Distribution box intelligent early warning method and device based on big data

Abstract

The invention discloses a distribution box intelligent early warning method and device based on big data, and relates to the technical field of intelligent early warning. The method comprises the steps of synchronously obtaining a thermodynamic diagram and a current pulse signal of a distribution box, calculating the proportion of an abnormal temperature region of the thermodynamic diagram to generate a first time sequence group, calculating the change rate of the current pulse signal to obtain a second time sequence group, extracting and accurately capturing fault signs of different dimensions according to differentiated features, and arranging the fault signs according to a time window to obtain the time sequence group. Time sequence association is constructed; obtaining a relation time sequence group through preprocessing, substituting the relation time sequence group into the bidirectional enhanced LSTM model, deeply mining a time sequence rule, and outputting a predicted value set; and the predicted value is compared with the threshold value for judgment and early warning to realize timely fault identification and response, so that multi-modal data is fused and the weight matrix of the bidirectional LSTM model is changed along with time to process the data, and the early warning accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent early warning technology, specifically relating to an intelligent early warning method and device for power distribution boxes based on big data. Background Technology

[0002] Big data-based intelligent early warning methods for distribution boxes play a crucial role in smart grids and industrial power distribution systems due to their strong real-time performance, high prediction accuracy, and wide adaptability. With the continuous expansion of its application scenarios, reliable early warning of distribution box operating status has become a core requirement for ensuring the continuity of power supply and the safety of equipment operation. During operation, faults such as line overload and component aging not only reduce power distribution efficiency but may also trigger power accidents. Existing distribution box operation early warning technologies have significant shortcomings in terms of timeliness and reliability. Traditional single-parameter detection modes cannot consider the interconnected fluctuations and coupled effects of multiple parameters under complex operating conditions.

[0003] Patent CN108964283A discloses a distribution box early warning system, relating to the technical field of electrical safety. The method includes: a data acquisition device, a GPRS wireless sensor network device, and a terminal device, with the GPRS wireless sensor network device connected to both the data acquisition device and the terminal device. The data acquisition device collects target data from the distribution box, including at least one of the following: temperature data and current data. The GPRS wireless sensor network device determines a transmission scheme for the target data based on the number and location of the data acquisition devices, and sends the target data to the terminal device based on the transmission scheme. The terminal device acquires the target data and determines whether the target data exceeds a preset threshold. If the target data exceeds the preset threshold, an early warning message is generated and processed. This invention solves the technical problem of low transmission efficiency in existing distribution box early warning systems. However, in practical applications, due to insufficient consideration of multimodal data fusion and lack of adaptation to changes in operating conditions, it is difficult to further improve the accuracy of early warnings in complex operating scenarios. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of low accuracy in early warning, and to propose an intelligent early warning method and device for power distribution boxes based on big data.

[0005] In a first aspect of this invention, a smart early warning method for distribution boxes based on big data is proposed, the method comprising: Synchronously acquire the heat map and current pulse signal of the distribution box in the time series; The proportion of abnormal temperature areas in the heat map is calculated to assess the probability of a fault occurring, thus obtaining a first time series group. The probability of a fault occurring is measured by calculating the rate of change of the current pulse signal, resulting in a second time series group. Arrange the first time series group and the second time series group according to the time window to obtain a time series group; The time series group is preprocessed to obtain the relational time series group; Substituting the aforementioned time series data into a bidirectional enhanced LSTM model yields a set of predicted values. The predicted values ​​in the predicted value set are compared with the threshold to determine whether an alert should be issued.

[0006] Optionally, the heat map can be divided into grids to obtain each first grid cell region; Traverse all first grid cell regions, define the first grid cell regions that meet the first preset conditions as the core region map, and define the first grid cell regions that do not meet the first preset conditions as the non-core region map; The non-core region map is divided into grids to obtain the second grid unit region; Traverse all second grid cell regions and set the second grid cell regions that meet the first preset condition as the focus region map.

[0007] Optionally, the difference in frequency of the current pulse signal between two adjacent moments can be calculated; The ratio of the difference to the time interval between two adjacent moments is taken as the rate of change; The rates of change obtained at each time point are arranged in chronological order to obtain the second time series group.

[0008] Optionally, for each time window, a fusion vector is obtained by fusing the input value of the current time window with the preceding time series through vector concatenation. The weight update increment is obtained by performing a nonlinear transformation on the fusion vector using a parameter update function; The time-varying weight matrix is ​​obtained by adding the weight update increment of the previous time window; The bidirectional enhanced LSTM model is obtained by replacing the weight matrix of the bidirectional LSTM model with the time-varying weight matrix. The forward time vector and the reverse time vector are obtained by traversing the time series from the beginning in a forward direction and from the end in a reverse direction, respectively. Substituting the forward time vector and the reverse time vector into the bidirectional enhanced LSTM model yields the prediction set.

[0009] In a second aspect of this invention, a smart early warning device for distribution boxes based on big data is proposed, comprising: The synchronous acquisition module is used to synchronously acquire the heat map and current pulse signal of the distribution box over a time series. The first time series group generation module is used to calculate the proportion of abnormal temperature areas in the heat map to measure the probability of a fault occurring and obtain the first time series group. The second time series group generation module is used to calculate the rate of change of the current pulse signal to measure the probability of a fault occurring and obtain a second time series group. A time series group generation module is used to arrange the first time series group and the second time series group according to a time window to obtain a time series group; A relational time series group generation module is used to preprocess the time series group to obtain the relational time series group; The prediction value set generation module is used to substitute the relationship time series group into the bidirectional enhanced LSTM model to obtain the prediction value set; The early warning module is used to compare the predicted values ​​in the predicted value set with the threshold to determine whether to issue an early warning.

[0010] Optionally, the first time series group generation module includes: The attention area map generation module is used to input the heat map into the importance assessment model to obtain the attention area map; The region area module is used to calculate the sum of the areas of all regions of interest to obtain the region area; A temperature feature value generation module is used to take the ratio of the area of ​​the region to the area of ​​the heat map as a temperature feature value; The first time series group module is used to arrange the temperature feature values ​​obtained at each time point in chronological order to obtain the first time series group.

[0011] Optionally, the region of interest map generation module includes: The first grid unit region generation module is used to perform grid-based segmentation processing on the heat map to obtain each first grid unit region; The judgment module is used to traverse all first grid cell regions, define the first grid cell regions that meet the first preset conditions as the core region map, and define the first grid cell regions that do not meet the first preset conditions as the non-core region map. The second grid cell region generation module is used to perform grid segmentation on the non-core region map to obtain the second grid cell region. The focus area map module is used to traverse all second grid cell regions and set the second grid cell regions that meet the first preset conditions as the focus area map.

[0012] Optionally, the second time series group generation module includes: The difference calculation module is used to calculate the difference in frequency of the current pulse signal between two adjacent moments; The transformation rate calculation module is used to take the ratio of the difference to the time interval between two adjacent moments as the transformation rate; The second time series group module is used to arrange the rates of change obtained at each time point in chronological order to obtain the second time series group.

[0013] Optionally, the predicted value set generation module includes: The fusion vector generation module is used to perform a fusion vector operation on the input value of the current time window and the preceding time series by concatenating vectors for each time window. The weight update increment generation module is used to perform a nonlinear transformation on the fusion vector through a parameter update function to obtain the weight update increment; The time-varying weight matrix generation module is used to add the weight matrix of the previous time window to the weight update increment to obtain the time-varying weight matrix; A bidirectional augmented LSTM model generation module is used to replace the weight matrix of the bidirectional LSTM model with the time-varying weight matrix to obtain a bidirectional augmented LSTM model; The vector generation module is used to traverse the relational time series from the beginning in a forward direction and from the end in a reverse direction to obtain a forward time vector and a reverse time vector, respectively. The prediction set module is used to substitute the forward time vector and the backward time vector into the bidirectional enhanced LSTM model to obtain the prediction set.

[0014] The beneficial effects of this invention are as follows: The heat map and current pulse signal of the distribution box are acquired synchronously over a time series; the proportion of abnormal temperature areas in the heat map is calculated to measure the probability of a fault, resulting in a first time series group; the rate of change of the current pulse signal is calculated to measure the probability of a fault, resulting in a second time series group; the first and second time series groups are arranged according to time windows to obtain a time series group; the time series groups are preprocessed to obtain a relational time series group; the relational time series group is substituted into a bidirectional enhanced LSTM model to obtain a set of predicted values; the predicted values ​​in the predicted value group are compared with a threshold to determine whether an early warning is issued. This invention proposes a big data-based intelligent early warning method for distribution boxes. By simultaneously acquiring heat maps and current pulse signals from the distribution box, a first time series group is generated by calculating the proportion of abnormal temperature areas in the heat map, and a second time series group is generated by calculating the rate of change of the current pulse signal. Differential feature extraction is used to accurately capture fault signs from different dimensions. The time series groups are arranged according to time windows to construct temporal correlations. The resulting relational time series groups are preprocessed and fed into a bidirectional enhanced LSTM model to deeply mine temporal patterns and output a set of predicted values. The predicted values ​​are compared with thresholds to determine early warnings, enabling timely fault identification and response. This method integrates multimodal data and processes the data by changing the weight matrix of the bidirectional LSTM model over time, improving the accuracy of early warnings. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 A flowchart illustrating a smart early warning method for distribution boxes based on big data, provided as an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a smart early warning device for a power distribution box based on big data, provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a method for intelligent early warning of power distribution boxes based on big data. See also... Figure 1 , Figure 1 A flowchart illustrating a big data-based intelligent early warning method for distribution boxes, provided as an embodiment of the present invention. The method includes the following steps: S101, synchronously acquire the heat map and current pulse signal of the distribution box in the time series; S102, the proportion of abnormal temperature areas is calculated from the heat map to measure the probability of a fault occurrence and obtain the first time series group; S103, calculate the rate of change of the current pulse signal to measure the probability of a fault occurring and obtain the second time series group; S104, Arrange the first time series group and the second time series group according to the time window to obtain the time series group; S105, preprocess the time series group to obtain the relational time series group; S106, Substitute the relational time series group into the bidirectional enhanced LSTM model to obtain the predicted value set; S107, compare the predicted values ​​in the predicted value set with the threshold to determine whether to issue a warning.

[0020] This invention provides a big data-based intelligent early warning method for distribution boxes. It simultaneously acquires heat maps and current pulse signals from the distribution box, calculates the proportion of abnormal temperature areas in the heat map to generate a first time series group, and calculates the rate of change of the current pulse signal to obtain a second time series group. Differential feature extraction accurately captures fault signs in different dimensions. The time series groups are arranged according to time windows to construct temporal correlations. Preprocessing yields relational time series groups, which are then substituted into a bidirectional enhanced LSTM model to deeply mine temporal patterns and output a set of predicted values. The predicted values ​​are compared with thresholds to determine early warnings, enabling timely fault identification and response. This method integrates multimodal data and processes the data by changing the weight matrix of the bidirectional LSTM model over time, improving the accuracy of early warnings.

[0021] In one implementation, a high-precision infrared thermal imager is used to perform non-contact temperature scanning of key areas inside the distribution box, and the surface temperature distribution data of each component is collected in real time to generate a heat map, which intuitively presents the temperature anomalies and heat distribution trends inside the distribution box. The current pulse signal is obtained by deploying a current sensor at the main incoming circuit or the main switch of the distribution box to collect the total incoming current of the entire distribution box in real time. This can reflect the instantaneous current change characteristics of the entire distribution box and cover the comprehensive current pulse status of all branches and equipment carried by the distribution box.

[0022] In one implementation, the preprocessing operation on the time series group includes performing a normalization operation followed by a convolutional pooling operation to obtain the relational time series group; wherein the normalization operation can be linear normalization or Z-Score standardization.

[0023] In one implementation, if there are predicted values ​​in the prediction set that are greater than or equal to the warning threshold, a first-level warning is triggered; if all predicted values ​​in the prediction set are less than the warning threshold, but half or more of the predicted values ​​in the prediction set are within a preset range, a second-level warning is triggered; wherein the warning threshold and the preset range are set by technical personnel.

[0024] One implementation method simultaneously acquires two core parameters, temperature and current, covering the key physical characteristics of distribution box faults and avoiding the limitations of a single data dimension. Synchronous acquisition ensures that the two types of data are aligned in the time dimension, laying the foundation for subsequent analysis of their correlation and avoiding misjudgments of features due to time misalignment.

[0025] In one implementation, the intuitive heat map is transformed into a quantifiable indicator of the proportion of abnormal temperature areas, avoiding the subjectivity of manual interpretation and making temperature features easier to use for subsequent modeling. The proportion of areas can reflect the range of abnormal temperature changes, such as from local overheating to large-area temperature rise, indirectly reflecting the degree of fault development, and has more early warning value than a single temperature value.

[0026] In one implementation, bidirectional LSTM can simultaneously learn the dependencies between the past and present of a sequence and between the present and the future, thereby predicting the future using historical data and mining temporal features more comprehensively than unidirectional LSTM.

[0027] In one embodiment, a heatmap is substituted into an importance assessment model to obtain a map of the region of interest; The sum of the areas of all regions of interest is used to obtain the area of ​​the region. The ratio of the region area to the heat map area is used as the temperature characteristic value; The temperature characteristic values ​​obtained at each time point are arranged in chronological order to obtain the first time series group.

[0028] In one implementation, the importance assessment model can automatically identify areas in the heat map that are highly correlated with the fault, filter temperature information from non-critical areas, and reduce interference from invalid data. The areas of interest selected by the model can prioritize capturing temperature changes in high-incidence fault locations, pointing to potential fault points more accurately than full-map analysis, and enhancing the correlation between temperature characteristics and faults. The summation of the areas of interest from the scattered areas of interest is transformed into a single quantitative indicator, which intuitively reflects the total scale of the temperature anomaly areas that need to be monitored. The spatial information of multiple areas is compressed into a single area value, reducing data complexity and laying a quantitative foundation for subsequent calculation of ratios and generation of time series.

[0029] In one embodiment, the heat map is divided into grids to obtain each first grid unit region; Traverse all first grid cell regions, define the first grid cell regions that meet the first preset conditions as the core region map, and define the first grid cell regions that do not meet the first preset conditions as the non-core region map; The non-core region map is divided into grids to obtain the second grid unit region; Traverse all second grid cell regions and set the second grid cell regions that meet the first preset condition as the focus region map.

[0030] In one implementation, the heatmap is treated as a whole and divided into numerous equally sized grid cells according to preset rules. These grid cells are the first grid cell regions. Each of these first grid cell regions is checked individually. If the data within a particular first grid cell region meets a first preset condition, it is designated as a focused region map. Conversely, if the first preset condition is not met, such as due to disordered data distribution or low density, the first grid cell region is designated as a non-core region map. The non-core region map is then further gridded. According to preset rules, the non-core region map is further divided into smaller grid cells, which are the second grid cell regions. This step further refines the non-core regions for more detailed data analysis. Similar to the processing of the first grid cell regions, all second grid cell regions are traversed, and each second grid cell region is checked to see if it meets the first preset condition. Second grid cell regions that meet the condition are designated as focused region maps. Through this step, potentially overlooked regions that meet specific conditions can be identified from the non-core region maps, further refining the designation of focused region maps. The preset rules are set by technical personnel, such as dividing the heat map into 9*9 grid units.

[0031] In one implementation, the first preset condition is that the number of samples in the grid must be greater than a density threshold R, and the samples must be continuously and uniformly distributed within the grid. The density threshold R is set by technicians; the uniformity of the sample distribution is determined by calculating the average value of the samples in the grid, the deviation from the grid center, and a measure of data dispersion. The centroid of the samples in the grid cell is calculated using the mean calculation formula, and then the center position of the grid cell range is calculated. By comparing the distance between the centroid and the center point, and then dividing by the range of the grid, the deviation ratio is obtained; by dividing the data dispersion measure by the grid range, a dispersion index is obtained, which is used to measure the dispersion of data within the grid. If the deviation ratio is less than the deviation threshold and the dispersion index is less than the dispersion threshold, the grid is considered a core region; wherein, the deviation ratio and the dispersion index are set by technicians.

[0032] In one implementation, the first grid segmentation of the heatmap can quickly filter out the core focus areas, avoid indiscriminate refinement of the entire map, reduce invalid calculations, and improve overall efficiency; the second grid segmentation, which is performed only on non-core areas, can improve accuracy in key non-core areas and ensure that no valuable details are missed.

[0033] In one embodiment, the difference in frequency of the current pulse signal between two adjacent moments is calculated; The ratio of the difference to the time interval between two adjacent moments is taken as the rate of change; The rates of change obtained at each time point are arranged in chronological order to obtain the second time series group.

[0034] In one implementation, the frequency difference directly reflects the change amplitude of the current pulse at adjacent moments, which can quickly identify sudden changes or abnormal fluctuations in the signal and is more sensitive to instantaneous faults in the circuit than the frequency value at a single moment. When the circuit is operating normally, the frequency usually remains stable and the difference tends to be close to 0. However, before a fault occurs, the frequency often fluctuates abnormally and the difference will deviate significantly from 0. The difference can effectively filter redundant information in a stable state and focus on key change points.

[0035] In one embodiment, for each time window, a fusion vector is obtained by fusing the input value of the current time window with the preceding time series through vector concatenation. The weight update increment is obtained by performing a nonlinear transformation on the fusion vector using a parameter update function; The time-varying weight matrix is ​​obtained by adding the weight matrix of the previous time window to the weight update increment; The bidirectional augmented LSTM model is obtained by replacing the weight matrix of the bidirectional LSTM model with the time-varying weight matrix. The forward and reverse time vectors are obtained by traversing the time series from the beginning and end in reverse order, respectively. Substituting the forward and backward time vectors into the bidirectional augmented LSTM model yields the prediction set.

[0036] In one implementation, for each time window, the input value x of the current time window is concatenated using vector concatenation. t Time series h with preceding relation t-1 Perform a fusion operation to obtain the fusion vector z t This vector provides the basis for time-varying adjustment of the weight parameters by jointly encoding information from all past time windows of the instantaneous input information.

[0037] In one implementation, the parameter update function adopts a hierarchical mapping structure. First, a linear transformation is used to project the feature space, then nonlinear activation units are used for information enhancement. Finally, the output is a weight update increment with the same dimension as the weight matrix. This increment defines the time-varying adjustment direction and magnitude of the weight matrix W, allowing the model parameters to adapt to the sequence distribution of the current time window. The parameter update function is F(x t h t-1 )=u(W1*z t +b1), where W1 is the linear transformation matrix, b1 is the bias vector of the linear transformation, u is the nonlinear activation function (u can be ReLU), W1 and b1 are learnable parameters, first initialized using random or advanced methods, then the gradient is calculated by backpropagation, and updated iteratively using a gradient descent-type optimizer, finally obtaining parameter values ​​adapted to the task without manual intervention; the weight update increment W = aF(x t h t-1 ), where 'a' is the attention weight and is set by the technical staff.

[0038] In one implementation, the time-varying weight matrix W t =W t-1 +W, W t W represents the time-varying weight matrix for time window t. t-1 The time-varying weight matrix represents the time window t-1.

[0039] In one implementation, nonlinear transformation can uncover the hidden nonlinear relationships in the fused vector, which better reflects the complex characteristics of actual fault features than linear transformation; the parameter update function can generate differentiated weight update increments based on the feature strength of the fused vector, making the model more sensitive to the response of key anomaly windows; by accumulating the preceding weights and update increments, the weight matrix of LSTM can dynamically change with the time window, which can adapt to the feature distribution at different time stages and improve the model's adaptability to the non-stationarity of time series data.

[0040] Based on the same inventive concept, this invention also provides a smart early warning device for power distribution boxes based on big data. See also Figure 2 , Figure 2 A schematic diagram of a smart early warning device for a distribution box based on big data, provided as an embodiment of the present invention, includes: The synchronous acquisition module is used to synchronously acquire the heat map and current pulse signal of the distribution box over a time series. The first time series group generation module is used to calculate the proportion of abnormal temperature areas in the heat map to measure the probability of fault occurrence and obtain the first time series group. The second time series group generation module is used to calculate the rate of change of the current pulse signal to measure the probability of a fault occurrence and obtain the second time series group. The time series group generation module is used to arrange the first time series group and the second time series group according to the time window to obtain a time series group. The relational time series group generation module is used to preprocess time series groups to obtain relational time series groups; The prediction set generation module is used to substitute the relational time series group into the bidirectional augmented LSTM model to obtain the prediction set; The early warning module is used to compare the predicted values ​​in the predicted value set with the threshold to determine whether to issue an early warning.

[0041] This invention provides an intelligent early warning device for distribution boxes based on big data. It simultaneously acquires heat maps and current pulse signals from the distribution box, calculates the proportion of abnormal temperature areas in the heat map to generate a first time series group, and calculates the rate of change of the current pulse signals to obtain a second time series group. Differential feature extraction accurately captures fault signs in different dimensions. The time series groups are arranged according to time windows to construct temporal correlations. Preprocessing yields relational time series groups, which are then substituted into a bidirectional enhanced LSTM model to deeply mine temporal patterns and output a set of predicted values. The predicted values ​​are compared with thresholds to determine early warnings, enabling timely fault identification and response. This method integrates multimodal data and processes the data by changing the weight matrix of the bidirectional LSTM model over time, improving the accuracy of early warnings.

[0042] In one embodiment, the first time series group generation module includes: The attention area map generation module is used to input the heat map into the importance assessment model to obtain the attention area map; The region area module is used to calculate the sum of the areas of all regions of interest to obtain the region area; The temperature feature value generation module is used to take the ratio of the area of ​​the region to the area of ​​the heat map as the temperature feature value. The first time series group module is used to arrange the temperature feature values ​​obtained at each time point in chronological order to obtain the first time series group.

[0043] In one embodiment, the region map generation module includes: The first grid cell region generation module is used to perform grid-based segmentation of the heat map to obtain each first grid cell region; The judgment module is used to traverse all first grid cell regions, define the first grid cell regions that meet the first preset conditions as the core region map, and define the first grid cell regions that do not meet the first preset conditions as the non-core region map. The second grid cell region generation module is used to perform grid segmentation on the non-core region map to obtain the second grid cell region. The Focused Region Graph module is used to traverse all second grid cell regions and set the second grid cell regions that meet the first preset condition as the Focused Region Graph.

[0044] In one embodiment, the second time series group generation module includes: The difference calculation module is used to calculate the difference in frequency of current pulse signals between two adjacent moments; The rate of change calculation module is used to take the ratio of the difference to the time interval between two adjacent moments as the rate of change. The second time series group module is used to arrange the rates of change obtained at each time point in chronological order to obtain the second time series group.

[0045] In one embodiment, the predicted value set generation module includes: The fusion vector generation module is used to perform a fusion vector operation on the input value of the current time window and the preceding time series by concatenating vectors for each time window. The weight update increment generation module is used to obtain the weight update increment by performing a nonlinear transformation on the fusion vector through a parameter update function. The time-varying weight matrix generation module is used to add the weight matrix of the previous time window to the weight update increment to obtain the time-varying weight matrix; The bidirectional augmented LSTM model generation module is used to replace the weight matrix of the bidirectional LSTM model with the time-varying weight matrix to obtain the bidirectional augmented LSTM model. The vector generation module is used to obtain the forward time vector and the reverse time vector by traversing the relational time series from the starting end in the forward direction and the ending end in the reverse direction, respectively. The prediction set module is used to input the forward and backward time vectors into the bidirectional augmented LSTM model to obtain the prediction set.

[0046] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A big data-based intelligent warning method for distribution box, characterized in that, The method comprises: synchronously acquiring a thermal map and a current pulse signal of a distribution box in a time sequence; calculating an abnormal temperature area proportion of the thermal map to measure the possibility of fault occurrence to obtain a first time sequence group; calculating a change rate of the current pulse signal to measure the possibility of fault occurrence to obtain a second time sequence group; arranging the first time sequence group and the second time sequence group according to a time window to obtain a time sequence group; preprocessing the time sequence group to obtain a relationship time sequence group; substituting the relationship time sequence group into a bidirectional enhanced LSTM model to obtain a prediction value set; comparing the prediction value in the prediction value set with a threshold to determine whether to issue a warning.

2. The big data-based intelligent warning method for distribution box according to claim 1, characterized in that, The first time sequence group is obtained by analyzing and calculating the thermal map, which comprises: substituting the thermal map into an importance evaluation model to obtain an attention region map; calculating the sum of areas of all attention region maps to obtain a region area; taking the ratio of the region area to the area of the thermal map as a temperature characteristic value; arranging the temperature characteristic values obtained at each time point in time sequence to obtain the first time sequence group.

3. The big data-based intelligent warning method for distribution box according to claim 2, characterized in that, The attention region map set is obtained by substituting the thermal map into the importance evaluation model, which comprises: performing grid segmentation processing on the thermal map to obtain each first grid unit region; traversing all the first grid unit regions, and setting a first grid unit region meeting a first preset condition as an attention region map, and setting a first grid unit region not meeting the first preset condition as a non-core region map; performing grid segmentation processing on the non-core region map to obtain a second grid unit region; traversing all the second grid unit regions, and setting a second grid unit region meeting the first preset condition as the attention region map.

4. The big data-based intelligent warning method for distribution box according to claim 1, characterized in that, The second time sequence group is obtained by analyzing and calculating the current pulse signal, which comprises: calculating the difference between the frequencies of the current pulse signal at adjacent two time points; taking the ratio of the difference to the time interval between adjacent two time points as a change rate; arranging the change rates obtained at each time point in time sequence to obtain the second time sequence group.

5. The big data-based intelligent warning method for distribution box according to claim 1, characterized in that, The prediction value set is obtained by substituting the relationship time sequence group into the bidirectional enhanced LSTM model, which comprises: for each time window, performing fusion operation on the input value of the current time window and the previous relationship time sequence by vector splicing to obtain a fusion vector; performing nonlinear transformation on the fusion vector by a parameter update function to obtain a weight update increment; adding the weight matrix of the previous time window and the weight update increment to obtain a time-varying weight matrix; replacing the weight matrix of the bidirectional LSTM model with the time-varying weight matrix to obtain the bidirectional enhanced LSTM model; traversing the relationship time sequence from the start end forward and the end end backward to obtain a forward time vector and a backward time vector; substituting the forward time vector and the backward time vector into the bidirectional enhanced LSTM model to obtain a prediction set.

6. A big data-based intelligent warning device for distribution box, characterized in that, The device comprises: a synchronous acquisition module, configured to synchronously acquire a thermal map and a current pulse signal of a distribution box in a time sequence; a first time sequence group generation module, configured to calculate an abnormal temperature area proportion of the thermal map to measure the possibility of fault occurrence to obtain a first time sequence group; The second time sequence set generation module is configured to calculate a change rate of the current pulse signal to obtain a second time sequence set. The time sequence set generation module is configured to arrange the first time sequence set and the second time sequence set according to a time window to obtain a time sequence set. The relationship time sequence set generation module is configured to pre-process the time sequence set to obtain a relationship time sequence set. The prediction value set generation module is configured to input the relationship time sequence set into a bidirectional enhanced LSTM model to obtain a prediction value set. The early warning module is configured to compare a prediction value in the prediction value set with a threshold to determine whether to issue an early warning.

7. The big data based intelligent early warning device for distribution box according to claim 6, characterized in that, The first time sequence set generation module includes: The attention region map generation module is configured to input the heat map into an importance evaluation model to obtain an attention region map. The region area module is configured to calculate a sum of areas of all attention region maps to obtain a region area. The temperature characteristic value generation module is configured to take a ratio of the region area to an area of the heat map as a temperature characteristic value. The first time sequence set module is configured to arrange the temperature characteristic values obtained at different time points in a time sequence to obtain a first time sequence set.

8. The big data based intelligent early warning device for distribution box according to claim 7, characterized in that, The attention region map generation module includes: The first grid unit region generation module is configured to perform grid segmentation processing on the heat map to obtain a plurality of first grid unit regions. The judgment module is configured to traverse all the first grid unit regions, and divide a first grid unit region meeting a first preset condition into an attention region map, and divide a first grid unit region not meeting the first preset condition into a non-core region map. The second grid unit region generation module is configured to perform grid segmentation processing on the non-core region map to obtain a second grid unit region. The attention region map module is configured to traverse all the second grid unit regions, and set a second grid unit region meeting the first preset condition as the attention region map.

9. The big data based intelligent early warning device for distribution box according to claim 6, characterized in that, The second time sequence set generation module includes: The difference value calculation module is configured to calculate a difference value of frequencies of the current pulse signal at adjacent two time points. The change rate calculation module is configured to take a ratio of the difference value to a time interval between the adjacent two time points as a change rate. The second time sequence set module is configured to arrange the change rates obtained at different time points in a time sequence to obtain a second time sequence set.

10. The smart early warning device for power distribution box based on big data according to claim 6, characterized in that, The prediction value set generation module includes: The fusion vector generation module is configured to, for each time window, perform a fusion operation on an input value of a current time window and a previous relationship time sequence by vector splicing to obtain a fusion vector. The weight update increment generation module is configured to perform a nonlinear transformation on the fusion vector by a parameter update function to obtain a weight update increment. The time-varying weight matrix generation module is configured to add the weight update increment to a weight matrix of a previous time window to obtain a time-varying weight matrix. The bidirectional enhanced LSTM model generation module is configured to replace a weight matrix of a bidirectional LSTM model with the time-varying weight matrix to obtain a bidirectional enhanced LSTM model. The vector generation module is configured to traverse the relationship time sequence from a start end in a forward direction and from an end end in a reverse direction to obtain a forward time vector and a reverse time vector, respectively. A prediction set module is configured to substitute the forward time vector and the backward time vector into the bidirectional enhanced LSTM model to obtain a prediction set.

Images