Edge real-time inference system for operation risk of distribution box in internet of things environment

Abstract

The application provides an edge real-time inference system for operation risk of a distribution box in an Internet of Things environment, and relates to the technical field of data processing, comprising: constructing a reference field according to three state anchor point coordinate parameters, dividing the reference field to obtain a grid set, extracting the grid density gradient in the grid set to obtain a compensation parameter, calibrating a preliminary risk state trajectory spectrum by using the compensation parameter to obtain a target risk state trajectory spectrum; being used for performing feature extraction operation on the target risk state trajectory spectrum, constructing a steady-state boundary set, and calculating the spatial feature deviation and cumulative abnormal confidence value of the risk trajectory in the target risk state trajectory spectrum relative to the steady-state boundary set; when the weighted fusion index of the spatial feature deviation and the cumulative abnormal confidence value exceeds the steady-state boundary set, performing edge-side real-time state transition judgment logic to obtain a distribution box operation risk level label. The application improves the risk identification capability and response speed of the operation state of the distribution box.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an edge-end real-time inference system for the operational risk of distribution boxes in an Internet of Things (IoT) environment. Background Technology

[0002] In the risk monitoring of distribution boxes at the end of urban power distribution networks, edge computing terminals based on the Internet of Things (IoT) have been gradually deployed to collect various data such as current, voltage, box door status, and internal temperature and humidity. Existing technologies employ static thresholds or fixed logic rules to independently determine the limits of these individual parameters; other solutions add basic data filtering to reduce transient noise interference. However, in actual operating conditions, distribution boxes sometimes face complex situations such as frequent load switching, continuous changes in ambient temperature and humidity, and transient fluctuations in communication links. For example, in a certain distribution box during midday in summer, the load current may slowly climb from 80A to 120A, while the internal temperature... In a short period, the temperature rises from 28°C to 46°C. Door magnetic signals occasionally experience millisecond-level transient jumps due to equipment vibration, and communication links may also experience brief packet losses. In such scenarios, the baseline of electrical parameters often drifts slowly with changes in temperature and load. However, using static thresholds can sometimes misjudge normal current increases exceeding 100A as overcurrent anomalies, and it's difficult to reflect the coupling effect between temperature and current. Simple door magnetic signal jumps are sometimes directly interpreted as illegal door opening alarms, resulting in a number of invalid alarms. Since existing methods mostly focus on discrete comparisons of instantaneous data, it's often difficult to construct dynamic reasoning mechanisms at the edge that reflect the temporal correlation and state transition process of multi-source data. Therefore, accurately tracking the aforementioned gradual, multi-factor coupled risk evolution trajectory, such as the process from heat accumulation to a slow decline in insulation performance, is usually difficult, easily leading to delayed risk level determination or false alarms. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide an edge-end real-time inference system for the operational risks of distribution boxes in an Internet of Things (IoT) environment, thereby improving the risk identification capability and response speed of the operational status of distribution boxes.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] The first aspect is the edge-end real-time inference system for the operational risks of distribution boxes in an IoT environment, including:

[0006] The acquisition module is used to acquire multi-source data streams, perform time-series alignment and dimensional normalization on the multi-source data streams to construct an initial state feature set, calculate the correlation weights between each dimension within the initial state feature set, construct a potential field model based on the correlation weights, and perform topological reorganization on the feature nodes within the potential field model to obtain the feature coupled state.

[0007] The extraction module is used to input the feature coupled state into the edge-side temporal inference network, extract the temporal change rate of each dimension in the feature coupled state, and construct the evolution model;

[0008] The computation module is used to perform state recursion, anomaly screening, gradient calculation and smoothing, deflection curvature extraction and spatial mapping reconstruction on the evolution model to construct a preliminary risk state trajectory spectrum; obtain the coordinate parameters of three state anchor points pre-deployed on the distribution box, construct a reference field based on the coordinate parameters of the three state anchor points, divide the reference field to obtain a grid set, extract the grid density gradient within the grid set to obtain compensation parameters, and use the compensation parameters to calibrate the preliminary risk state trajectory spectrum to obtain the target risk state trajectory spectrum;

[0009] The output module is used to perform feature extraction on the target risk state trajectory spectrum, construct a steady-state boundary set, and calculate the spatial feature deviation and cumulative anomaly confidence value of the risk trajectory in the target risk state trajectory spectrum relative to the steady-state boundary set. When the weighted fusion index of spatial feature deviation and cumulative anomaly confidence value exceeds the steady-state boundary set, the edge-side real-time state transition judgment logic is executed to obtain the distribution box operation risk level label.

[0010] In a second aspect, a computing device includes:

[0011] One or more processors;

[0012] A storage device for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the system.

[0013] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the system.

[0014] The above-described solution of the present invention has at least the following beneficial effects:

[0015] By calculating the correlation weights between various dimensions, constructing a potential field model, and completing the topological reorganization of feature nodes, deep coupling and feature fusion of multi-source data can be achieved, helping to uncover the intrinsic correlations between data such as electrical parameters and equipment status. Compared with the method of independently judging a single parameter and ignoring the influence of parameter coupling, this method is beneficial to avoid parameter interference under complex operating conditions such as load switching and environmental changes, and reduce misjudgments caused by parameter baseline drift. The feature coupled state is input into the edge-side time-series extrapolation network to extract the time-series change rate of each dimension and construct an evolution model. The entire data processing, feature extraction, and risk extrapolation process is completed at the edge, without the need to upload massive amounts of raw data to the cloud for processing, which can effectively reduce data transmission latency and network bandwidth consumption. At the same time, by calculating the propagation gradient and deflection curvature of abnormal features and combining it with time-series inertial smoothing, the risk evolution trajectory can be tracked in real time, which can better capture gradual, multi-factor coupled risk changes, such as insulation performance degradation caused by heat accumulation, and shorten the response time for risk identification and level determination. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an edge-end real-time inference system for assessing the operational risks of distribution boxes in an Internet of Things (IoT) environment, provided by an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the process provided by an embodiment of the present invention, which involves inputting the feature coupled state into the edge-side temporal inference network, extracting the temporal change rate of each dimension in the feature coupled state, and constructing an evolution model. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] like Figure 1 As shown, embodiments of the present invention propose an edge-end real-time inference system for assessing the operational risks of distribution boxes in an Internet of Things (IoT) environment, comprising:

[0020] The acquisition module is used to acquire multi-source data streams, perform time-series alignment and dimensional normalization on the multi-source data streams to construct an initial state feature set, calculate the correlation weights between each dimension within the initial state feature set, construct a potential field model based on the correlation weights, and perform topological reorganization on the feature nodes within the potential field model to obtain the feature coupled state.

[0021] The extraction module is used to input the feature coupled state into the edge-side temporal inference network, extract the temporal change rate of each dimension in the feature coupled state, and construct the evolution model;

[0022] The computation module is used to perform state recursion, anomaly screening, gradient calculation and smoothing, deflection curvature extraction and spatial mapping reconstruction on the evolution model to construct a preliminary risk state trajectory spectrum; obtain the coordinate parameters of three state anchor points pre-deployed on the distribution box, construct a reference field based on the coordinate parameters of the three state anchor points, divide the reference field to obtain a grid set, extract the grid density gradient within the grid set to obtain compensation parameters, and use the compensation parameters to calibrate the preliminary risk state trajectory spectrum to obtain the target risk state trajectory spectrum;

[0023] The output module is used to perform feature extraction on the target risk state trajectory spectrum, construct a steady-state boundary set, and calculate the spatial feature deviation and cumulative anomaly confidence value of the risk trajectory in the target risk state trajectory spectrum relative to the steady-state boundary set. When the weighted fusion index of spatial feature deviation and cumulative anomaly confidence value exceeds the steady-state boundary set, the edge-side real-time state transition judgment logic is executed to obtain the distribution box operation risk level label.

[0024] In this embodiment of the invention, by calculating the correlation weights between various dimensions, constructing a potential field model, and completing the topological reorganization of feature nodes, deep coupling and feature fusion of multi-source data can be achieved, helping to uncover the intrinsic correlation between data such as electrical parameters and equipment status. Compared with the method of independent judgment of a single parameter and ignoring the influence of parameter coupling, this method is beneficial to avoid parameter interference under complex working conditions such as load switching and environmental changes, and reduce misjudgments caused by parameter baseline drift. The feature coupling state is input into the edge-side time-series inference network, the time-series change rate of each dimension is extracted, and an evolution model is constructed. The entire data processing, feature extraction, and risk inference process are completed at the edge, without the need to upload massive amounts of raw data to the cloud for processing, which can effectively reduce data transmission latency and network bandwidth consumption. At the same time, by calculating the propagation gradient and deflection curvature of abnormal features and combining it with time-series inertial smoothing, the risk evolution trajectory can be tracked in real time, which can better capture gradual, multi-factor coupled risk changes, such as insulation performance degradation caused by heat accumulation, and shorten the response time for risk identification and level determination.

[0025] In a preferred embodiment of the present invention, multi-source data streams are acquired, and time-series alignment and dimensional normalization are performed on the multi-source data streams to construct an initial state feature set. Correlation weights between dimensions within the initial state feature set are calculated. A potential field model is constructed based on the correlation weights, and the feature nodes within the potential field model are topologically reorganized to obtain a feature coupled state, which may include:

[0026] The system collects three-phase voltage, three-phase current, active power, grid frequency, and door status signals during the operation of the distribution box to obtain a multi-source data stream. It then parses the timestamps of the multi-source data stream and performs interpolation to fill in any missing time segments, resulting in a continuous time-series data stream. Specifically, this involves establishing stable data connections with voltage sensors, current sensors, power sensors, frequency sensors, and the magnetic door switch installed inside the distribution box via an edge computing terminal deployed within the box. The voltage sensors collect real-time instantaneous sampling data of the voltage of each phase in the three-phase circuit of the distribution box, sampling every 10 milliseconds to ensure the capture of instantaneous voltage fluctuations. The current sensors collect real-time instantaneous sampling data of the current of each phase in the three-phase circuit. The sampling interval is consistent with the voltage sampling interval to achieve synchronous acquisition of voltage and current data; the power sensor is used to collect real-time instantaneous active power data of the distribution box, and synchronously collects active power data of each phase and total active power data; the frequency sensor is used to collect real-time frequency data of the power grid, with an acquisition interval of 50 milliseconds to ensure timely capture of minute changes in the power grid frequency; the magnetic control switch of the box door is used to collect the door status signal of the metering box, using high and low level signals to distinguish the door status, outputting a low level signal when the door is closed and a high level signal when the door is open, with a signal acquisition interval of 100 milliseconds, and simultaneously capturing instantaneous changes in the door status; the edge computing terminal organizes and integrates all the above-mentioned collected data according to the acquisition time, parameter type, and parameter value format to form a complete multi-source data stream.

[0027] The edge computing terminal parses the millisecond-level timestamps carried by each data item in the multi-source data stream, sorts all the data according to the chronological order of the timestamps to form an ordered data sequence, and then compares the timestamps of all the sorted data one by one to identify time gaps in the data sequence, that is, areas where corresponding collected data is missing within a certain time period. The criterion for determining a missing segment is that the time interval between two adjacent data timestamps is greater than the preset collection interval of the corresponding parameter. For example, if the time interval between two adjacent timestamps of voltage data is greater than 10 milliseconds, then that time period is determined to be a missing segment of voltage data. For each time gap segment, the two adjacent normal data points before and after the gap segment are extracted, and the values ​​of these two data points and their corresponding timestamps are recorded to determine the change trend of the data before and after the gap segment. Linear interpolation is then used. The filling method is based on values. The specific calculation process is as follows: First, calculate the value of the data point after the missing segment and subtract the value of the data point before the missing segment to obtain the value difference. Then, calculate the timestamp of the missing position and subtract the timestamp of the data point before the missing segment to obtain the time difference. Divide this time difference by the total time difference between the data point after the missing segment and the data point before the missing segment to obtain the time interval percentage. Then, add the value of the data point before the missing segment to the value difference and multiply by the time interval percentage to calculate the filling value corresponding to the missing position. If there are multiple missing positions within the missing segment, calculate the filling value for each missing position in turn according to the above calculation method. Fill all missing data points in turn, supplementing the originally discontinuous data sequence into a continuous and uninterrupted data sequence, and finally obtain a continuous time-series data stream.

[0028] Using the master clock channel of the continuous time-series data stream as the synchronization reference, a fixed-step resampling process is performed on the continuous time-series data stream to obtain a time-aligned sequence. The maximum and minimum values ​​of each electrical parameter channel in the time-aligned sequence are extracted to obtain a range set. Based on the range set, a dimensional scaling process is performed on the time-aligned sequence to obtain a dimensional normalized matrix. Specifically, all electrical parameter channels include three-phase voltage channels, three-phase current channels, active power channels, and grid frequency channels, as well as a door status signal channel for monitoring the door status. The grid frequency channel is selected as the master clock channel because the grid frequency has the highest stability during the operation of the distribution box, is least affected by load changes and environmental interference, and is less prone to large fluctuations, thus providing a stable time synchronization reference. The sampling step size of the master clock channel is first determined, i.e., the sampling interval of the grid frequency is 50 milliseconds, and this sampling step size is used as the unified resampling step size. For all data channels in the continuous time-series data stream, including three-phase voltage, three-phase current, active power, and door status signals, resampling is performed at a fixed step size. The specific resampling process is as follows: a resampling time point is set at a time interval of 50 milliseconds. For each resampling time point, if the corresponding data channel has original sampled data, the original data is directly extracted as the value of that resampling time point. If the corresponding data channel does not have original sampled data, linear interpolation is used to supplement the corresponding data. The interpolation calculation method is the same as the linear interpolation method described above, that is, the supplementary value of that time point is calculated based on the two adjacent original data points before and after the resampling time point. Through the above resampling operation, all data channels have corresponding sampled values ​​at the same time point, completely eliminating the time deviation between different channels, ensuring that the data of each dimension is completely synchronized in time, and finally obtaining a time-aligned sequence that is completely aligned in time.

[0029] First, a fixed statistical period is determined, set to match the regular operation monitoring cycle of the distribution box, i.e., 24 hours, to ensure coverage of all operating conditions of the distribution box throughout the day. Within this statistical period, all sampled values ​​of each electrical parameter channel in the time-series alignment sequence are extracted sequentially. For each electrical parameter channel, all sampled values ​​within that channel are compared one by one to find the maximum and minimum values. The range of each channel is calculated by subtracting the minimum value from the maximum value. The range reflects the numerical fluctuation range of the data in that channel. The ranges of all electrical parameter channels are integrated to form a range set, which contains information on the numerical fluctuation range of each channel. For each data point in the time-series alignment sequence, a dimensional scaling transformation is performed. Specifically, the value of the data point is subtracted from the minimum value of the corresponding channel to obtain the numerical difference, which is then divided by the range of the corresponding channel. Through this calculation, different dimensions and different numerical ranges are transformed. The electrical parameters are uniformly mapped to the same numerical range of 0 to 1. For example, the voltage parameter range is 190V to 240V with a range of 50V. A certain voltage data point is 220V. After dimensional scaling, the normalized value of this data point is (220-190)÷50=0.6. Through dimensional scaling, the calculation deviation caused by dimensional differences is completely eliminated, and the large differences in the numerical ranges of parameters such as voltage, current, and power are avoided. Finally, a dimensional normalization matrix is ​​obtained. The rows and columns of the dimensional normalization matrix are clearly defined. Specifically, each row of the matrix corresponds to a resampling time point, that is, each row of data corresponds to the normalized value of all electrical parameter channels after dimensional scaling at a fixed time point in the time-aligned sequence. Each column of the matrix corresponds to an electrical monitoring dimension (i.e., an electrical parameter channel), specifically arranged in the order of three-phase voltage, three-phase current, active power, grid frequency, and door status signal. The data in each column corresponds to the normalized value of the corresponding electrical parameter channel at all resampling time points.

[0030] The dimensionally normalized matrix is ​​column-wise concatenated according to the preset electrical monitoring dimensions to construct the initial state feature set. Iterating through any two dimension data columns within the initial state feature set, the corresponding data points of these two dimension data columns within a preset sliding window are extracted and covariance matching is performed to obtain the dimension covariance matrix. The diagonal elements of the dimension covariance matrix are extracted and square root transformation is performed to obtain the dimension standard deviation vector. Specifically, this includes: following a preset fixed order of electrical monitoring dimensions, the preset order being three-phase voltage, three-phase current, active power, grid frequency, and box door status signal. This order is set according to the importance of electrical parameters to the operational risks of the distribution box, ensuring that subsequent correlation calculations prioritize key parameters; the dimensionally normalized matrix... The data from each dimension are concatenated in a column-wise order according to the preset sequence. That is, all the data from each dimension are treated as a column and arranged and concatenated in the order of three-phase voltage, three-phase current, active power, grid frequency, and box door status signal to form a complete initial state feature set. The initial state feature set contains the normalized data of all monitoring dimensions and the preliminary correlation between each dimension, providing the basic data for the subsequent calculation of correlation weights. Traverse any two different dimension data columns in the initial state feature set and set a preset sliding window. The length of the sliding window is set to cover the short-term change trend of the data. Combined with the fluctuation characteristics of the distribution box operation data, the length of the sliding window is set to 100 sampling points, that is, 5 seconds of data, to ensure that the short-term correlation characteristics of the data can be captured.

[0031] Within each preset sliding window, all data points corresponding to the two dimensions are extracted. For each set of corresponding data points, the difference between the data point in the first dimension and the mean of all data points within the sliding window for that dimension is calculated. Then, the difference between the data point in the second dimension and the mean of all data points within the sliding window for that dimension is calculated. These two differences are multiplied to obtain the correlation product of the data points in that set. All correlation products of corresponding data points in all sets within the sliding window are summed to obtain a total sum. This total sum is then divided by the number of data points within the sliding window to obtain the covariance of the two dimensions within that sliding window. The covariance reflects the degree of linear correlation between the two dimensions; a positive covariance indicates a positive correlation between the two dimensions. The data in each dimension are positively correlated. A negative covariance indicates a negative correlation between the two dimensions, while a covariance of 0 indicates no linear correlation. Following the above method, the covariance between each pair of dimensions is calculated sequentially. All covariances are then integrated to form a dimensional covariance matrix. The number of rows and columns in the dimensional covariance matrix is ​​equal to the number of monitored dimensions, and the elements in the matrix correspond to the covariance between two dimensions. All elements on the diagonal of the dimensional covariance matrix are extracted. Each diagonal element corresponds to the covariance of a dimension itself. The square root operation is performed on each diagonal element to obtain the standard deviation of each dimension. The standard deviation is used to reflect the dispersion of the data in that dimension. The standard deviations of all dimensions are then integrated to form a dimensional standard deviation vector.

[0032] The off-diagonal elements of the dimensional covariance matrix are mapped proportionally to the corresponding components of the dimensional standard deviation vector to obtain the dimensional correlation coefficient matrix. Exponential decay weighting and non-negative truncation are then applied to the dimensional correlation coefficient matrix to obtain the dimensional interaction weight matrix. Specifically, each off-diagonal element in the dimensional covariance matrix is ​​divided by the product of the standard deviations of the corresponding two dimensions in the dimensional standard deviation vector to calculate the correlation coefficient between the two dimensions. The correlation coefficient ranges from -1 to 1; the closer the absolute value of the correlation coefficient is to 1, the stronger the correlation between the two dimensions; the closer the correlation coefficient is to 0, the weaker the correlation. All pairwise correlation coefficients are integrated to form the dimensional correlation coefficient matrix. Exponential decay weighting is then applied to each value in the dimensional correlation coefficient matrix, specifically using the exponential decay expression: ,in This represents the data value after exponential decay weighting. This represents the original values ​​in the dimensional correlation coefficient matrix, where e represents the natural constant (approximately 2.71828). This represents the attenuation coefficient (in conjunction with the length of the sliding window). (Set to 0.053) It represents the time step (the unit is the number of sliding windows, that is, the number of intervals between the current sliding window and the sliding window containing the target data; the longer the time, the larger the value of t).

[0033] Weighting is achieved through this exponential expression, and the exponential decay coefficient (i.e.) Based on the temporal correlation of the data, the further back in time the related data is, the better. The larger the value, the smaller the exponential decay coefficient, and the higher the weighted data value. The smaller the value, the more focused the attention is on recent data correlations; combined with the length of the sliding window (100 sampling points, i.e., 5 seconds), the attenuation coefficient... The value is set to 0.053 to ensure that the association weight decays by about 5% for each sliding window (i.e. every 5 seconds). Then, the values ​​that are less than zero after the calculation are truncated to zero to avoid negative associations interfering with subsequent calculations, thus obtaining the dimension interaction weight matrix.

[0034] The process involves summing the data in each row of the dimensional interaction weight matrix to obtain the total row sum. Then, the data in each row of the dimensional interaction weight matrix is ​​normalized to the total row sum. Finally, the correlation weights between the dimensions in the initial state feature set are calculated. Specifically, this includes summing all the values ​​in each row of the dimensional interaction weight matrix to obtain the total row sum for that row, and then dividing each value in that row by the total row sum for the corresponding row to complete the normalization process. This ensures that the sum of the weights in each row is 1. Finally, the correlation weights between the dimensions in the initial state feature set are calculated. These correlation weights can accurately reflect the degree of mutual influence between the electrical parameters.

[0035] The correlation weights are matched with the initial state feature set through dimensional mapping, and the potential field intensity parameters corresponding to each electrical monitoring dimension are extracted to obtain the potential field initialization parameter set. Based on the potential field initialization parameter set, spatial potential energy projection processing is performed on the initial state feature set to convert the data of each dimension in the initial state feature set into coordinate nodes in the potential field space. A potential field space basis is then established based on the coordinate nodes to construct the potential field model and the initial distribution set of feature nodes. Specifically, this includes: matching the correlation weights between each dimension in the initial state feature set with each electrical monitoring dimension in the initial state feature set through a one-to-one dimensional mapping. The matching rules strictly follow the preset dimensional order of the initial state feature set, i.e., three-phase voltage... The order of three-phase current, active power, grid frequency, and door status signals is used to ensure that each electrical monitoring dimension corresponds to a unique associated weight, preventing mismatches between dimensions and associated weights. After mapping and matching, the associated weight corresponding to each electrical monitoring dimension is extracted and used as the potential field strength parameter for that dimension. The potential field strength parameters of all electrical monitoring dimensions are integrated to form a complete potential field initialization parameter set. The potential field initialization parameter set clarifies the influence strength of each electrical monitoring dimension in the subsequent potential field space construction. The higher the associated weight of a dimension, the larger its corresponding potential field strength parameter, and the higher its influence weight in the potential field space. Based on the initial potential field parameter set, spatial potential energy projection processing is performed on the initial state feature set. Specifically, for each set of data in the initial state feature set, the normalized values ​​of the data in each electrical monitoring dimension are extracted one by one. The normalized value of each dimension is multiplied by the potential field intensity parameter corresponding to that dimension to calculate the projection coordinate value of the data in the potential field space. Each dimension corresponds to a coordinate dimension in the potential field space. That is, three-phase voltage, three-phase current, active power, grid frequency, and door status signal correspond to five coordinate dimensions in the potential field space, respectively. After projection processing, each set of data will form a five-dimensional coordinate point, which is a feature node in the potential field space. After projection processing, a potential field space basis is established based on all generated feature nodes. Specifically, the node with the largest and smallest coordinate values ​​among all feature nodes is selected to determine the value range of each coordinate dimension of the potential field space. The value range of all coordinate dimensions is consistent with the range of dimensionally normalized data, i.e., 0 to 1. Based on this value range, a five-dimensional basic framework of the potential field space is built, clarifying the electrical monitoring dimension corresponding to each coordinate dimension, and completing the construction of the potential field model. This potential field model can intuitively present the correlation and data distribution pattern of each electrical monitoring dimension. At the same time, the specific coordinate positions of all feature nodes in the potential field space are recorded, and the feature nodes are sorted according to the time order of the initial state feature set to form the initial distribution set of feature nodes.

[0036] Based on the initial distribution set of feature nodes and their associated weights, potential energy gradient inference is performed between nodes to obtain the potential energy gradient matrix. Threshold filtering and adjacency determination are then performed on the potential energy gradient matrix, retaining node connection paths with potential energy gradients higher than a preset threshold, resulting in a potential field topology graph. Specifically, this involves: combining the initial distribution set of feature nodes and the inter-dimensional association weights, performing potential energy gradient inference calculations between adjacent feature nodes in the potential field space one by one. The criterion for determining adjacent nodes is that the coordinate distance between two nodes in the potential field space is less than a preset distance threshold, set to 0.1, to ensure accurate identification of closely related adjacent nodes. The specific calculation process for the potential energy gradient is as follows: first, the coordinate difference between two adjacent nodes in the potential field space is calculated, i.e., the difference between the corresponding coordinates of the two nodes... The difference in coordinates is calculated, and then multiplied by the correlation weight of the corresponding dimension of the two nodes to obtain the potential gradient between the two nodes. The potential gradient between all adjacent nodes is calculated in turn, and integrated to form a complete potential gradient matrix. The rows and columns of the potential gradient matrix are clearly defined. Specifically, each row of the matrix corresponds to a feature node, and each column also corresponds to a feature node. Each element in the matrix corresponds to the potential gradient value between the feature node represented by the row and the feature node represented by the column. If two feature nodes are not adjacent (the coordinate distance is greater than or equal to the preset distance threshold of 0.1), the value of the element at that position is set to 0, which is used to distinguish the correlation between adjacent nodes and non-adjacent nodes. The number of rows and columns of the matrix are equal to the total number of feature nodes in the initial distribution set of feature nodes.

[0037] Based on the actual risk assessment requirements for the operation of distribution boxes, specifically referring to the need for accurate identification and prediction of safety risks caused by abnormal coupling of electrical parameters during the operation of distribution boxes, the core is to capture strong correlations and abnormal changes among electrical parameters such as three-phase voltage, three-phase current, active power, and grid frequency within the distribution box, as well as the coupling correlation between abnormal box door status and abnormal electrical parameters. This avoids safety hazards such as short circuits, overloads, and leakage caused by the failure to identify abnormal parameter correlations in a timely manner, ensuring the stable and safe operation of the distribution box. Based on this actual risk assessment requirement, a potential... The gradient critical value can distinguish between strong and weak correlations between nodes. Combined with the operating characteristics of the distribution box, these characteristics include stable and controllable coupling relationships among electrical parameters during operation. Core parameters such as three-phase voltage, three-phase current, active power, and grid frequency exhibit fixed inherent correlation patterns. Under normal operating conditions, the correlation strength of each parameter is at a moderate level, while under abnormal operating conditions, the strong correlation will significantly increase. The distribution box is susceptible to load fluctuations and environmental interference, causing slight fluctuations in parameter correlations. However, the boundary between strong and weak correlations is clear, and no significant fluctuations occur. In cases of fuzzy superposition, safety risks are mostly caused by anomalies in strong correlations of multiple parameters. Fluctuations in weak correlations usually do not lead to safety hazards and do not require special attention. After dimensional normalization, the values ​​of correlation weights and potential gradients are all in the range of 0 to 1, with a concentrated and regular distribution, making it easy to set a fixed critical value for screening. The specific reason for setting the critical value to 0.5 is that, combined with the calculation logic of the correlation weights, after proportional normalization, the correlation weights are all in the range of 0 to 1, and the corresponding potential gradient values ​​are also synchronously in the range of 0 to 1. 0.5 is in the middle of this range, which can clarify... Clearly define the boundaries between strong and weak correlations to avoid the problem of missing strong correlations due to excessively high threshold values ​​and misjudging weak correlations due to excessively low threshold values. Based on actual operation test data of the distribution box, the potential energy gradient values ​​of adjacent feature nodes are mostly concentrated below 0.5 under normal operating conditions. Under abnormal operating conditions (such as voltage and current imbalance, power surges, etc.), the potential energy gradient values ​​of closely correlated nodes are all above 0.5. This threshold value can accurately capture strong correlations under abnormal operating conditions, while eliminating weak correlation interference under normal operating conditions. By matching the potential field intensity parameters and spatial projection logic, the potential energy gradient value reaches 0.When the value is 5 or higher, the corresponding node correlation strength can reflect the substantial mutual influence between parameters, which meets the core requirement of focusing on key strong correlations in the risk assessment of distribution boxes. This provides a precise screening basis for subsequent topology connection graph construction and node reorganization. All potential gradient values ​​in the potential gradient matrix are filtered, retaining only node connections with values ​​higher than a preset threshold and eliminating weakly correlated connections with values ​​lower than the threshold. Simultaneously, the adjacency status of the nodes corresponding to the retained connections is determined to confirm whether there is a direct correlation between adjacent nodes, i.e., whether there is a direct interaction between the dimensions corresponding to the two nodes. All node connection paths that meet the conditions are integrated to form a potential field topology connection graph.

[0038] Based on the potential field topology connection graph, the initial distribution set of feature nodes is subjected to spatial position offset and high-dimensional aggregation processing to obtain the topologically reconstructed feature set. The topologically reconstructed feature set and the potential field topology connection graph are then subjected to structured mapping and fusion processing to obtain the feature coupling state. Specifically, according to the connection relationship of the potential field topology connection graph, all nodes in the initial distribution set of feature nodes are adjusted by spatial position offset. The specific adjustment rule is as follows: nodes with high association weights are offset in a direction closer to each other, and the offset distance is proportional to the association weight. The higher the association weight, the greater the offset distance, and the maximum offset distance is set to 0.05. Nodes with low association weights are offset in a direction farther away, and the offset distance is inversely proportional to the association weight. The lower the association weight, the greater the offset distance, ensuring that closely associated nodes can be clustered together and loosely associated nodes can be separated from each other. Simultaneously, closely related nodes in the potential field topology connection graph—that is, nodes connected by strong correlation paths—are aggregated towards the center of the high-dimensional space. During the aggregation process, the node with the highest correlation weight is used as the aggregation center, and other closely related nodes move towards this center, integrating the scattered nodes into multiple closely related node clusters. Each node cluster corresponds to a set of closely related electrical parameters, resulting in a topology-reorganized feature set. The topology-reorganized feature set and the potential field topology connection graph are then fully mapped and fused according to the structural relationship of the potential field space. That is, each node cluster in the topology-reorganized feature set is mapped one-to-one with the corresponding node connection path in the potential field topology connection graph. The coordinate data of the node clusters and the corresponding topology connection relationships are integrated into a unified whole, so that the fused feature data contains both the numerical information of the nodes and the correlation information between the nodes. Finally, a feature coupled state that can comprehensively reflect the temporal correlation and spatial coupling characteristics of multi-dimensional data is obtained.

[0039] In this embodiment, a potential field topology connection graph is constructed, retaining strong correlations and eliminating weak correlations to improve the accuracy of risk assessment; spatial offset and aggregation of feature nodes are performed to strengthen key features and improve the effectiveness of feature data.

[0040] like Figure 2As shown, in another preferred embodiment of the present invention, inputting the feature coupled state into the edge-side temporal inference network, extracting the temporal change rate of each dimension in the feature coupled state, and constructing an evolutionary model may include:

[0041] The feature-coupled state is input into the edge-side time-series inference network for channel feature parsing processing to obtain the network's initial response sequence. Based on the initial response sequence, numerical difference features of continuous time steps are extracted to obtain a dimensional difference feature set. Specifically, this includes: inputting the feature-coupled state into the edge-side time-series inference network, extracting the temporal change rate of each dimension in the feature-coupled state, and constructing an evolutionary model. The edge-side time-series inference network is a lightweight data processing network deployed on the edge computing terminal. Its core is to achieve efficient processing of the feature-coupled state based on preset feature parsing rules, adapting to the real-time computing needs of the edge end, and fitting the actual working conditions of limited computing power and the need for rapid data processing on the edge of the distribution box. The edge computing terminal has a computing power ≤2 TOPS and a latency ≤50ms. The feature-coupled state is a multi-dimensional feature set formed by integrating normalized data from various electrical monitoring dimensions and the correlations between dimensions. After the complete feature coupled state is input into the network, the edge-side time-series extrapolation network performs channel feature parsing processing on the input feature coupled state. It extracts the time-series data of each electrical monitoring dimension from the feature coupled state, separating the numerical information and correlation information of each dimension. It removes irrelevant redundant information such as random noise generated during data acquisition and interference data unrelated to changes in electrical parameters, retaining the core features that reflect parameter changes. Then, the parsed core features of each dimension are arranged in chronological order to obtain the network's initial response sequence. Based on this initial response sequence, numerical difference features of consecutive time steps are extracted to form a dimensional difference feature set. The interval between two consecutive adjacent time steps is consistent with the resampling step size, both being 50ms to ensure temporal consistency. Let the two consecutive time steps be t and t+1, and the core feature values ​​corresponding to the same electrical monitoring dimension be... and ,in This represents the core feature value of the electrical monitoring dimension at time step t and time step d, where t represents the time step number (ranging from 1 to T) and d represents the electrical monitoring dimension number (ranging from 1 to D). D is the total number of electrical monitoring dimensions, which is set to 8 based on actual monitoring needs: three-phase voltage, three-phase current, active power, and grid frequency. The formula for calculating the difference between the values ​​of a single dimension and two adjacent time steps is as follows: = ,in This represents the numerical difference between the d-th electrical monitoring dimension and the t-th and t+1-th time steps. A positive value indicates that the value of that dimension is increasing, and a negative value indicates that the value is decreasing. The above calculation is performed sequentially for all continuous time steps (t from 1 to T-1) and all electrical monitoring dimensions (d from 1 to D). All numerical differences are arranged in the order of dimension and time step to form the dimensional difference feature set.

[0042] A rate mapping process within a sliding window is performed on the dimensional difference feature set to obtain the temporal change rate of each dimension. Based on the temporal change rate of each dimension, a state transition relationship between adjacent time steps is constructed to obtain the dimensional state transition matrix. Specifically, this includes: performing a rate mapping process within a sliding window on the dimensional difference feature set to obtain the temporal change rate of each dimension. The sliding window length is set to 100 sampling points, corresponding to a time length of 5 seconds, i.e., 100 sampling points × 50ms / sampling points = 5000ms = 5s. This length can cover the short-term fluctuation period of electrical parameters while also considering the computational efficiency at the edge, avoiding excessive computational latency due to an excessively long window and misjudgment of data fluctuations due to an excessively short window. The dimensional difference feature set is divided into multiple consecutive sliding windows in chronological order, each window containing 100 numerical difference data points. There is no overlap between windows to ensure that each data point participates in the calculation only once. Let the k-th sliding window be k = 1, 2, ..., K, where K is the total number of sliding windows, and the numerical differences of all the electrical monitoring dimensions within the d-th dimension are... Where t ranges from (k-1)×100+1 to k×100, the formula for calculating the average of the numerical differences is: ,in Let represent the average difference of the numerical values ​​for the k-th sliding window and the d-th electrical monitoring dimension. Dividing the average difference by the sliding window length of 5 seconds yields the average rate of change of that dimension within the window, which is also the time-series rate of change. The calculation formula is: ,in This represents the time-series change rate of the k-th sliding window and the d-th electrical monitoring dimension, in normalized values ​​per second. It reflects the rate and direction of change of the dimension value over time. The above calculation is performed sequentially for all sliding windows and all electrical monitoring dimensions. All time-series change rates are then sorted by window order + dimension order to obtain the set of time-series change rates for each dimension.

[0043] Based on the temporal change rate of each dimension, the state transition relationship between adjacent time steps is constructed to obtain the dimensional state transition matrix. First, the correlation weights of each dimension are extracted. ,in This represents the correlation weight of the d-th electrical monitoring dimension, ranging from 0 to 1. The sum of the correlation weights of all dimensions is 1. The correlation weights are calculated based on the degree of influence of each dimension on the operational risk of the distribution box. The specific calculation process is as follows: A hierarchical structure is constructed, consisting of a target layer, a criterion layer, and a scheme layer. The target layer determines the correlation weights of each electrical monitoring dimension on the operational risk of the distribution box. The criterion layer contains the core evaluation indicators affecting the operational risk of the distribution box, specifically including three criteria: parameter stability, fault correlation, and degree of safety impact. The scheme layer contains eight electrical monitoring dimensions (three-phase voltage, three-phase current, active power, and grid frequency). A judgment matrix is ​​constructed based on a 1-9 scale (where 1 indicates that two indicators are the same). Equal importance, where 3 indicates the former is slightly more important than the latter, 5 indicates the former is significantly more important than the latter, 7 indicates the former is strongly more important than the latter, 9 indicates the former is extremely more important than the latter, 2, 4, 6, and 8 are the median values ​​of the above adjacent scales, and the reciprocal indicates the degree of importance of the latter compared to the former), construct the judgment matrix A1 of the target layer to the criterion layer, and the judgment matrices B1 (parameter stability corresponds to the scheme layer), B2 (fault correlation corresponds to the scheme layer), and B3 (safety impact degree corresponds to the scheme layer) of each criterion layer index to the scheme layer, where the judgment matrix A1 is a 3×3 matrix, and the judgment matrices B1, B2, and B3 are all 8×8 matrices. For example, the judgment matrix A1 of the target layer to the criterion layer is set as follows: The scale values ​​of each element in the judgment matrix correspond to the quantification of the importance of each indicator in the criterion layer relative to the target layer. Elements A1(1,3)=3, A1(1,2)=2, and A1(2,3)=2 correspond to the scale definitions of significantly important and slightly important in the 1-9 scale method, respectively, and are used to quantify the relative importance differences among the indicators in the criterion layer. The maximum eigenvalue of each judgment matrix is ​​calculated. The corresponding feature vectors are then normalized to obtain the weights of each element in each layer relative to the corresponding element in the previous layer. Simultaneously, the consistency index CI is calculated. (where n is the order of the judgment matrix), find the average random consistency index RI for the corresponding order, and calculate the consistency ratio CR = If CR < 0.1, the judgment matrix satisfies the consistency requirement; otherwise, the elements of the judgment matrix are adjusted until the requirement is met. The weights of the criterion layer relative to the target layer and the weights of the scheme layer relative to each indicator of the criterion layer are weighted and summed to obtain the total weight of each monitoring dimension of the scheme layer relative to the target layer, i.e., the correlation weight of each dimension. At the same time, the overall ranking consistency ratio is calculated to ensure that CR < 0.1, thus verifying the rationality of the weight allocation.

[0044] Based on the above calculation process and combined with the actual operating conditions of the distribution box, the correlation weights for three-phase voltage and three-phase current were ultimately determined to be 0.15 each, and the correlation weights for active power and grid frequency were each 0.2, ensuring higher weights for key parameters. Two adjacent sliding windows, k and k+1, were selected to extract the time-series rate of change corresponding to the d-th dimension. and The state transition coefficient is obtained by multiplying the temporal change rate of the next window by the association weight of that dimension. The calculation formula is as follows: ,in Let represent the state transition coefficients between the d-th dimension and the k-th and (k+1)-th sliding windows. Arrange all state transition coefficients by sliding window index (row) + dimension index (column) to form a dimension-state transition matrix: ;

[0045] The matrix has K rows. 1 represents the number of adjacent window pairs, D represents the number of monitoring dimensions, and the matrix elements are the state transition coefficients between corresponding windows and corresponding dimensions.

[0046] Dynamic weight allocation is performed on the dimensional state transition matrix to obtain temporal evolution weight parameters. Based on these weight parameters, a multi-step state recursive mapping process is performed to construct the evolutionary model. Specifically, this includes using each state transition coefficient from the dimensional state transition matrix... Multiply by the association weight of the corresponding dimension The weighted state transition coefficients are obtained, and the calculation formula is as follows: ,in Let represent the weighted state transition coefficient of the k-th window and the d-th dimension. The purpose is to strengthen the influence of state transitions in key dimensions. For each sliding window k, the weighted state transition coefficients of all dimensions within that window are summed to obtain the total weight of that window. The calculation formula is as follows: ,in This represents the total weight of the k-th sliding window, used for subsequent weight normalization, using the state transition coefficients after weighting. Divide by the total weight of the corresponding window Normalization is performed to ensure that the sum of the weights of all dimensions within each window is 1. The calculation formula is as follows: ,in This represents the temporal evolution weight parameter for the k-th window and the d-th dimension.

[0047] The final state of the characteristic coupling state is used as the initial recursive state. This initial recursive state contains normalized values ​​from all electrical monitoring dimensions, ensuring the recursion starting point closely matches the actual operating state. The number of recursion steps is set based on the typical cycle of risk evolution in the distribution box. Considering the actual operating conditions of the distribution box, i.e., the average cycle from risk budding to manifestation is 20 seconds, the number of recursion steps is set to 20 steps, with each step corresponding to a time of 1 second, ensuring coverage of the complete risk evolution process. Let the value of the d-th dimension in the i-th recursion step (i=1,2,...,20) be... The temporal evolution weight parameters for the i-th window and the d-th dimension are: Then, after the i-th step of recursion, the formula for calculating the predicted value of this dimension is: ,in This represents the predicted value of the i-th step recursion and the d-th dimension. The recursion is performed in 20 steps. All dimension values ​​obtained in each step are arranged in chronological order to form a complete state evolution sequence. With the state evolution sequence as the core, combined with the temporal change rate of each dimension and the temporal evolution weight parameters, an evolution model is constructed.

[0048] This embodiment, through the analytic hierarchy process (AHP) and after consistency verification, combines the weights assigned based on the actual operating conditions of the distribution box, and can relatively reasonably reflect the impact of each electrical monitoring dimension on operational risks, thereby strengthening the role of key parameters to a certain extent.

[0049] In a preferred embodiment of the present invention, the evolutionary model is subjected to state recursion, anomaly screening, gradient calculation and smoothing, deflection curvature extraction and spatial mapping reconstruction to construct a preliminary risk state trajectory spectrum; the coordinate parameters of three state anchor points pre-deployed on the distribution box are obtained, a reference field is constructed based on the coordinate parameters of the three state anchor points, the reference field is subdivided to obtain a grid set, the grid density gradient within the grid set is extracted to obtain compensation parameters, and the preliminary risk state trajectory spectrum is calibrated using the compensation parameters to obtain the target risk state trajectory spectrum, which may include:

[0050] The evolutionary model is recursively calculated, and the state data generated by the evolutionary model at each recursion step are arranged into a sequence in chronological order to obtain a state evolution sequence. The state parameters at each moment in the state evolution sequence are compared with the preset steady-state threshold range to filter out state nodes with out-of-range values, thus obtaining an abnormal feature sequence. Specifically, this includes: performing recursive calculation on the constructed evolutionary model, with the number of recursion steps consistent with the number of recursion steps when the evolutionary model was constructed, both being 20 steps; organizing all electrical monitoring dimension state data generated by the evolutionary model at each recursion step in chronological order to obtain a state evolution sequence, which fully presents the evolution process of each electrical parameter over time; comparing the state parameters at each moment in the state evolution sequence with the preset steady-state threshold range to filter out state nodes with out-of-range values ​​to obtain an abnormal feature sequence. The preset steady-state threshold range is set according to the standard range of each electrical parameter when the distribution box is operating normally, combined with the dimensional normalization processing result, i.e., the numerical range of 0~1. Specifically, the three-phase voltage dimension threshold range is set to [0.85, [0.95], the threshold range for the three-phase current dimension is [0.80, 0.90], the threshold range for the active power dimension is [0.75, 0.95], and the threshold range for the grid frequency dimension is [0.90, 1.00]. During the over-limit comparison, the state parameters of each dimension at each moment in the state evolution sequence are compared with the preset steady-state threshold range of the corresponding dimension. If the parameter value of a certain dimension at a certain moment is less than the lower limit of the interval or greater than the upper limit of the interval, the state node at that moment is determined to be an abnormal state node. All abnormal state nodes are sorted in chronological order to obtain an abnormal feature sequence.

[0051] Extracting consecutive adjacent state node pairs from the anomaly feature sequence, calculating the set of differences between the node value at the next time step and the node value at the previous time step, yields the anomaly feature propagation gradient. Extracting the gradient vectors from consecutive adjacent time steps within the anomaly feature propagation gradient yields a gradient vector sequence. Specifically, this involves selecting two consecutive adjacent state nodes from the anomaly feature sequence as node pairs, with the time interval consistent with the recursive step size, both being 1 second. Let the two consecutive adjacent time steps be j and j+1, where j=1,2,... -1, The total number of nodes in abnormal states, with the corresponding values ​​for the d-th dimension being... and ,in Let represent the value of the j-th outlier node in the d-th dimension. The formula for calculating the difference is: ,in Let represent the difference between the d-th dimension and the j-th and (j+1)-th anomalous nodes. Perform the above calculation sequentially on all consecutive adjacent node pairs and all electrical monitoring dimensions. Arrange all differences in node pair order + dimension order to obtain the anomalous feature propagation gradient. This gradient reflects the propagation speed and direction of the anomalous feature. Calculate the gradient vector sequence and the gradient direction change sequence. Arrange the anomalous feature propagation gradient in chronological order. The anomalous feature propagation gradient at each time j is taken as a gradient vector. This vector contains the gradient values ​​of all electrical monitoring dimensions, denoted as . ,in Let represent the gradient vector corresponding to the j-th node. Arrange all gradient vectors in chronological order to obtain the gradient vector sequence.

[0052] Calculate the spatial angle between adjacent vectors in the gradient vector sequence to obtain the gradient direction change sequence; calculate the numerical difference of consecutive angles based on the gradient direction change sequence to extract the trajectory turning rate and obtain the deflection curvature; extract historical gradient data within a preset historical time window based on the gradient propagation of anomaly features to obtain the historical gradient sequence, specifically including: selecting two consecutive adjacent gradient vectors in the gradient vector sequence. and Calculate the spatial angle between the two, and calculate the dot product of the two gradient vectors. The formula is: Then calculate the magnitudes of the two gradient vectors separately, using the following formula: and The formula for calculating the cosine of the included angle is: ,in This represents the spatial angle between two gradient vectors. Finally, the numerical value of the angle is calculated using the following formula: The units are radians. All included angle values ​​are arranged chronologically to obtain a gradient direction change sequence. The numerical differences between consecutive included angles are calculated based on this sequence. The trajectory turning rate is extracted to obtain the deflection curvature. Two consecutive included angle values ​​from the gradient direction change sequence are selected. and The formula for calculating the difference is: ,in This represents the difference between two adjacent angles. The time interval corresponding to the angle is set to 1 second, which is consistent with the time interval of abnormal nodes. The formula for calculating the turning rate is: ,in This represents the trajectory turning rate within the j-th time interval, expressed in radians per second. It is calculated by dividing the trajectory turning rate by the magnitude of the gradient vector at the corresponding time point. The deflection curvature is obtained, and the calculation formula is: ,in The value represents the deflection curvature within the j-th time interval. The larger the value, the more obvious the turning point of the abnormal feature trajectory. The above calculation is performed sequentially for all continuous angles to obtain a complete deflection curvature sequence. The gradient inertia compensation and smooth propagation gradient are calculated. The preset historical time window length and sliding window length are the same, both 5 seconds, corresponding to 5 abnormal nodes. For the gradient vector at each time j, all gradient data within 5 seconds before that time are extracted and arranged in chronological order to obtain the historical gradient sequence.

[0053] The historical gradient sequence is subjected to end-value weighted accumulation processing to obtain the gradient inertia compensation amount. Momentum fusion processing is then performed by superimposing the current value of the gradient propagation based on anomaly features onto the gradient inertia compensation amount to obtain the smoothed propagation gradient. Specifically, this includes calculating the gradient inertia compensation amount using an exponentially decaying weighted accumulation method. The weight coefficients are based on the influence of historical gradients on the current gradient and decay exponentially over time. Setting it to 0.053, this coefficient balances the influence of historical gradients and the current gradient, avoiding excessive or insufficient interference from historical data. In the specific calculation, the first value in the historical gradient data is calculated... The weights of the gradient vectors are given by the following formula: ,in This represents the time interval between the gradient vector and the current time j, in seconds. The weighted gradient value is then calculated using the following formula: ,in This represents the weighted gradient value. The gradient inertia compensation is obtained by summing the weighted gradient values, as shown in the formula: ,That This represents the gradient inertia compensation amount at time j and in the d-th dimension, with a momentum coefficient set. =0.9, this coefficient can balance gradient smoothness and response speed. Too large a value can easily lead to gradient lag. If the gradient is too small, it cannot effectively eliminate fluctuations. The formula for calculating the smooth propagation gradient is: = ,in This represents the smooth propagation gradient.

[0054] Numerical accumulation mapping along the time axis is performed on the smooth propagation gradient to obtain the gradient magnitude path; spatial coordinate alignment is performed based on the gradient magnitude path by applying deflection curvature to obtain trajectory space mapping data; the trajectory space mapping data is then serialized and reconstructed to construct a preliminary risk state trajectory spectrum, specifically including: starting from the initial time j=1, cumulative calculation along the time axis is performed on the smooth propagation gradient, with the initial cumulative value set to 0, and the accumulation formula is: ,in Let represent the cumulative gradient magnitude at time j and dimension d. All cumulative magnitudes are arranged in chronological order to obtain the gradient magnitude path. The cumulative gradient magnitude is used as the basic dimension of the spatial coordinates. Three spatial coordinate dimensions are set, corresponding to the cumulative gradient magnitudes of the three core monitoring dimensions: three-phase voltage, three-phase current, and active power. Combined with the trajectory deflection curvature corresponding to each time moment, the directions of the three spatial coordinate dimensions are adjusted. The adjustment rules are as follows: the value of the first spatial coordinate dimension (corresponding to the cumulative gradient magnitude of three-phase voltage) is equal to its original cumulative gradient magnitude multiplied by the cosine of the deflection curvature; the value of the second spatial coordinate dimension (corresponding to the cumulative gradient magnitude of three-phase current) is equal to its original cumulative gradient magnitude multiplied by the sine of the deflection curvature; and the value of the third spatial coordinate dimension (corresponding to the cumulative gradient magnitude of active power) is equal to its original cumulative gradient magnitude multiplied by (1 plus the sum of the deflection curvatures). The three spatial coordinate values ​​obtained after the above adjustments are the adjusted spatial coordinates. All the adjusted spatial coordinates at all times are arranged in chronological order to obtain the trajectory spatial mapping data.

[0055] The trajectory spatial mapping data at all times, i.e., the adjusted 3D spatial coordinates corresponding to each time point, are uniformly arranged in chronological order. A clear redundancy judgment standard is set: if the absolute value of the coordinate difference in the three axes of two adjacent adjusted 3D spatial coordinates is less than 0.001 (normalized value), it means that the spatial position of these two times has hardly changed, and the coordinate data of the latter time point is invalid and redundant data. Only the coordinate data of the former time point is retained. In this way, redundant information is eliminated, reducing the amount of data while fully preserving the core features of the trajectory and avoiding redundant data from interfering with subsequent analysis. For each time point, the retained adjusted 3D spatial coordinates are matched one by one with the complete electrical parameter anomaly values ​​of the corresponding time point, i.e., the anomaly parameter values ​​of all electrical monitoring dimensions at that time point. This ensures that each spatial coordinate can accurately correspond to the specific electrical parameter anomaly state, realizing a one-to-one association between spatial position and electrical anomaly situation. This ensures that the recombined data contains both spatial information and corresponding electrical anomaly information. The sequence structure is organized by integrating all the data after time-series sorting, redundancy removal, and supplementation of related data in chronological order into a continuous sequence structure. Each time point corresponds to a complete set of data, including the adjusted three-dimensional spatial coordinates of that time point and the abnormal values ​​of electrical parameters at the corresponding time. Through the above processing, a preliminary risk state trajectory spectrum is finally obtained.

[0056] The three-dimensional spatial coordinates of the center point of the main circuit breaker terminal block fixing bolt at the incoming end, the geometric center point of the secondary side terminal block of the current transformer at the outgoing end, and the reference positioning corner point of the magnetic control switch mounting bracket inside the metering box are extracted and used as coordinate parameters for three pre-deployed state anchor points on the distribution box. Specifically, this includes extracting the state anchor point coordinates and constructing a reference field. The state anchor points are three fixed points pre-installed inside the distribution box to construct a spatial reference, corresponding to the incoming side, the outgoing side, and the box door status monitoring side, respectively. The incoming main circuit breaker terminal block fixing bolt center point, the outgoing side, and the metering box door status monitoring side are all extracted. The center point of the circuit breaker terminal block fixing bolt is the first state anchor point. Its three-dimensional coordinates are measured using a high-precision spatial positioning device with a positioning accuracy of ±0.1mm. These three-dimensional coordinates include the coordinate values ​​of each of the three spatial axes. The geometric center point of the secondary side terminal block of the current transformer at the outgoing end is the second state anchor point. The measurement method is the same as that for the first state anchor point, measuring the coordinate values ​​of its three spatial axes. The reference positioning corner point of the magnetic control switch mounting bracket inside the metering box door is the third state anchor point. The measurement method is also the same as that for the first state anchor point, measuring the coordinate values ​​of its three spatial axes.

[0057] Based on the coordinate parameters of the three state anchor points, a spatial bounding surface is calculated. The spatial points corresponding to each coordinate parameter are connected to generate a closed spatial boundary, defining the effective range of the risk state deduction and obtaining the reference field. The reference field is then divided into multiple discrete 3D mesh units along the orthogonal spatial axes, resulting in a mesh set. Specifically, using the three state anchor points as the three vertices of a spatial triangle, the projected polygons of this spatial triangle on the three projection planes are first solved. The projected coordinates of the three state anchor points on these three projection planes are extracted, and these coordinates are connected in a clockwise order to form the projected polygons of the triangle. Then, the algorithm for finding the maximum inscribed rectangle of each projected polygon is applied. Taking a projection plane as an example, first calculate the minimum bounding rectangle of the projection polygon to obtain the value intervals of the rectangle in the two corresponding spatial axis directions, namely the value interval of the first spatial axis (minimum value of the first spatial axis, maximum value of the first spatial axis) and the value interval of the second spatial axis (minimum value of the second spatial axis, maximum value of the second spatial axis). The minimum value of the first spatial axis is the minimum value of the coordinates of the three state anchor points on the spatial axis, and the maximum value of the first spatial axis is the maximum value of the coordinates of the three state anchor points on the spatial axis. The minimum value of the second spatial axis is the minimum value of the coordinates of the three state anchor points on the spatial axis, and the maximum value of the second spatial axis is the maximum value of the coordinates of the three state anchor points on the spatial axis.

[0058] Set the search step size to 0.1 mm to fit the subsequent mesh generation. Traverse all candidate rectangles within the search range whose edges are parallel to the coordinate axes, ensuring that the vertices of each candidate rectangle are inside the projected polygon or on its boundary. The formula for calculating the area of ​​the rectangle is: = ,in These are the spatial coordinates of the first axis of each of the two diagonal vertices of the rectangle. The coordinates of the second spatial axis of each of the two opposite vertices of the rectangle are used to select the rectangle with the largest area, which is then used as the maximum inscribed rectangle of the projection plane. The parameters of the maximum inscribed rectangle of the three projection planes are integrated, and the intersection of the ranges of each spatial axis is taken to calculate the center and radius of the circumsphere of the spatial triangle with the three state anchor points as vertices. The center of the circumsphere is the intersection of the perpendicular bisectors of the three sides of the spatial triangle, and the space inside this sphere is the reference field. Equal-interval meshing is performed along the three orthogonal spatial axes to generate a mesh set. The meshing interval is set to 10mm, which balances spatial accuracy and computational efficiency, accurately capturing trajectory nodes without causing insufficient computational power at the edges due to an excessive number of meshes. In the specific calculation, the number of meshes for each of the three spatial axes is first calculated. Let the three spatial axes be axis 1, axis 2, and axis 3, and the minimum values ​​for each axis are... The maximum values ​​are respectively If the split spacing is h=10mm, then the formulas for calculating the number of splits for each shaft are as follows: , ,in This indicates a rounding up operation, ensuring that the entire spatial extent is covered by the grid. The reference field is then divided into... There are three discrete 3D mesh elements. The spatial range of each mesh element is the 3D space formed by the intervals between two adjacent partition points on axis 1, the intervals between two adjacent partition points on axis 2, and the intervals between two adjacent partition points on axis 3. The coordinates of the p-th partition point on each axis are calculated using the following formulas: Axis 1: Axis 2: , Axis 3: p is the subdivision point number; all discrete 3D mesh elements are arranged in spatial coordinate order to obtain a mesh set.

[0059] The coordinates of trajectory state nodes arranged in time sequence are extracted from the preliminary risk state trajectory spectrum. These coordinates are then mapped and projected one by one onto a grid set to obtain the trajectory projection distribution set. The trajectory projection distribution set is traversed to count the number of trajectory state node coordinates falling within each 3D grid cell, resulting in a sequence of trajectory state node counts. Specifically, considering the actual needs of distribution box operation status monitoring, the preliminary risk state trajectory spectrum is a spatial sequence set that, after time-series sorting, redundancy removal, and supplementation with associated data, can present the evolution trajectory and change patterns of abnormal characteristics. Its core data consists of the three-dimensional spatial coordinates corresponding to each moment, corresponding to the cumulative gradient amplitude adjusted data of the three core monitoring dimensions: three-phase voltage, three-phase current, and active power. From the preliminary risk state trajectory spectrum, all trajectory state node coordinates arranged in chronological order are completely extracted. Each trajectory state node coordinate is a three-dimensional spatial coordinate, and the extraction process strictly adheres to the requirement of time series continuity. To ensure that the coordinate data at each moment is complete, without omission, duplication, or disorder, and fully corresponds to the temporal logic of the preliminary risk state trajectory spectrum, the extracted coordinate data is guaranteed to accurately reflect the temporal evolution of the abnormal characteristics. The coordinates of each extracted trajectory state node are mapped and projected one by one onto a pre-constructed grid set. The grid set is a set of discrete 3D grid cells obtained by equally spaced subdivision of the reference field. The specific mapping and projection operation involves determining which 3D coordinate value of each trajectory state node falls within the spatial range of a given 3D grid cell. The criterion is that if the 3D coordinate values ​​of a trajectory state node are all within the range of the three spatial axes of a given 3D grid cell, then the trajectory state node is considered to fall within that 3D grid cell. After completing the mapping and projection of all trajectory state nodes, the grid cell information corresponding to each trajectory state node is associated and organized with the node coordinates themselves to form a trajectory projection distribution set.

[0060] Based on the need for trajectory distribution analysis in the risk monitoring of distribution boxes, it is necessary to clarify the density of trajectory nodes in each grid unit. Therefore, the trajectory projection distribution set is traversed one by one according to the spatial order of each three-dimensional grid unit in the grid set. During the traversal, the number of trajectory status nodes in each three-dimensional grid unit is counted. When counting, each trajectory status node in the trajectory projection distribution set is checked one by one to confirm the three-dimensional grid unit corresponding to the node. Each time a node is confirmed to fall into the grid unit, the number of nodes in the grid unit is incremented by one. This operation is continued until all trajectory status nodes are checked and counted. After the count is completed, the number of nodes corresponding to each three-dimensional grid unit is arranged in the spatial order of each three-dimensional grid unit in the grid set to form a sequence of trajectory status node numbers.

[0061] The local proportion of trajectory state node coordinates within each 3D grid cell is calculated based on the sequence of trajectory state node numbers. These local proportions are then arranged in spatial order to obtain a grid density distribution sequence. Difference extraction is performed on the local proportions of adjacent 3D grid cells within the grid density distribution sequence to obtain the grid density gradient. Specifically, this involves: calculating the local proportion of trajectory nodes within each 3D grid cell to accurately reflect their density; calculating the total number of trajectory state nodes by summing the node counts for all 3D grid cells in the sequence, which serves as the basis for calculating the local proportion; and dividing the number of trajectory state nodes within each 3D grid cell by the total number of trajectory state nodes to obtain the local proportion of the coordinates of the trajectory state nodes within that 3D grid cell. After calculating the local proportions of all grid cells, these local proportions are arranged in spatial order within the grid set to form a grid density distribution sequence.

[0062] The grid density gradient reflects the spatial variation trend of trajectory node density within the grid set, providing a core basis for subsequent compensation parameter calculations. Therefore, it is necessary to perform difference extraction processing on the grid density distribution sequence. Following the order of the grid density distribution sequence, the local proportion values ​​corresponding to two adjacent 3D grid cells are selected one by one. The local proportion value of the latter 3D grid cell is subtracted from the local proportion value of the former 3D grid cell to obtain the local proportion difference between the two adjacent 3D grid cells. This difference reflects the magnitude and direction of the change in trajectory node density between the two adjacent grid cells. Since the grid set is a three-dimensional spatial structure, the above difference extraction processing needs to be performed on all adjacent 3D grid cells along the three spatial axes to ensure that the variation of grid density throughout the entire three-dimensional space can be fully captured. After completing the difference extraction of all adjacent grid cells, all extracted local proportion differences are arranged according to spatial direction and grid order to obtain the grid density gradient.

[0063] Numerical scaling is performed based on the grid density gradient, and the scaled values ​​are mapped to a preset correction interval to obtain compensation parameters. Spatial position deviation correction calculations are then performed on the trajectory state node coordinates based on these compensation parameters to obtain a calibrated trajectory coordinate sequence. Node continuum reorganization processing is then performed on the calibrated trajectory coordinate sequence to construct the target risk state trajectory spectrum. Specifically, considering the inconsistent numerical range of the grid density gradient, which may exhibit both positive and negative differences and large numerical spans, it cannot be directly used as a basis for trajectory deviation compensation. Therefore, scaling is required to bring it into a uniform quantization range and map it to a preset correction interval to obtain compensation parameters for subsequent trajectory node coordinate deviation correction. A comprehensive statistical analysis of all difference values ​​in the grid density gradient is performed, comparing each difference value individually to accurately identify the maximum and minimum values, clarifying the specific numerical range of the grid density gradient, and laying a precise data foundation for subsequent scaling. Scaling is then performed on each grid density gradient value separately. The specific method involves subtracting the minimum value of the current grid density gradient from the current value, and then dividing the difference by the difference between the maximum and minimum values ​​of the grid density gradient. This calculation method maps all grid density gradient values ​​to a uniform range of 0 to 1, ensuring that the scaled values ​​are within a consistent quantization range and eliminating the influence of different gradient value spans, thus facilitating subsequent mapping processing. After scaling, the scaled values ​​are mapped to a preset correction range, which is preset to 0.9 to 1.1. This range setting aligns with the actual needs of trajectory node coordinate deviation correction, effectively correcting deviations while avoiding over-correction that could lead to trajectory distortion. The mapping method involves multiplying the scaled value by 0.2 and then adding 0.9. Multiplying by 0.2 stretches the scaled value in the 0 to 1 range to the 0 to 0.2 range, which matches the correction range span, while adding 0.9 shifts the stretched value to the preset correction range of 0.9 to 1.1. This calculation accurately obtains the compensation parameters corresponding to each grid cell.

[0064] Because the coordinates of trajectory status nodes may be slightly deviated during the generation process due to factors such as abnormal gradient fluctuations, spatial mapping deviations, and minor errors in monitoring equipment, these deviations can affect the accuracy of subsequent risk trajectory analysis. Therefore, targeted corrections using compensation parameters are necessary to ensure that the trajectory coordinates accurately reflect the actual situation of the distribution box's risk evolution. The process involves determining the 3D mesh cell corresponding to each trajectory status node, accurately matching the 3D mesh cell to which each trajectory status node belongs based on the correlation between nodes and mesh cells recorded in the trajectory projection distribution, and then obtaining the corresponding compensation parameters for that 3D mesh cell to ensure that each node... The corresponding compensation parameters can be accurately matched. For the three-dimensional coordinates of each trajectory state node, the corresponding compensation parameter is multiplied by the coordinate values ​​of the three spatial axes of the node (X-axis, Y-axis, and Z-axis) to obtain the corrected values ​​of each spatial axis coordinate. After the three spatial axis coordinates are corrected, the three corrected coordinate values ​​are integrated to form the calibrated coordinates of the trajectory state node. The calculation accuracy is strictly controlled during the correction process to avoid coordinate deviations caused by calculation errors. After the coordinate correction of all trajectory state nodes is completed, the calibrated coordinates of all trajectory state nodes are sorted out one by one in strict chronological order to obtain the calibrated trajectory coordinate sequence.

[0065] Although the calibrated trajectory coordinate sequence has undergone deviation correction and improved coordinate accuracy, there may still be discontinuities due to excessively large time intervals between adjacent nodes. This prevents the complete representation of the continuous trajectory of risk evolution, hindering subsequent risk feature extraction and risk level determination. Therefore, node continuum reorganization is necessary to construct a target risk state trajectory spectrum that accurately reflects the spatial trajectory of the distribution box's risk evolution. A time-series verification of the calibrated trajectory coordinate sequence is performed, checking the time order of each calibrated trajectory node coordinate to ensure that all coordinates are strictly arranged in chronological order, without any temporal discrepancies, missing elements, or duplications. If any temporal anomalies are found, adjustments are made promptly to ensure that all coordinate data are continuous and consistent in the time dimension. Subsequent continuous processing lays a solid foundation. Continuous processing is performed on the calibrated trajectory coordinate sequence. For every two adjacent trajectory node coordinates, transition coordinates are added between them. The transition coordinates are determined by generating transition points at even intervals based on the X-axis, Y-axis, and Z-axis coordinate values ​​of the two adjacent nodes. The number of transition points is reasonably set according to the coordinate difference between adjacent nodes. When the coordinate difference is large, the number of transition points is appropriately increased, and when the coordinate difference is small, the number of transition points is appropriately reduced. This ensures the continuity of the trajectory while avoiding data redundancy due to too many transition points, ensuring that the trajectory can completely present the continuous process of risk evolution. All calibrated trajectory node coordinates and supplemented transition coordinates are strictly integrated and recombined according to the time series to form a complete target risk state trajectory spectrum.

[0066] In this embodiment, a reference field is constructed based on three preset state anchor points of the distribution box and a mesh set is generated. The compensation parameters are calculated by combining the mesh density gradient, so as to achieve accurate calibration of the preliminary trajectory spectrum, effectively correct coordinate deviations caused by gradient fluctuations, equipment errors, etc., improve the accuracy of trajectory coordinates, and make the risk trajectory more in line with the actual operating conditions.

[0067] In a preferred embodiment of the present invention, feature extraction is performed on the target risk state trajectory spectrum to construct a steady-state boundary set, and the spatial feature deviation and cumulative anomaly confidence value of the risk trajectory in the target risk state trajectory spectrum relative to the steady-state boundary set are calculated; when the weighted fusion index of the spatial feature deviation and cumulative anomaly confidence value exceeds the steady-state boundary set, the edge-side real-time state transition judgment logic is executed to obtain the distribution box operation risk level label, which may include:

[0068] Extract the trajectory coordinate data of the target risk state trajectory spectrum arranged continuously in time series to obtain the trajectory feature sequence; perform numerical distribution statistical processing on the trajectory feature sequence to extract the statistical envelope parameters of the normal operation interval, and obtain the steady-state envelope parameter set. Specifically, the target risk state trajectory spectrum is a data set that accurately reflects the spatial trajectory of the risk evolution of the distribution box after calibration and continuous reconstruction. Its core data is the calibrated three-dimensional spatial coordinates corresponding to each moment, corresponding to the cumulative gradient amplitude adjusted data of the three core monitoring dimensions of three-phase voltage, three-phase current, and active power; in order to successfully construct the steady-state boundary set and calculate the risk trajectory deviation, it is necessary to... From the target risk state trajectory spectrum, all trajectory coordinate data arranged continuously in chronological order are extracted. Each trajectory coordinate data is a three-dimensional spatial coordinate after deviation correction. The extraction process strictly follows the requirement of time series continuity, characterizing the trajectory coordinates of each time point one by one to ensure that no trajectory coordinates of any time point are missed and that there is no duplicate extraction. The extracted trajectory coordinate data are organized one by one according to the time series to form a trajectory feature sequence. For all trajectory coordinate data in the trajectory feature sequence, numerical distribution statistical processing is performed on the three spatial axes of X-axis, Y-axis, and Z-axis, and the distribution of coordinate values ​​on each spatial axis is statistically analyzed. This includes the central tendency, dispersion, and extreme value distribution of coordinate values, comprehensively understanding the distribution patterns of coordinate values ​​on each spatial axis, and clarifying the fluctuation range of coordinate values ​​under normal operating conditions. Specifically, the statistical parameters for central tendency include the mean (range 0.1~0.8, corresponding to the dimensionally normalized value) and median (range 0.08~0.75); the statistical parameters for dispersion include the standard deviation (range 0.05~0.3) and variance (range 0.0025~0.09); and the statistical parameters for extreme value distribution include the maximum and minimum values, with a range consistent with the dimensionally normalized coordinate value range, 0~1. This is combined with the distribution box... The standard ranges of various electrical parameters during normal operation (standard range of three-phase voltage is 220V±10%, standard range of three-phase current is 0~100A, standard range of active power is 0~22kW), and the dimensionally normalized numerical range (0~1), are used to accurately extract statistical envelope parameters that reflect the normal operation range from the statistical results. These statistical envelope parameters can comprehensively cover the core range of trajectory coordinate values ​​under normal operation, accurately define the coordinate fluctuation range during normal operation, and avoid deviations in the definition of the normal range due to missing parameters. These statistical envelope parameters are classified and organized according to the three spatial axes of X-axis, Y-axis, and Z-axis to form a steady-state envelope parameter set.

[0069] Based on the steady-state envelope parameter set, upper and lower threshold values ​​are extracted. These threshold values ​​are then processed to construct intervals, resulting in a steady-state boundary set. The trajectory coordinate data is then mapped and compared with the steady-state boundary set. The shortest vertical distance from each trajectory coordinate data point to its corresponding steady-state boundary is calculated, yielding a normal distance sequence. Specifically, for each spatial axis in the steady-state envelope parameter set, the upper and lower threshold values ​​for the spatial axis coordinates are extracted one by one. The upper threshold is the maximum value of the spatial axis coordinate under normal operating conditions, ranging from 0.7 to 1.0 (after dimension normalization), fully covering the maximum fluctuation range of the axis coordinate under normal operating conditions. The lower threshold is the minimum value of the spatial axis coordinate under normal operating conditions, ranging from 0 to 0.3 (after dimension normalization), fully covering the minimum fluctuation range of the axis coordinate under normal operating conditions. During the extraction process, the threshold values ​​are strictly compared with the statistical results and repeatedly checked to ensure accuracy and conformity to the statistical characteristics of the normal operating range, avoiding misjudgments in subsequent risk assessments due to threshold deviations. After the threshold extraction is completed, the upper and lower thresholds of each spatial axis are constructed into intervals to form the normal operation interval corresponding to each spatial axis. This interval can completely cover all possible values ​​of the axis coordinates during normal operation, without omissions or exceeding the normal range. The normal operation intervals of the three spatial axes are integrated to form the steady-state boundary set of the three-dimensional spatial range.

[0070] Each trajectory coordinate data in the trajectory feature sequence is compared with the steady-state boundary set through position mapping. The positional relationship of each trajectory coordinate data relative to the steady-state boundary in the three spatial axes (X-axis, Y-axis, and Z-axis) is characterized one by one to determine whether the trajectory coordinate data is inside the steady-state boundary set. For each trajectory coordinate data, the shortest vertical distance to the steady-state boundary of the corresponding spatial axis is calculated. The calculation method is as follows: if the trajectory coordinate data is inside the steady-state boundary set, it means that the risk trajectory is within the normal operating range at that moment and there is no deviation, and the distance value is 0; if the trajectory coordinate data is outside the steady-state boundary set, it means that the risk trajectory deviates from the normal operating range at that moment, and the distance value is the vertical distance from the trajectory coordinate data to the steady-state boundary, with a value range of 0.01~1.0 (after dimension normalization). The shortest vertical distances corresponding to all trajectory coordinate data are organized one by one according to the time series to obtain the normal distance sequence.

[0071] The normal distance sequence undergoes boundary violation screening to extract distance value segments located outside the steady-state boundary set, resulting in a boundary violation distance sequence. Based on this sequence, the time span corresponding to each distance value segment is extracted to determine the duration of the deviation. Specifically, this involves setting clear boundary violation criteria. Combining this with the definition of the steady-state boundary set, when a distance value in the normal distance sequence is greater than 0, it indicates that the corresponding trajectory coordinate data is outside the steady-state boundary set and is considered boundary violation data; at this moment, the risk trajectory deviates from the normal operating range. When a distance value is equal to 0, it indicates that the corresponding trajectory coordinate data is inside the steady-state boundary set and is not considered boundary violation data; at this moment, the risk trajectory... The trajectory is within the normal operating range; according to this judgment criterion, each distance value in the normal distance sequence is filtered one by one, and all distance values ​​greater than 0 are accurately filtered out. The value range of this type of value is 0.01~1.0 (after dimension normalization). These out-of-bounds distance values ​​are organized into a time series to form continuous distance value segments. Each segment corresponds to a continuous out-of-bounds time, which can clearly reflect the situation of the risk trajectory continuously going out of bounds within a certain time period. The start time of the segment is the time corresponding to the first out-of-bounds distance value, and the end time of the segment is the time corresponding to the last out-of-bounds distance value; after integrating all segments, the out-of-bounds distance sequence is obtained.

[0072] The duration of deviation reflects the length of time a risk trajectory crosses the boundary and is an important basis for judging the severity of the risk. The longer the duration, the more prominent the risk. Therefore, it is necessary to extract the time span corresponding to each distance value segment from the boundary distance sequence. For each continuous distance value segment in the boundary distance sequence, the start time point and end time point corresponding to the segment are determined one by one. The start time point is the time corresponding to the first boundary distance value of the segment, and the end time point is the time corresponding to the last boundary distance value of the segment. During the determination process, the time points are strictly compared with the time sequence and repeatedly checked to ensure that the time points are accurate and without deviation. Subtracting the start time point from the end time point yields the time span corresponding to the distance value segment, which is the duration of deviation of the boundary trajectory, with a value range of 1s to 60s.

[0073] The comprehensive deviation magnitude of each trajectory coordinate data relative to the steady-state boundary set is calculated based on the out-of-bounds distance sequence to obtain the spatial feature deviation. The spatial feature deviation is then continuously accumulated along the time evolution direction, and confidence weight mapping is performed in conjunction with the deviation duration to obtain the cumulative anomaly confidence value. Specifically, for each out-of-bounds distance value in the out-of-bounds distance sequence, the comprehensive deviation magnitude is calculated based on the deviation in the X, Y, and Z spatial axes corresponding to that value. The calculation method involves accumulating the out-of-bounds distance values ​​of the trajectory coordinate data in the three spatial axes to obtain the sum of the deviation distances in the three axes, and then dividing this sum by 3 to obtain the average value. This average value is the comprehensive deviation magnitude of the trajectory coordinate data, which is the spatial feature deviation at that time point, and its value range is 0.01~1.0 (after dimension normalization). Starting from the beginning of the time series, the spatial characteristic deviation is continuously accumulated along the time evolution direction. The accumulation method is that the cumulative value at each time point is equal to the cumulative value at the previous time point plus the spatial characteristic deviation at the current time point. If there is no out-of-bounds situation at the current time point, that is, the spatial characteristic deviation is 0, the cumulative value remains unchanged. Through this accumulation method, the accumulation process of risk anomalies can be clearly reflected. The larger the cumulative value, the more serious the risk anomaly accumulation. The cumulative anomaly confidence value ranges from 0 to 10.0. The larger the value, the higher the degree of anomaly accumulation. Simultaneously, confidence weight mapping is performed based on the duration of the deviation. The mapping rule is that the longer the duration of the deviation, the greater the corresponding confidence weight. The confidence weight ranges from 0.5 to 1.0. Specifically, the mapping relationship is as follows: 0.5s~5s corresponds to a confidence weight of 0.5~0.6; 5s~20s corresponds to a confidence weight of 0.6~0.8; and 20s~60s corresponds to a confidence weight of 0.8~1.0. This is because the longer the duration of the deviation, the higher the credibility of the risk anomaly and the better it reflects the real risk hazard, and vice versa. The spatial characteristic deviation at each time point is multiplied by the corresponding confidence weight and then included in the continuous cumulative calculation to obtain the cumulative anomaly confidence value at each time point.

[0074] Spatial feature deviation and cumulative anomaly confidence value are extracted. A weighted summation is performed based on preset deviation weight coefficients and confidence weight coefficients to obtain a weighted fusion index. The weighted fusion index is then compared numerically with the upper threshold of the steady-state boundary set. The difference between the weighted fusion index and the upper threshold is calculated to obtain the out-of-bounds deviation difference. Specifically, this involves: extracting the spatial feature deviation and cumulative anomaly confidence value for each time point, verifying each index to ensure a one-to-one correspondence in the time dimension without misalignment or omission, providing an accurate data foundation for the weighted summation; and presetting deviation weight coefficients and confidence weight coefficients, with the sum of the two coefficients being 1, where the deviation weight coefficient ranges from 0. The confidence weight coefficient ranges from 0.6 to 0.7, with the weight coefficient set based on the degree of influence of each indicator on risk assessment. The confidence weight coefficient is greater than the deviation weight coefficient, highlighting the impact of the degree of anomaly accumulation on risk assessment. This is because the accumulation of risk anomalies often reflects the true risk hazards and the development trend of risks better than deviations at a single moment. For each time point, the spatial feature deviation is multiplied by the deviation weight coefficient to obtain the weighted value of the spatial feature deviation. Then, the cumulative anomaly confidence value is multiplied by the confidence weight coefficient to obtain the weighted value of the cumulative anomaly confidence value. The two weighted values ​​are added together to obtain the weighted fusion index for that time point, with a value range of 0 to 7.0.

[0075] To determine whether the risk exceeds the normal range and to clarify the degree of exceedance, it is necessary to compare the weighted fusion index with the upper limit threshold of the steady-state boundary set and calculate the deviation difference. The upper limit threshold of the weighted fusion index corresponding to the steady-state boundary set is extracted. This upper limit threshold is calculated based on the steady-state envelope parameter set, with a value range of 1.0 to 2.0, which can accurately define the critical value between normal and abnormal risks, ensuring the reasonableness and accuracy of the threshold, without being too high or too low. The weighted fusion index at each time point is compared with this upper limit threshold. If the weighted fusion index is less than or equal to the upper limit threshold, it indicates that the risk at that time point is within the normal range, with no exceedance, and the deviation difference is 0. If the weighted fusion index is greater than the upper limit threshold, it indicates that the risk at that time point exceeds the normal range. The difference between the weighted fusion index and the upper limit threshold is the deviation difference at that time point, with a value range of 0.01 to 5.0.

[0076] Based on the deviation difference from the boundary, a level mapping process is performed according to the preset risk classification threshold range to convert the deviation difference into the corresponding risk quantification level, obtaining the initial risk judgment result. The initial risk judgment result is then subjected to a time-series continuity check process with the deviation duration input edge-side real-time state transition judgment logic to obtain a steady-state risk transition flag. Based on the steady-state risk transition flag, the corresponding risk level code is extracted, and tag generation and encapsulation processes are performed to obtain the distribution box operation risk level tag. Specifically, this includes: a preset risk classification threshold range, which is set according to the actual judgment requirements of different risk levels of the distribution box, divided into multiple continuous ranges. Each range corresponds to a risk quantification level, with risk quantification levels corresponding to different risk degrees from low to high. The larger the deviation difference from the boundary, the higher the corresponding risk level. The higher the risk quantification level, the more severe the risk. The division of the intervals aligns with the actual risk control needs of the distribution box operation, ensuring the rationality of the level mapping. The specific grading intervals and corresponding quantification levels are as follows: deviation difference of 0~0.5 corresponds to risk quantification level 1 (low risk); 0.5~2.0 corresponds to risk quantification level 2 (medium risk); 2.0~5.0 corresponds to risk quantification level 3 (high risk). For the deviation difference at each time point, the interval in which the deviation difference is located is determined one by one by referring to the preset risk grading threshold interval, and then the corresponding risk quantification level is accurately mapped. During the mapping process, the interval range is strictly checked to ensure that the mapping is accurate and there is no level mismatch. The risk quantification levels of all time points are organized in time sequence to obtain the initial risk judgment result.

[0077] The initial risk assessment result only reflects the preliminary risk level at a single point in time. It may be misjudged due to factors such as short-term fluctuations and monitoring errors, and cannot accurately reflect the true trend of risk level changes. Therefore, it is necessary to combine the deviation from the continuous period and perform time sequence continuity verification through the real-time state transition judgment logic on the edge side to obtain the steady-state risk transition indicator and eliminate the influence of short-term fluctuations. The real-time state transition judgment logic on the edge side is deployed on the edge computing terminal, which is adapted to the needs of real-time computing and limited computing power at the edge. The core is to perform verification processing, which fits the actual working conditions of limited computing power and the need for rapid response on the edge side of the distribution box, and can realize rapid verification of risk level. During the verification process, it is determined whether the risk quantification level in the initial risk assessment result shows a continuous increase or remains at a high-risk level. Simultaneously, considering the deviation duration, a preset threshold for the duration of the risk quantification level is set at 2-5 seconds. If the duration of a certain risk quantification level reaches this preset threshold (2-5 seconds) and there is a clear level transition (e.g., from level 1 to level 2 or above, or from level 2 to level 3), it indicates that the risk level has undergone a stable change and is not caused by short-term fluctuations, thus generating a steady-state risk transition indicator. If the above conditions are not met, it indicates that the risk level change may be caused by short-term fluctuations, and no steady-state risk transition indicator is generated. The steady-state risk transition indicator is used to identify whether a stable transition in the risk level has occurred, avoiding misjudgments caused by short-term fluctuations.

[0078] Each risk quantification level is pre-coded with a corresponding risk level code. The code uses a simple and easily recognizable character combination, corresponding one-to-one with the risk level. Specifically, risk quantification level 1 (low risk) corresponds to code L01, risk quantification level 2 (medium risk) corresponds to code M02, and risk quantification level 3 (high risk) corresponds to code H03. This clearly distinguishes different risk levels, facilitating data transmission at the edge and rapid identification by staff. The coding is tailored to the needs of edge data transmission, ensuring efficient and error-free transmission. Based on steady-state risk transition markers, the corresponding risk level code is extracted. If a steady-state risk transition marker exists, it is extracted... The risk level code after the transition is obtained. This code corresponds to the risk level after the stable transition and can reflect the true trend of risk change. If there is no steady-state risk transition indicator, the current continuous risk level code is extracted. This code corresponds to the current stable risk level. The extracted risk level code is processed to generate a label that can clearly identify the risk level. The label contains the risk level name and the corresponding code information. The label is encapsulated to ensure that it can be adapted to edge-side data transmission and subsequent risk warning processing, avoiding loss or errors during data transmission. Finally, the distribution box operation risk level label is obtained.

[0079] This embodiment effectively improves the accuracy and reliability of risk trajectory monitoring for distribution boxes through a complete trajectory processing, calibration, and risk assessment process.

[0080] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0081] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0082] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An edge-end real-time inference system for the operational risk of distribution boxes in an Internet of Things (IoT) environment, characterized in that, include: The acquisition module is used to acquire multi-source data streams, perform time-series alignment and dimensional normalization on the multi-source data streams to construct an initial state feature set, and calculate the correlation weights between each dimension within the initial state feature set. A potential field model is constructed based on the correlation weights, and the feature nodes within the potential field model are topologically reorganized to obtain the feature coupled state. The extraction module is used to input the feature coupled state into the edge-side temporal inference network, extract the temporal change rate of each dimension in the feature coupled state, and construct the evolution model; The computation module is used to perform state recursion, anomaly screening, gradient calculation and smoothing, deflection curvature extraction and spatial mapping reconstruction on the evolution model to construct a preliminary risk state trajectory spectrum. Obtain the coordinate parameters of three pre-deployed state anchor points on the distribution box, and construct a reference field based on the coordinate parameters of the three state anchor points, including: The three-dimensional spatial coordinates of the center point of the fixing bolt of the main circuit breaker terminal block at the incoming end, the geometric center point of the secondary side terminal block of the current transformer at the outgoing end, and the reference positioning corner point of the magnetic control switch mounting bracket inside the metering box are extracted and used as coordinate parameters for three state anchor points pre-laid on the distribution box. Based on the coordinate parameters of the three state anchor points, spatial bounding surface calculation is performed, and the spatial position points corresponding to each coordinate parameter are connected to generate a closed spatial boundary, delineating the effective range of risk state deduction and obtaining the reference field. The reference field is divided into grid sets, the grid density gradient within the grid sets is extracted to obtain compensation parameters, and the preliminary risk state trajectory spectrum is calibrated using the compensation parameters to obtain the target risk state trajectory spectrum. The output module is used to perform feature extraction on the target risk state trajectory spectrum, construct a steady-state boundary set, and calculate the spatial feature deviation and cumulative anomaly confidence value of the risk trajectory in the target risk state trajectory spectrum relative to the steady-state boundary set. When the weighted fusion index of spatial feature deviation and cumulative anomaly confidence value exceeds the steady-state boundary set, the edge-side real-time state transition judgment logic is executed to obtain the distribution box operation risk level label.

2. The edge-end real-time inference system for the operational risk of distribution boxes in an IoT environment according to claim 1, characterized in that, Collect multi-source data streams, perform time-series alignment and dimensional normalization on the multi-source data streams to construct an initial state feature set, and calculate the correlation weights between each dimension within the initial state feature set, including: The system collects three-phase voltage sampling values, three-phase current sampling values, active power values, grid frequency values, and door status signals during the operation of the distribution box to obtain a multi-source data stream. The system then parses the timestamps of the multi-source data stream and performs interpolation filling on the time-missing segments in the multi-source data stream to obtain a continuous time-series data stream. Using the master clock channel of the continuous time-series data stream as the synchronization reference, a fixed-step resampling process is performed on the continuous time-series data stream to obtain a time-aligned sequence; the maximum and minimum values ​​of each electrical parameter channel in the time-aligned sequence are extracted and the range set is obtained; the time-aligned sequence is subjected to dimensional scaling based on the range set to obtain a dimensional normalized matrix. The dimensionally normalized matrix is ​​column-wise concatenated according to the preset electrical monitoring dimensions to construct the initial state feature set; any two dimension data columns in the initial state feature set are traversed, and the corresponding data points of any two dimension data columns in the preset sliding window are extracted and covariance matching is performed to obtain the dimension covariance matrix; the diagonal elements of the dimension covariance matrix are extracted and square root transformation is performed to obtain the dimension standard deviation vector. The off-diagonal elements of the dimensional covariance matrix are mapped proportionally to the corresponding components of the dimensional standard deviation vector to obtain the dimensional correlation coefficient matrix; the dimensional correlation coefficient matrix is ​​then subjected to exponential decay weighting and non-negative truncation to obtain the dimensional interaction weight matrix. The data in each row of the dimensional interaction weight matrix is ​​summed to obtain the total row sum. The data in each row of the dimensional interaction weight matrix is ​​then normalized to the ratio of the total row sum to calculate the correlation weight between each dimension in the initial feature set.

3. The edge-end real-time inference system for the operational risk of distribution boxes in an IoT environment according to claim 2, characterized in that, A potential field model is constructed based on the correlation weights. The feature nodes within the potential field model are then topologically reorganized to obtain the feature coupled states, including: The associated weights are matched with the initial state feature set by dimension mapping, and the potential field intensity parameters corresponding to each electrical monitoring dimension are extracted to obtain the potential field initialization parameter set. Based on the potential field initialization parameter set, the initial state feature set is processed by spatial potential energy projection, and the data of each dimension in the initial state feature set is converted into coordinate nodes in the potential field space. Based on the coordinate nodes, the potential field space basis is established to construct the potential field model and the initial distribution set of feature nodes. Based on the initial distribution set of feature nodes and the associated weights, the potential energy gradient between nodes is deduced to obtain the potential energy gradient matrix. Threshold filtering and adjacency determination are performed on the potential energy gradient matrix to retain the node connection paths with potential energy gradients higher than the preset critical value, thus obtaining the potential field topology connection graph. Based on the potential field topology connection graph, the initial distribution set of feature nodes is subjected to spatial position offset and high-dimensional aggregation processing to obtain the topology reorganized feature set; the topology reorganized feature set and the potential field topology connection graph are subjected to structured mapping and fusion processing to obtain the feature coupled state.

4. The edge-end real-time inference system for the operational risk of distribution boxes in an IoT environment according to claim 3, characterized in that, The feature-coupled state is input into the edge-side temporal extrapolation network to extract the temporal change rate of each dimension in the feature-coupled state and construct an evolutionary model, including: The feature-coupled state is input into the edge-side temporal extrapolation network to perform channel feature parsing processing, resulting in the network's initial response sequence. Numerical difference features of continuous time steps are extracted from the network's initial response sequence to obtain the dimensional difference feature set. A rate mapping process within a sliding window is performed on the dimensional difference feature set to obtain the temporal change rate of each dimension; the state transition relationship between adjacent time steps is constructed based on the temporal change rate of each dimension to obtain the dimensional state transition matrix; Dynamic weight allocation is performed on the dimensional state transition matrix to obtain temporal evolution weight parameters; multi-step state recursive mapping is performed based on the temporal evolution weight parameters to construct the evolution model.

5. The edge-end real-time inference system for the operational risk of distribution boxes in an IoT environment according to claim 4, characterized in that, The evolutionary model undergoes state recursion, anomaly screening, gradient calculation and smoothing, deflection curvature extraction, and spatial mapping reconstruction to construct a preliminary risk state trajectory spectrum, including: The evolution model is recursively calculated, and the state data generated by the evolution model at each recursion step are arranged into a sequence in chronological order to obtain the state evolution sequence. The state parameters at each time step in the state evolution sequence are compared with the preset steady-state threshold range to filter out the state nodes with out-of-range values ​​and obtain the abnormal feature sequence. Extract the state node pairs of consecutive adjacent time steps in the anomaly feature sequence, calculate the set of differences between the node value of the next time step and the node value of the previous time step, and obtain the anomaly feature propagation gradient; extract the gradient vectors of consecutive adjacent time steps in the anomaly feature propagation gradient to obtain the gradient vector sequence. Calculate the spatial angle between adjacent vectors in the gradient vector sequence to obtain the gradient direction change sequence; calculate the numerical difference of consecutive angles based on the gradient direction change sequence, extract the trajectory turning rate, and obtain the deflection curvature; extract historical gradient data within a preset historical time window based on the gradient propagation of abnormal features to obtain the historical gradient sequence. The historical gradient sequence is subjected to end-value weighted accumulation processing to obtain the gradient inertia compensation amount; the current value of the gradient propagation based on the anomaly features is superimposed with the gradient inertia compensation amount to perform momentum fusion processing to obtain the smooth propagation gradient; Numerical cumulative mapping along the time axis is performed on the smooth propagation gradient to obtain the gradient magnitude path; spatial coordinate alignment is performed based on the gradient magnitude path by loading deflection curvature to obtain trajectory spatial mapping data; the trajectory spatial mapping data is serialized and reconstructed to construct the preliminary risk state trajectory spectrum.

6. The edge-end real-time inference system for the operational risk of distribution boxes in an Internet of Things environment according to claim 5, characterized in that, The reference field is partitioned to obtain a mesh set. The mesh density gradient within the mesh set is extracted to obtain compensation parameters. The preliminary risk state trajectory spectrum is calibrated using the compensation parameters to obtain the target risk state trajectory spectrum, including: Based on the reference field, perform equal-spacing partitioning along the spatial orthogonal axis to cut the effective range into multiple discrete three-dimensional mesh units, thus obtaining a mesh set; Extract the coordinates of trajectory state nodes arranged in time sequence from the preliminary risk state trajectory spectrum, and map and project the coordinates of trajectory state nodes one by one into the grid set to obtain the trajectory projection distribution set; traverse the trajectory projection distribution set to count the number of trajectory state node coordinates falling in each three-dimensional grid cell to obtain the trajectory state node number sequence. The local proportion of the coordinates of the trajectory state nodes in each 3D grid cell is calculated based on the sequence of the number of trajectory state nodes. The local proportion values ​​are arranged in the order of the grid space to obtain the grid density distribution sequence. The difference extraction process is performed on the local proportion values ​​of adjacent 3D grid cells in the grid density distribution sequence to obtain the grid density gradient. Numerical scaling is performed based on the grid density gradient, and the scaled values ​​are mapped to a preset correction range to obtain compensation parameters. Spatial position deviation correction calculations are performed on the trajectory state node coordinates based on the compensation parameters to obtain a calibrated trajectory coordinate sequence. Node continuum recombination processing is performed based on the calibrated trajectory coordinate sequence to construct the target risk state trajectory spectrum.

7. The edge-end real-time inference system for the operational risk of distribution boxes in an Internet of Things environment according to claim 6, characterized in that, Feature extraction is performed on the target risk state trajectory spectrum to construct a steady-state boundary set. The spatial feature deviation and cumulative anomaly confidence value of the risk trajectory relative to the steady-state boundary set are then calculated, including: Extract trajectory coordinate data arranged continuously in time series from the trajectory spectrum of the target risk state to obtain the trajectory feature sequence; perform numerical distribution statistical processing on the trajectory feature sequence to extract the statistical envelope parameters of the normal operation range to obtain the steady-state envelope parameter set; Based on the steady-state envelope parameter set, extract the upper and lower thresholds, and perform interval construction processing on the upper and lower thresholds to obtain the steady-state boundary set; perform position mapping and comparison processing on the trajectory coordinate data and the steady-state boundary set, and calculate the shortest vertical distance from each trajectory coordinate data to the corresponding steady-state boundary to obtain the normal distance sequence. The normal distance sequence is subjected to boundary violation detection and filtering, and the distance value segments located outside the steady-state boundary set are extracted to obtain the boundary violation distance sequence; the time span corresponding to each distance value segment is extracted from the boundary violation distance sequence to obtain the deviation duration period; The comprehensive deviation magnitude of each trajectory coordinate data relative to the steady-state boundary set is calculated based on the out-of-bounds distance sequence to obtain the spatial feature deviation degree. The spatial feature deviation degree is continuously accumulated along the time evolution direction, and confidence weight mapping is performed in combination with the deviation duration period to obtain the cumulative anomaly confidence value.

8. The edge-end real-time inference system for the operational risk of distribution boxes in an Internet of Things environment according to claim 7, characterized in that, When the weighted fusion index of spatial feature deviation and cumulative anomaly confidence value exceeds the steady-state boundary set, the edge-side real-time state transition judgment logic is executed to obtain the distribution box operation risk level label, including: Spatial feature deviation and cumulative anomaly confidence value are extracted. Weighted summation is performed based on preset deviation weight coefficient and confidence weight coefficient to obtain a weighted fusion index. The weighted fusion index is compared with the upper limit threshold of the steady-state boundary set to calculate the difference between the weighted fusion index and the upper limit threshold, thus obtaining the out-of-bounds deviation difference. Based on the deviation difference from the boundary, a level mapping process is performed according to the preset risk classification threshold range to convert the deviation difference into the corresponding risk quantification level, thus obtaining the initial risk judgment result. The initial risk judgment result is then subjected to a time sequence continuity verification process with the deviation duration input edge side real-time state transition judgment logic to obtain a steady-state risk transition flag. Based on the steady-state risk transition flag, the corresponding risk level code is extracted, and a label generation and encapsulation process is performed to obtain the distribution box operation risk level label.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the system as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the system as described in any one of claims 1 to 8.

Images