A cable branch box contact temperature early warning method and system

Abstract

The application provides a cable branch box contact temperature early warning method and system, including obtaining temperature time series data of each monitoring contact of the cable branch box, taking the contact as a graph node, integrating normalized spatial distance, real-time temperature difference and distance mapping historical temperature correlation, calculating thermal difference degree as edge weight, constructing an undirected graph and generating a minimum spanning tree; extracting time-frequency and statistical features of the temperature sequence to construct a multi-dimensional state feature vector; calculating the minimum spanning tree edge weight statistics, combining the mean difference degree of the state features of the two sub-trees after cutting the edge to generate a pruning coefficient, and pruning the edge to obtain an initial cluster; constructing a hybrid distance based on the Euclidean distance of the nodes in the initial cluster and the minimum spanning tree path distance, and obtaining a contact state cluster through secondary clustering; constructing a reference historical normal state distribution model for each cluster, calculating the deviation degree of the current state centroid and the reference centroid, combining the model parameters to determine the early warning threshold, and outputting the temperature early warning if the threshold is exceeded.

Description

Technical Field

[0001] This application belongs to the field of early warning, and in particular relates to a method and system for early warning of contact temperature in cable branch boxes. Background Technology

[0002] During long-term operation, the contacts inside cable branch boxes are prone to abnormal heating due to factors such as increased contact resistance and insulation aging. If not warned in time, the continuous rise in contact temperature may lead to insulation breakdown or even fire, causing widespread power outages. Current technologies typically utilize online temperature monitoring systems to collect contact temperature data and rely primarily on fixed thresholds or temperature rise rates at single points for alarm judgment. However, the confined space inside cable branch boxes and the significant thermal radiation and convection coupling between contacts make single-point judgment methods ineffective in identifying the hidden early abnormal temperature rises caused by poor local contact, resulting in delayed warnings. In recent years, joint diagnostic methods based on graph theory, time series analysis, and clustering algorithms have been increasingly applied to electrical equipment condition monitoring. By constructing a graph network of measuring points and performing cluster analysis, the collaborative change patterns among multiple measuring points can be uncovered at the system level. However, existing methods fail to normalize and integrate spatial span, real-time temperature differences, and historical temporal correlations when constructing the measurement point association network, making it difficult for the graph model to accurately represent the comprehensive thermal differences between contacts. When clustering based on minimum spanning trees, inaccurate initial cluster divisions are easily caused; furthermore, most schemes only perform a single coarse clustering, lacking a secondary partitioning mechanism that integrates topological path distance and state feature spatial distance. Furthermore, existing state assessment and early warning mechanisms fail to establish a dynamic baseline based on the contact group clustering results and struggle to adaptively generate early warning thresholds based on historical normal state distribution model parameters. This makes it difficult to meet the practical engineering needs for accurate early warning of cable branch box contact temperatures. Summary of the Invention

[0003] To improve the accuracy and sensitivity of cable branch box contact temperature early warning and the overall operational safety of power grid equipment, according to a first aspect of the present invention, a cable branch box contact temperature early warning method is proposed, comprising: Acquire the temperature time series data of each monitoring contact in the cable branch box, take each monitoring contact as a graph node, integrate the spatial distance between two nodes after normalization, the real-time temperature difference and the correlation of historical temperature sequence after distance mapping to calculate the thermal difference as the edge weight, construct an undirected graph containing all the monitoring contacts and generate the corresponding minimum spanning tree. Extract the time-frequency features, basic temperature features, and statistical features of the temperature time series data, construct a multi-dimensional state feature vector for each monitoring contact, calculate the edge weight statistics of the minimum spanning tree, and generate a pruning coefficient by combining the mean difference of the state feature vectors between the two subtrees formed after cutting the edges in a way that is negatively correlated with the mean difference. Use the pruning coefficient to prune the edges of the minimum spanning tree to obtain the initial cluster. Based on the Euclidean distance between the state feature vectors of each node in the initial cluster and the path distance of the minimum spanning tree, a hybrid distance is constructed. The touch point state clusters are obtained by secondary partitioning through a clustering algorithm. A reference historical normal state distribution model is constructed according to the historical normal state features of each node in each touch point state cluster. The deviation between the centroid of the current state feature vector and the centroid of the reference historical state is calculated. The corresponding warning threshold is determined by combining the parameters of the reference historical normal state distribution model. When the deviation is greater than the warning threshold, the touch point temperature warning information is output.

[0004] According to a second aspect of the present invention, a cable branch box contact temperature early warning system is provided, comprising: The generation module is used to acquire the temperature time series data of each monitoring contact in the cable branch box, take each monitoring contact as a graph node, integrate the spatial distance between two nodes after normalization, the real-time temperature difference and the correlation of the historical temperature sequence after distance mapping to calculate the thermal difference as the edge weight, construct an undirected graph containing all the monitoring contacts and generate the corresponding minimum spanning tree. The processing module is used to extract the time-frequency features, basic temperature features, and statistical features of the temperature time series data, construct the multi-dimensional state feature vector of each monitoring contact, calculate the edge weight statistics of the minimum spanning tree, generate a pruning coefficient in a way that is negatively correlated with the mean difference degree of the state feature vector between the two subtrees formed after cutting the edge, and use the pruning coefficient to perform edge pruning on the minimum spanning tree to obtain the initial cluster. The early warning module is used to construct a hybrid distance based on the Euclidean distance between the state feature vectors of each node in the initial cluster and the path distance of the minimum spanning tree. It then performs a secondary partitioning using a clustering algorithm to obtain contact state clusters. Based on the historical normal state features of each node in each contact state cluster, it constructs a reference historical normal state distribution model. It calculates the deviation between the centroid of the current state feature vector and the centroid of the reference historical state, and determines the corresponding early warning threshold by combining the parameters of the reference historical normal state distribution model. When the deviation is greater than the early warning threshold, it outputs contact temperature early warning information.

[0005] This invention comprehensively monitors the spatial distance between contacts, real-time temperature difference, and historical temperature sequence correlation to calculate thermal difference and construct a minimum spanning tree, thereby characterizing the thermal correlation state between contacts within a cable branch box. Based on this, a multi-dimensional state feature vector of the temperature time series is extracted, and a pruning coefficient is generated by combining the feature mean difference between the two subtrees formed after edge pruning. Adaptive edge pruning is then applied to the minimum spanning tree to remove weakly correlated or heterogeneous connections. Furthermore, a hybrid distance is constructed by fusing state feature Euclidean distance and tree structure path distance, followed by secondary clustering to obtain contact state clusters with higher state consistency. Based on the historical normal state distribution model corresponding to each contact state cluster and the current deviation calculation results, an adaptive matching early warning threshold is applied, which improves the accuracy, sensitivity, and false alarm resistance of cable branch box contact temperature early warning, enhancing the safety and reliability of power grid equipment operation. Attached Figure Description

[0006] Figure 1 A flowchart for a cable branch box contact temperature early warning method; Figure 2 This is a schematic diagram of the time decay function curve; Figure 3 This is a schematic diagram of the state characteristic deviation monitoring curve; Figure 4 This diagram illustrates the comparison of fault identification performance between different models. Detailed Implementation

[0007] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0008] It should be understood that the terms “comprising” and “having”, and any variations thereof, in the embodiments of this specification are intended to cover but not exclude inclusion. For example, a product or device that includes a series of components is not necessarily limited to those components that are explicitly listed, but may include other components that are not explicitly listed or that are inherent to such product or device.

[0009] Example 1 In Embodiment 1 of the present invention, a method for early warning of contact temperature in cable branch boxes is proposed, such as... Figure 1 As shown, it includes: S1. Obtain the temperature time series data of each monitoring contact in the cable branch box. Take each monitoring contact as a graph node, integrate the spatial distance between two nodes after normalization, the real-time temperature difference, and the correlation of historical temperature sequences after distance mapping to calculate the thermal difference as the edge weight, construct an undirected graph containing all the monitoring contacts, and generate the corresponding minimum spanning tree.

[0010] The data acquisition module reads real-time temperature data transmitted from temperature sensors deployed inside the cable branch box and stores it as a time series. This time series data is then read from a local cache, historical database, or real-time acquisition queue to serve as the base dataset.

[0011] All monitoring points are defined as a set of nodes in a graph structure. The spatial distance between nodes is calculated using Euclidean distance in a three-dimensional coordinate system, and then normalized to a minimum and maximum value. The absolute difference between the temperature values ​​of two nodes at the same sampling time is calculated as the real-time temperature difference, and then normalized in the same way.

[0012] Calculate the Pearson correlation coefficient between the historical temperature time series of two nodes. The Pearson correlation coefficient characterizes the degree of linear correlation between the two temperature series, and its value ranges from [-1, 1]. Then, apply the Pearson correlation coefficient according to... A distance mapping is performed to obtain the correlation distance of historical temperature sequences, so that nodes with higher positive correlation have smaller distances and nodes with higher negative correlation have larger distances, representing the correlation of historical temperature sequences. Spatial distance weighting coefficients, real-time temperature difference weighting coefficients, and historical correlation weighting coefficients are set, and the weighted sum of the three is calculated as the thermodynamic difference.

[0013] The thermal difference is used as the edge weight, which is then assigned to the connecting edges between nodes. A complete undirected weighted graph containing all monitored touchpoints is constructed. Kruskal's algorithm is executed to generate a minimum spanning tree that connects all graph nodes and minimizes the sum of edge weights.

[0014] In some embodiments, the calculation of thermal dissimilarity as edge weights based on the integrated and normalized spatial distance between two nodes, the real-time temperature difference, and the correlation of historical temperature sequences after distance mapping is included, comprising: Calculate the absolute values ​​of the spatial distance and real-time temperature difference between the two nodes; The covariance and variance terms at each historical moment are weighted using a time decay function to calculate the weighted Pearson correlation coefficient of the historical temperature series of the two nodes. The decay rate of the time decay function is adjusted by the absolute value or nonnegation result of the current average temperature change rate of all the monitoring points. The thermal difference between the two nodes is obtained by weighting and combining the normalized spatial distance, the absolute value of the normalized real-time temperature difference, and the weighted Pearson correlation coefficient after distance mapping.

[0015] The coordinates of the two nodes in the three-dimensional coordinate system are obtained, and the Euclidean distance is calculated as the spatial distance D, which typically ranges from 10 to 150 cm. Simultaneously, the temperature measurements of the two nodes at the current sampling moment are obtained, and the absolute value of their difference is calculated as the real-time temperature difference. .

[0016] Using the attenuation rate Time decay function Where k is the number of historical time steps, such as extracting 60 sampling points from the past hour for weighted calculation, and the decay rate is a non-negative value. The time decay function curves under different decay rates are shown below. Figure 2 The diagram illustrates the impact of decay rate on the weighting of historical data. Decay Rate The system is adjusted based on the average temperature change rate of all monitored contacts over the most recent 10 cycles: when the average temperature change rate is greater than 0.5℃ / min, the system is in a state of change, and a set decay rate is established. The preferred range is [0.1, 0.3], which gives higher weight to recent data; if the average temperature change rate is less than 0.1℃ / min, then... The preferred range is [0.01, 0.05], and an example value can be set as follows: Using the time-weighted sequence, calculate the weighted covariance and variance, and then calculate the weighted Pearson correlation coefficient r between the two nodes, which ranges from [-1, 1].

[0017] The calculated spatial distance D and the real-time temperature difference Normalization is mapped to the [0,1] interval; simultaneously, the weighted Pearson correlation coefficient r is mapped using the distance-based formula. This is converted into a correlation distance distributed in [0,1]. This distance mapping does not take the absolute value of r, so that strongly negatively correlated nodes with opposite temperature change trends are assigned a larger difference distance on the historical correlation dimension.

[0018] The thermal difference edge weights are calculated using a linear weighted fusion method with a preset weighting coefficient combination (α, β, γ), with spatial distance weights being the preferred setting. Temperature difference weight Historical correlation weight For example, if the normalized spatial distance between two nodes is 0.3, the normalized temperature difference is 0.4, and the historical weighted correlation coefficient is 0.8 (i.e., the distance-based mapping is 0.1), then the calculated thermal dissimilarity degree corresponding to this edge is 0.29. This parameter can represent the degree of thermal correlation difference between two nodes in the spatial and heat conduction dimensions. Since all input factors are normalized to [0,1] and α+β+γ=1, the resulting thermal dissimilarity degree edge weights are limited to the range [0,1].

[0019] In some embodiments, acquiring the temperature time series data of each monitoring contact within the cable branch box includes: The communication command channel is triggered to send query commands to the temperature sensors corresponding to each monitoring contact point within a fixed interval by a preset hardware timing interrupt program. Receive the actual measured temperature data and corresponding timestamp at the current moment returned by each of the temperature sensors; The actual measured temperature data is stored in the historical feature sequence group according to the time sequence to generate temperature time series data for each monitoring contact.

[0020] A high-precision timer interrupt service routine is installed on the microcore controller board mounted at the front of the branch box. This routine controls the acquisition timer interrupt overflow period through register prescaler control, for example, setting it to a fixed interrupt interval of 60 seconds. Whenever the microprocessor triggers a hardware interrupt service response, it immediately sends a message task via the internally bound Modbus-RTU protocol through the industrial RS-485 communication bus or LoRa wireless radio frequency transmission channel.

[0021] Using a polling mechanism, hexadecimal addressing query commands are sequentially sent to the high-precision contact temperature sensors attached to the copper busbars of the cable joints and branch boxes. After the commands are sent to the external hardware, the control receiving and processing unit of the data bus starts a listening and receiving program with a timeout of approximately 200ms.

[0022] After each frame of temperature measurement return data from the sensor is read and parsed, a CRC check is performed. The corresponding actual current temperature data is then read from this payload, such as a real-time value of 45.3℃. Simultaneously, the kernel clock management module reads the current local chip's RTC time and appends a timestamp accurate to the second for combination and encapsulation.

[0023] A circular first-in-first-out (FIFO) buffer queue is used to push the data into a built-in two-dimensional array record list for rolling coverage. Once the accumulated sample sliding length reaches at least 512 consecutive fixed-point sequences, the record value from the initial time endpoint is discarded. This satisfies both the short-time window required for historical correlation weighting calculations and the 512-point feature extraction window required for subsequent wavelet packet decomposition, maintaining a traceable, structured one-dimensional digital array with rigorous equidistant time intervals.

[0024] S2, extract the time-frequency features, basic temperature features, and statistical features of the temperature time series data, construct the multi-dimensional state feature vector of each monitoring contact, calculate the edge weight statistics of the minimum spanning tree, and generate a pruning coefficient by combining the mean difference of the state feature vectors between the two subtrees formed after cutting the edges in a way that is negatively correlated with the mean difference. Use the pruning coefficient to perform edge pruning on the minimum spanning tree to obtain the initial cluster.

[0025] For the temperature time series data of each node, wavelet packet decomposition is used to extract wavelet energy entropy as a time-frequency feature, and the average temperature within the sliding window is calculated as the basic temperature feature, along with kurtosis and skewness as statistical features. The average temperature, wavelet energy entropy, kurtosis, and skewness are concatenated and combined into a four-dimensional state feature vector.

[0026] Calculate the upper quartile Q3, lower quartile Q1, and interquartile range IQR of all edge weights in the minimum spanning tree generated above, and use them as edge weight statistics.

[0027] Traverse each edge of the minimum spanning tree and simulate cutting the edge, then obtain the two connected branches obtained after the cutting as two subtrees.

[0028] For each subtree, calculate the mean vector of the multidimensional state feature vectors of all its contained nodes. Calculate the Euclidean distance between the two mean vectors as the mean difference of the state feature vectors.

[0029] The mean difference of the state feature vector is input into a preset negative correlation mapping function to generate the pruning coefficient corresponding to the edge. The larger the mean difference, the smaller the pruning coefficient, so that the pruning threshold decreases as the subtree difference increases, making it easier to prune heterogeneous state connection edges.

[0030] The pruning threshold for each edge is calculated using the formula Threshold = Q3 + pruning coefficient × IQR. Q3 is the upper quartile of all edge weights, representing a relatively high reference level among edge weights; IQR is the interquartile range, equal to Q3 - Q1, indicating the dispersion of edge weights; the pruning coefficient is generated based on the mean difference in state characteristics between the two subtrees after pruning the edge, used to adjust the strictness of edge pruning; and Threshold is the pruning threshold corresponding to that edge. If the thermal difference edge weight of an edge is greater than the pruning threshold, then that edge is actually deleted from the minimum spanning tree.

[0031] After completing all traversal and deletion operations, extract all remaining independent connected components in the current graph as the initial clusters.

[0032] In some embodiments, the step of extracting the time-frequency features, baseline temperature features, and statistical features of the temperature time series data to construct a multidimensional state feature vector for each monitoring contact includes: The temperature time series data of each monitoring point is decomposed into multiple scales using wavelet packet transform, and a wavelet energy entropy is calculated based on the energy ratio of each sub-band as the time-frequency feature. The average temperature of the time series data of the temperature of each monitoring contact is calculated as the basic temperature feature. The kurtosis and skewness of the temperature time series data for each monitoring contact are calculated as the statistical features. The average temperature, wavelet energy entropy, kurtosis, and skewness are concatenated and combined to form a multidimensional state feature vector of the monitoring contact point.

[0033] A fixed-length sliding window for historical temperature time series is selected, for example, the window length is set to include temperature data with 512 consecutive sampling points, and the sampling interval is set to 1 minute. This window length is consistent with the minimum capacity of the front-end circular buffer queue to avoid insufficient samples in the underlying buffer, which could prevent wavelet packet decomposition from being performed.

[0034] The data within the window is subjected to a three-level wavelet packet decomposition transform, uniformly dividing the temperature signal frequency band into eight discrete sub-bands. For each sub-band, the sum of the squares of all its decomposition coefficients is calculated to obtain the sub-band energy value, and the total system energy is obtained by summing the energies of all bands. Using the ratio of the energy value of each sub-band to the total energy, the wavelet energy entropy is calculated using the Shannon information entropy formula, thus representing the degree of disorder in the frequency domain energy distribution of the temperature signal.

[0035] Under stable temperature conditions, the wavelet energy entropy is typically around 0.8. However, when short-cycle temperature fluctuations occur due to poor contact, localized overheating, or load disturbances, the entropy value can rise to around 2.5. After time-frequency feature extraction, the basic temperature characteristics and statistical features in the time domain are calculated using a sliding window on the same temperature time series data. The average value of the sequence is calculated to obtain the average temperature, which characterizes the absolute heating level of the contact point. The ratio of the fourth central moment to the square of the variance of the sequence is calculated to obtain the kurtosis, which measures the long-tailed characteristics and abrupt spikes in the temperature fluctuation distribution; under normal operating conditions, this value is close to 3. Simultaneously, the ratio of the third central moment to the cube of the standard deviation is calculated to obtain the skewness, which represents the asymmetry of temperature changes; under normal conditions, this value tends to be close to 0.

[0036] The extracted dimensional features are structurally concatenated and combined to obtain a 4-dimensional multidimensional state feature vector V = [average temperature, wavelet energy entropy, kurtosis, skewness]. For example, the state feature vector of a certain monitoring point at the current moment can be output as [45.3, 1.25, 3.12, 0.05]. The multidimensional vector setting realizes a comprehensive health status representation of the monitoring point from macroscopic absolute temperature rise and statistical regularity to short-period temperature rise fluctuation characteristics.

[0037] In some embodiments, the step of using the pruning coefficient to prune edges in the minimum spanning tree to obtain an initial cluster includes: Calculate the interquartile range of all edge weights in the minimum spanning tree; For each edge, calculate the mean difference of the state feature vectors of all nodes in the two subtrees formed by the cut, and input the mean difference into a preset negative correlation mapping function to generate the pruning coefficient; Determine whether the weight of each edge is greater than the sum of the product of the upper quartile, the pruning coefficient, and the interquartile range. If so, prune the corresponding edge and break the minimum spanning tree to obtain the initial cluster.

[0038] Extract the weight values ​​of all candidate edges in the currently constructed minimum spanning tree and form them into a set. Use a sorting algorithm to extract the 75th percentile (upper quartile) Q3 and the 25th percentile (lower quartile) Q1 in this set. Calculate the interquartile range (IQR) by subtracting the two. For example, if Q3 of the spanning tree edge is 0.65 and Q1 is 0.25, then the calculated IQR is 0.40.

[0039] Perform a virtual cut on each edge in the minimum spanning tree to divide the original graph into subtree A and subtree B. Calculate the average value of the state feature vectors of all nodes contained in subtrees A and B respectively. Use the L2 norm to find the Euclidean distance between the two mean vectors as the mean difference.

[0040] The mean difference is input into a preset mapping function of negative correlation (Sigmoid form) to generate a pruning coefficient, for example, by setting... Where C is the pruning factor and Δ is the mean variability. and Here, Δ represents the lower and upper limits of the pruning coefficient, k is the positive slope parameter, and θ is the center of difference parameter. The larger Δ is, the closer C is to... The smaller Δ is, the closer C is to... For example, when the difference between the feature mean of two subtrees reaches 1.2, the mapping can smoothly output a lower pruning coefficient of 0.30, thus reducing the pruning threshold; when the difference is extremely small, the pruning coefficient can converge and remain in the high range of about 1.50, making it less likely for normal homogeneous connection edges to be mistakenly pruned.

[0041] After obtaining the parameters, the specific edge-cutting decision-making logic is executed. For each edge participating in the judgment, the pruning threshold formula is calculated. The actual thermal difference edge weight of the edge is compared with the calculated pruning threshold. If the condition is met, the edge is identified as an abnormal bridging edge connecting heterogeneous thermal communities. For example, if the mean difference of the subtree corresponding to a candidate edge is large, the pruning coefficient is mapped to 0.30; if Q3=0.655 and IQR=0.40, then the pruning threshold is 0.655+0.30×0.40=0.775, which is approximately 0.78. When the actual thermal difference edge weight of the edge is 0.85, the condition of being greater than the threshold is met, and the system prunes it from the minimum spanning tree. After traversing all edge weights for verification and executing the corresponding actions, the original connected graph will be deconstructed and cut into multiple state-isolated connected subgraphs, which serve as the initial cluster output for the basic clustering data.

[0042] S3. Based on the Euclidean distance between the state feature vectors of each node in the initial cluster and the path distance of the minimum spanning tree, a hybrid distance is constructed. The touch point state clusters are obtained by secondary partitioning through a clustering algorithm. A reference historical normal state distribution model is constructed according to the historical normal state features of each node in each touch point state cluster. The deviation between the centroid of the current state feature vector and the centroid of the reference historical state is calculated. The corresponding warning threshold is determined by combining the parameters of the reference historical normal state distribution model. When the deviation is greater than the warning threshold, the touch point temperature warning information is output.

[0043] Within each initial cluster, the characteristic Euclidean distance matrix of the multidimensional state feature vectors between nodes is calculated.

[0044] Calculate the weighted shortest path length between nodes on the original minimum spanning tree, and use it as the path distance matrix; the weighted shortest path length is the sum of the thermal difference edge weights of all necessary edges on the connected path between the two nodes.

[0045] After normalizing the feature Euclidean distance matrix and the path distance matrix respectively, they are linearly added according to a preset weight ratio to construct a hybrid distance matrix.

[0046] A hierarchical clustering algorithm or a density peak clustering algorithm based on local density and node distance can be used to perform secondary clustering of the mixed distance matrix to obtain a finer-grained cluster of contact states; in a specific embodiment, a density peak clustering algorithm based on local density and node distance is used to perform secondary partitioning.

[0047] Extract the multidimensional state feature vectors of all nodes within each contact state cluster during the fault-free operation history. Calculate the historical feature mean vector and the historical feature covariance matrix. To avoid singularities in the reference historical covariance matrix due to insufficient sample size or excessive feature correlation, a regularization term is added to the reference historical covariance matrix before inversion, resulting in... Where η is the regularization coefficient and I is the identity matrix; subsequent steps are based on... The Mahalanobis distance is calculated from the inverse matrix, and based on... The determinant value is used to determine the dispersion adjustment term in the warning threshold. Using this historical feature mean vector and the historical feature covariance matrix or its regularization matrix, a multivariate Gaussian distribution model is jointly constructed as a reference historical normal state distribution model.

[0048] Obtain the multi-dimensional state feature vectors of all nodes within the current real-time monitoring cluster, and calculate their average vector as the centroid of the state feature vector at the current moment.

[0049] Calculate the regularized reference history covariance matrix. The inverse matrix of the state is used to calculate the Mahalanobis distance between the centroid of the current state feature vector and the centroid of the reference historical state, which is then used as the deviation.

[0050] Based on the dimension p of the multidimensional state feature vector and the number of nodes within the contact state cluster And the determinant of the reference history covariance matrix after regularization Construct early warning thresholds at the distance scale The critical quantile of the chi-square distribution is first used on the square Mahalanobis distance scale, and then converted to the Mahalanobis distance scale through the square root, thus avoiding a misaligned comparison between the first-order Mahalanobis distance and the chi-square quantile.

[0051] Specifically, a log-threshold decision function can be used: ,in Here are the chi-square distribution quantiles with degrees of freedom p and confidence level q. As a base threshold on the same scale as the Mahalanobis distance, a and b are non-negative sensitivity parameters. ε is the baseline scale for the historical normal covariance determinant, and ε is a stability constant to prevent the denominator from being zero. This function gives a more sensitive threshold to stable homogeneous clusters with a larger number of nodes and smaller covariance dispersion, while giving a more lenient threshold to clusters with larger dispersion.

[0052] If the calculated Mahalanobis distance deviation is greater than the aforementioned distance scale warning threshold If this occurs, an exception response procedure will be triggered; equivalently, the squared Mahalanobis distance value can also be compared to see if it is greater than 1. The logging module is invoked to output warning logs, and the contact temperature warning information, which includes the contact position and deviation value, is pushed to the terminal interface through the message queue service.

[0053] In some embodiments, the step of constructing a hybrid distance based on the Euclidean distance between the state feature vectors of each node within the initial cluster and the path distance of the minimum spanning tree, and then performing a secondary partitioning using a clustering algorithm to obtain the contact point state clusters, includes: Within each initial cluster, the Euclidean distance between the state feature vectors of the nodes is calculated, and the path distance of the nodes on the minimum spanning tree is obtained. The normalized Euclidean distance and the normalized path distance are linearly combined with preset weights to obtain the mixed distance; Based on the mixed distance, a density peak clustering algorithm based on local density and node distance is used to perform secondary partitioning and merging of nodes within each initial cluster to obtain the contact state cluster.

[0054] Extract all 4-dimensional state feature vectors containing nodes within each initial cluster that has been cut into shape. Calculate the Euclidean distance between any two state feature vectors within the cluster to represent the difference in thermodynamic state fluctuations. At the same time, apply a breadth-first search algorithm to trace the connected paths of each node on the previously built minimum spanning tree, and accumulate the weights of all necessary edges in the path as the path distance at the topological level.

[0055] After obtaining the two-dimensional matrices, normalization is used to scale them to the [0,1] interval to eliminate dimensional deviations. The matrices are then linearly added with set weights to construct the hybrid distance. In engineering implementation, emphasis is usually placed on the consistency of features at the presentation level; therefore, a preferred Euclidean distance feature weight of 0.7 and a path distance topological weight of 0.3 are selected. If the Euclidean normalized distance between the feature vectors of two touchpoints is 0.4 and the path distance of the tree structure is 0.8, then the hybrid distance for evaluating their interrelationship is 0.52.

[0056] After establishing the hybrid distance correlation matrix, a density peak clustering algorithm based on local density and node distance is input to initiate secondary subdivision of the clusters. The cutoff distance is set to the top 2% quantile of the off-diagonal distance values ​​in the hybrid distance matrix, sorted in ascending order. The number of local neighbor points for each monitored contact is counted to obtain the local density. The minimum hybrid distance from each monitored contact to a node with higher local density is calculated as the center distance. Nodes whose local density and center distance both meet preset conditions are identified as candidate cluster centers. Nodes with lower local density and larger hybrid distances to existing candidate cluster centers are marked as isolated anomalous nodes. For candidate cluster centers, neighboring non-core contacts are absorbed to form state subclusters according to the principle of closest hybrid distance. For nodes marked as isolated anomalous, a separate contact state cluster is formed to be verified. This is subsequently confirmed by combining the deviation warning results, thus completing the secondary clustering cutting and merging operation. This reshuffles the initial clusters with minor potential temperature rise misalignments, outputting contact state clusters with higher consistency in node state attributes.

[0057] In some embodiments, the step of constructing a reference historical normal state distribution model based on the historical normal state characteristics of each node within each contact state cluster, calculating the deviation between the centroid of the current state feature vector and the centroid of the reference historical state, and determining the corresponding warning threshold in combination with the parameters of the reference historical normal state distribution model includes: Extract the state feature vector samples of each monitoring contact in each contact state cluster under historical normal working conditions and merge them into a reference historical sample set. Calculate the reference historical normal state distribution model containing the reference historical state centroid and the reference historical covariance matrix based on the reference historical sample set. Calculate the Mahalanobis distance between the centroid of the current state feature vector of the current contact state cluster and the centroid of the reference historical state, and use it as the deviation. The number of nodes within the contact state cluster and the determinant of the reference historical covariance matrix or its regularized matrix are used as inputs to the threshold decision function to calculate the warning threshold of the contact state cluster.

[0058] The system retrieves all node sets within each specific touchpoint status cluster from the background database, including all time window feature sample data that were previously confirmed to be in a healthy and normal operating period. The features are then horizontally recombined to form a large matrix of reference historical sample sets with dimensions of N rows and 4 columns, where N represents the total number of samples extracted.

[0059] The average expected value of each column of the 4-dimensional vector is obtained to obtain the centroid vector of the reference historical state representing the historical benchmark feature position of the cluster. Then, the 4×4 reference historical covariance matrix calculated from the sample offset is derived. Together, they form a benchmark evaluation model in multidimensional Gaussian space to represent the steady-state distribution.

[0060] Based on this, the centroid of the current feature presented in the latest state is obtained by calculating the mean of the real-time state features of all nodes in the cluster at the current acquisition time. After calculating the current feature centroid, the regularized reference historical covariance matrix is ​​used. Substituting the inverse matrix into the distance calculation formula, the Mahalanobis distance between the latest centroid and the reference historical centroid is obtained. The distance variable, standardized with a uniform variance scale, is output as the deviation. For example, the deviation of a normally fluctuating random discrete state generally does not exceed 1.5. However, if poor contact occurs locally, causing thermal imbalance, the deviation can surge and exceed the 6.0 limit. The state characteristic deviation monitoring curve is shown below. Figure 3 As shown, the variation characteristics of the Mahalanobis distance under normal operating conditions and abnormal heating conditions are presented.

[0061] The discrimination threshold is determined using an internally deployed logarithmic threshold decision function. This function uses the actual number of nodes in the current state cluster, the determinant of the reference historical covariance matrix, and the critical quantile of the chi-square distribution corresponding to the feature dimension to jointly control the fault tolerance range, and converts the chi-square quantile into a Mahalanobis distance-scaled threshold. Preferably, p=4 and q=0.99. (0.99) is approximately 13.277, sqrt( (0.99)) is approximately 3.64; when the number of nodes in the cluster is 8 and the determinant dispersion is extremely small, and the basic empirical constant a is 0.4 and the dispersion adjustment parameter b is 0.1, the floating alarm warning threshold limit of approximately 4.1 can be obtained.

[0062] The experimental conditions were set by using six consecutive months of monitoring data from cable branch boxes in a city's power distribution network as the test set, containing a total of 1200 temperature time series samples, including 1000 samples under normal operating conditions and 200 samples of abnormal temperature rise faults induced by poor contact, local overheating, and insulation degradation. The sampling interval was set to 1 minute. The control group's baseline model relied solely on spatial distance and real-time temperature difference for conventional K-means clustering and fixed threshold early warning. The ablation model, based on the complete scheme, removed the minimum spanning tree edge pruning based on the pruning coefficient and the secondary density peak clustering module. The experimental group's complete scheme incorporated multi-dimensional state feature vector construction, thermal difference edge weight spanning tree pruning, and a dynamic reference early warning mechanism. All models underwent unified cross-validation on the same computing server.

[0063] The specific data and results of the experiment show that the baseline model achieved a fault identification accuracy of 75.2% on the validation set, with a false positive rate of 16.4% and a false negative rate of 22.5%. The ablation model, stripped of the core clustering module, exhibited moderate performance, with an identification accuracy improved to 86.7%, a false positive rate reduced to 9.2%, and a false negative rate of 11.8%. Under the experimental conditions set in this embodiment, the complete scheme model employing all the technical features of this application achieved superior performance, with an overall fault identification accuracy of 98.95%, a false positive rate reduced to 1.1%, and a false negative rate of 0.8%, while the single computation time for feature extraction and early warning decision-making remained within 150ms. A bar chart comparing the fault identification performance of different models is shown below. Figure 4 As shown, the differences between the baseline model, the ablation model, and the complete scheme in terms of accuracy, false alarm rate, and false negative rate are illustrated. Compared to the ablation model, the complete scheme improves accuracy by 12.25 percentage points, confirming that using the pruning coefficient to isolate the initial clusters and employing mixed distance for secondary subdivision can effectively isolate the core centroids of heterogeneous groups. A multi-dimensional state feature vector combining absolute temperature and time-frequency characteristics can identify short-period temperature rise fluctuations caused by stable overheating, poor contact, localized overheating, or load disturbances. Coupled with a Mahalanobis distance reference warning threshold mechanism, this reduces the risk of false alarms and false negatives caused by traditional fixed thresholds under complex operating conditions, improving the system's health warning sensitivity.

[0064] Example 2 Embodiment 2 of the present invention proposes a cable branch box contact temperature early warning system, comprising: The generation module is used to acquire the temperature time series data of each monitoring contact in the cable branch box, take each monitoring contact as a graph node, integrate the spatial distance between two nodes after normalization, the real-time temperature difference and the correlation of the historical temperature sequence after distance mapping to calculate the thermal difference as the edge weight, construct an undirected graph containing all the monitoring contacts and generate the corresponding minimum spanning tree. The processing module is used to extract the time-frequency features, basic temperature features, and statistical features of the temperature time series data, construct the multi-dimensional state feature vector of each monitoring contact, calculate the edge weight statistics of the minimum spanning tree, generate a pruning coefficient in a way that is negatively correlated with the mean difference degree of the state feature vector between the two subtrees formed after cutting the edge, and use the pruning coefficient to perform edge pruning on the minimum spanning tree to obtain the initial cluster. The early warning module is used to construct a hybrid distance based on the Euclidean distance between the state feature vectors of each node in the initial cluster and the path distance of the minimum spanning tree. It then performs a secondary partitioning using a clustering algorithm to obtain contact state clusters. Based on the historical normal state features of each node in each contact state cluster, it constructs a reference historical normal state distribution model. It calculates the deviation between the centroid of the current state feature vector and the centroid of the reference historical state, and determines the corresponding early warning threshold by combining the parameters of the reference historical normal state distribution model. When the deviation is greater than the early warning threshold, it outputs contact temperature early warning information.

[0065] In some embodiments, the calculation of thermal dissimilarity as edge weights based on the integrated and normalized spatial distance between two nodes, the real-time temperature difference, and the correlation of historical temperature sequences after distance mapping is included, comprising: Calculate the absolute values ​​of the spatial distance and real-time temperature difference between the two nodes; The covariance and variance terms at each historical moment are weighted using a time decay function to calculate the weighted Pearson correlation coefficient of the historical temperature series of the two nodes. The decay rate of the time decay function is adjusted by the absolute value or nonnegation result of the current average temperature change rate of all the monitoring points. The thermal difference between the two nodes is obtained by weighting and combining the normalized spatial distance, the absolute value of the normalized real-time temperature difference, and the weighted Pearson correlation coefficient after distance mapping.

[0066] The extraction of time-frequency features, baseline temperature features, and statistical features from the temperature time-series data, and the construction of a multi-dimensional state feature vector for each monitoring contact point, includes: The temperature time series data of each monitoring point is decomposed into multiple scales using wavelet packet transform, and a wavelet energy entropy is calculated based on the energy ratio of each sub-band as the time-frequency feature. The average temperature of the time series data of the temperature of each monitoring contact is calculated as the basic temperature feature. The kurtosis and skewness of the temperature time series data for each monitoring contact are calculated as the statistical features. The average temperature, wavelet energy entropy, kurtosis, and skewness are concatenated and combined to form a multidimensional state feature vector of the monitoring contact point.

[0067] In some embodiments, the step of using the pruning coefficient to prune edges in the minimum spanning tree to obtain an initial cluster includes: Calculate the interquartile range of all edge weights in the minimum spanning tree; For each edge, calculate the mean difference of the state feature vectors of all nodes in the two subtrees formed by the cut, and input the mean difference into a preset negative correlation mapping function to generate the pruning coefficient. Determine whether the weight of each edge is greater than the sum of the product of the upper quartile, the pruning coefficient, and the interquartile range. If so, prune the corresponding edge and break the minimum spanning tree to obtain the initial cluster.

[0068] In some embodiments, the step of constructing a hybrid distance based on the Euclidean distance between the state feature vectors of each node within the initial cluster and the path distance of the minimum spanning tree, and then performing a secondary partitioning using a clustering algorithm to obtain the contact point state clusters, includes: Within each initial cluster, the Euclidean distance between the state feature vectors of the nodes is calculated, and the path distance of the nodes on the minimum spanning tree is obtained. The normalized Euclidean distance and the normalized path distance are linearly combined with preset weights to obtain the mixed distance; Based on the mixed distance, a density peak clustering algorithm based on local density and node distance is used to perform secondary partitioning and merging of nodes within each initial cluster to obtain the contact state cluster.

[0069] In some embodiments, the step of constructing a reference historical normal state distribution model based on the historical normal state characteristics of each node within each contact state cluster, calculating the deviation between the centroid of the current state feature vector and the centroid of the reference historical state, and determining the corresponding warning threshold in combination with the parameters of the reference historical normal state distribution model includes: Extract the state feature vector samples of each monitoring contact in each contact state cluster under historical normal working conditions and merge them into a reference historical sample set. Calculate the reference historical normal state distribution model containing the reference historical state centroid and the reference historical covariance matrix based on the reference historical sample set. Calculate the Mahalanobis distance between the centroid of the current state feature vector of the current contact state cluster and the centroid of the reference historical state, and use it as the deviation. The number of nodes within the contact state cluster and the determinant of the reference historical covariance matrix or its regularized matrix are used as inputs to the threshold decision function to calculate the warning threshold of the contact state cluster.

[0070] In some embodiments, acquiring the temperature time series data of each monitoring contact within the cable branch box includes: The communication command channel is triggered to send query commands to the temperature sensors corresponding to each monitoring contact point within a fixed interval by a preset hardware timing interrupt program. Receive the actual measured temperature data and corresponding timestamp at the current moment returned by each of the temperature sensors; The actual measured temperature data is stored in the historical feature sequence group according to the time sequence to generate temperature time series data for each monitoring contact.

[0071] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and to aid in understanding a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art, upon reading this specification, may readily identify some of the devices as separate embodiments. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. And the content of each secondary embodiment is valid even if it contains fewer than all the features of a single foregoing disclosed embodiment.

Claims

1. A method for early warning of contact temperature in a cable branch box, characterized in that, Includes the following steps: Acquire the temperature time series data of each monitoring contact in the cable branch box, take each monitoring contact as a graph node, integrate the spatial distance between two nodes after normalization, the real-time temperature difference and the correlation of historical temperature sequence after distance mapping to calculate the thermal difference as the edge weight, construct an undirected graph containing all the monitoring contacts and generate the corresponding minimum spanning tree. Extract the time-frequency features, basic temperature features, and statistical features of the temperature time series data, construct a multi-dimensional state feature vector for each monitoring contact, calculate the edge weight statistics of the minimum spanning tree, and generate a pruning coefficient by combining the mean difference of the state feature vectors between the two subtrees formed after cutting the edges in a way that is negatively correlated with the mean difference. Use the pruning coefficient to prune the edges of the minimum spanning tree to obtain the initial cluster. Based on the Euclidean distance between the state feature vectors of each node in the initial cluster and the path distance of the minimum spanning tree, a hybrid distance is constructed. The touch point state clusters are obtained by secondary partitioning through a clustering algorithm. A reference historical normal state distribution model is constructed according to the historical normal state features of each node in each touch point state cluster. The deviation between the centroid of the current state feature vector and the centroid of the reference historical state is calculated. The corresponding warning threshold is determined by combining the parameters of the reference historical normal state distribution model. When the deviation is greater than the warning threshold, the touch point temperature warning information is output.

2. The method according to claim 1, characterized in that, The spatial distance between the two nodes after integration and normalization, the real-time temperature difference, and the correlation of the historical temperature sequence after distance mapping are used to calculate the thermodynamic difference as the edge weight, including: Calculate the absolute values ​​of the spatial distance and real-time temperature difference between the two nodes; The covariance and variance terms at each historical moment are weighted using a time decay function to calculate the weighted Pearson correlation coefficient of the historical temperature series of the two nodes. The decay rate of the time decay function is adjusted by the absolute value or nonnegation result of the current average temperature change rate of all the monitoring points. The thermal difference between the two nodes is obtained by weighting and combining the normalized spatial distance, the absolute value of the normalized real-time temperature difference, and the weighted Pearson correlation coefficient after distance mapping.

3. The method according to claim 1 or 2, characterized in that, The extraction of time-frequency features, baseline temperature features, and statistical features from the temperature time-series data, and the construction of a multi-dimensional state feature vector for each monitoring contact point, includes: The temperature time series data of each monitoring point is decomposed into multiple scales using wavelet packet transform, and a wavelet energy entropy is calculated based on the energy ratio of each sub-band as the time-frequency feature. The average temperature of the time series data of the temperature of each monitoring contact is calculated as the basic temperature feature. The kurtosis and skewness of the temperature time series data for each monitoring contact are calculated as the statistical features. The average temperature, wavelet energy entropy, kurtosis, and skewness are concatenated and combined to form a multidimensional state feature vector of the monitoring contact point.

4. The method according to claim 1, characterized in that, The process of pruning the minimum spanning tree using the pruning coefficient to obtain the initial cluster includes: Calculate the interquartile range of all edge weights in the minimum spanning tree; For each edge, calculate the mean difference of the state feature vectors of all nodes in the two subtrees formed by the cut, and input the mean difference into a preset negative correlation mapping function to generate the pruning coefficient. Determine whether the weight of each edge is greater than the sum of the product of the upper quartile, the pruning coefficient, and the interquartile range. If so, prune the corresponding edge and break the minimum spanning tree to obtain the initial cluster.

5. The method according to claim 1 or 4, characterized in that, The process involves constructing a hybrid distance based on the Euclidean distance between the state feature vectors of each node within the initial cluster and the path distance of the minimum spanning tree. This hybrid distance is then used to perform a secondary partitioning using a clustering algorithm to obtain contact point state clusters, including: Within each initial cluster, the Euclidean distance between the state feature vectors of the nodes is calculated, and the path distance of the nodes on the minimum spanning tree is obtained. The normalized Euclidean distance and the normalized path distance are linearly combined with preset weights to obtain the mixed distance; Based on the mixed distance, a density peak clustering algorithm based on local density and node distance is used to perform secondary partitioning and merging of nodes within each initial cluster to obtain the contact state cluster.

6. The method according to claim 1, characterized in that, The step of constructing a reference historical normal state distribution model based on the historical normal state characteristics of each node within each contact point state cluster, calculating the deviation between the centroid of the current state feature vector and the centroid of the reference historical state, and determining the corresponding warning threshold based on the parameters of the reference historical normal state distribution model includes: Extract the state feature vector samples of each monitoring contact in each contact state cluster under historical normal working conditions and merge them into a reference historical sample set. Calculate the reference historical normal state distribution model containing the reference historical state centroid and the reference historical covariance matrix based on the reference historical sample set. Calculate the Mahalanobis distance between the centroid of the current state feature vector of the current contact state cluster and the centroid of the reference historical state, and use it as the deviation. The number of nodes within the contact state cluster and the determinant of the reference historical covariance matrix or its regularized matrix are used as inputs to the threshold decision function to calculate the warning threshold of the contact state cluster.

7. The method according to claim 1, characterized in that, The acquisition of temperature time series data for each monitoring contact point within the cable branch box includes: The communication command channel is triggered to send query commands to the temperature sensors corresponding to each monitoring contact point within a fixed interval by a preset hardware timing interrupt program. Receive the actual measured temperature data and corresponding timestamp at the current moment returned by each of the temperature sensors; The actual measured temperature data is stored in the historical feature sequence group according to the time sequence to generate temperature time series data for each monitoring contact.

8. A cable branch box contact temperature early warning system, characterized in that, Includes the following modules: The generation module is used to acquire the temperature time series data of each monitoring contact in the cable branch box, take each monitoring contact as a graph node, integrate the spatial distance between two nodes after normalization, the real-time temperature difference and the correlation of the historical temperature sequence after distance mapping to calculate the thermal difference as the edge weight, construct an undirected graph containing all the monitoring contacts and generate the corresponding minimum spanning tree. The processing module is used to extract the time-frequency features, basic temperature features, and statistical features of the temperature time series data, construct the multi-dimensional state feature vector of each monitoring contact, calculate the edge weight statistics of the minimum spanning tree, generate a pruning coefficient in a way that is negatively correlated with the mean difference degree of the state feature vector between the two subtrees formed after cutting the edge, and use the pruning coefficient to perform edge pruning on the minimum spanning tree to obtain the initial cluster. The early warning module is used to construct a hybrid distance based on the Euclidean distance between the state feature vectors of each node in the initial cluster and the path distance of the minimum spanning tree. It then performs a secondary partitioning using a clustering algorithm to obtain contact state clusters. Based on the historical normal state features of each node in each contact state cluster, it constructs a reference historical normal state distribution model. It calculates the deviation between the centroid of the current state feature vector and the centroid of the reference historical state, and determines the corresponding early warning threshold by combining the parameters of the reference historical normal state distribution model. When the deviation is greater than the early warning threshold, it outputs contact temperature early warning information.

9. The system according to claim 8, characterized in that, The spatial distance between the two nodes after integration and normalization, the real-time temperature difference, and the correlation of the historical temperature sequence after distance mapping are used to calculate the thermodynamic difference as the edge weight, including: Calculate the absolute values ​​of the spatial distance and real-time temperature difference between the two nodes; The covariance and variance terms at each historical moment are weighted using a time decay function to calculate the weighted Pearson correlation coefficient of the historical temperature series of the two nodes. The decay rate of the time decay function is adjusted by the absolute value or nonnegation result of the current average temperature change rate of all the monitoring points. The thermal difference between the two nodes is obtained by weighting and combining the normalized spatial distance, the absolute value of the normalized real-time temperature difference, and the weighted Pearson correlation coefficient after distance mapping.

10. The system according to claim 8, characterized in that, The extraction of time-frequency and statistical features from the temperature time-series data, and the construction of a multi-dimensional state feature vector for each monitoring contact point, includes: The temperature time series data of each monitoring point is decomposed into multiple scales using wavelet packet transform, and a wavelet energy entropy is calculated based on the energy ratio of each sub-band as the time-frequency feature. The kurtosis and skewness of the temperature time series data for each monitoring contact are calculated as the statistical features. The wavelet energy entropy, kurtosis, and skewness are concatenated and combined to form the multidimensional state feature vector of the monitoring contact point.

Images