence path. The state perturbation diffusion index is defined as an exponential decay function of the perturbation intensity multiplied by the total path length, that is, the perturbation diffusion index is equal to the perturbation intensity multiplied by the path probability weighting coefficient and then multiplied by the negative exponential function of the number of hops. The calculation formula is: D = S × (∑p_i × w_i) × exp(-λ × h), where D is the state perturbation diffusion index, S is the current state deviation score, p_i is the propagation probability of path i, w_i is the normalized weight of the state transfer factor in the path, h is the number of path hops, and λ is the path attenuation coefficient, and its value is set according to the experience of the device hierarchical structure; For the collaborative behavior deviation index, the state evolution sequences of multiple neighboring devices are regarded as distribution sequences, and a deviation measurement model based on the entropy weight distribution of time series features is constructed. The dynamic weighted information entropy between the current device state evolution sequence and the first-layer neighboring device sequence is calculated in three dimensions: the current eigenvalue, the change slope, and the trajectory deviation degree. The collaborative behavior deviation index is defined as: C = ∑(E_j × δ_j), where C is the collaborative behavior deviation index, E_j is the relative information entropy of the jth feature dimension, and δ_j is the mean square error between the current device and the neighboring device in this dimension. The sequences are calculated after normalizing each dimension using the historical mean.

9. The state monitoring method for a distribution cable branch box based on state recognition according to claim 8, characterized in that, The pre-trained risk identification model is a fuzzy logic unit, which receives the state perturbation diffusion index and the collaborative behavior deviation index as input variables, and uses a fuzzy inference mechanism based on the membership function to construct the mapping relationship from input to output, for generating the risk level interval and the cause probability distribution vector of the current state; Among them, the state perturbation diffusion index and the collaborative behavior deviation index are respectively defined as input dimensions, and are divided into three grade intervals: low, medium, and high through a three-segment membership function. Each input variable corresponds to three fuzzy linguistic terms, and nine input state rule combinations are formed by combining the input spaces. The fuzzy logic unit uses an inference engine based on a fuzzy rule table to map each combined state to a five-level risk level output. Among them, the risk level interval is the final judgment result based on the interval where the fuzzy membership value is the largest. At the same time, according to the position of the fuzzy aggregation center point of the output layer, the cause probability distribution vector is constructed by combining the statistical frequency of the event cause labels of each rule in the training samples, indicating the attribution probability of each type of abnormal cause in this input state.

Images