Repeated call ticket identification method based on customer appeal correlation analysis

Abstract

This invention discloses a method for identifying repeat call work orders based on customer complaint correlation analysis, including: acquiring customer complaint work order records, power grid topology impedance weighted diagrams, and protection device action logs; fusing text semantics and topology feature mapping to fault mode probability distributions; calculating spatiotemporal correlation degree using fault type as a control parameter and fusing weighted results to obtain joint correlation degree; screening candidate work order pairs and extracting ordered sequences of protection devices along the path; calculating physical reachability through protection device action logs and eliminating invalid work order pairs isolated by protected actions; clustering effective correlated work order pairs into correlated work order groups, fusing multidimensional signals of fault phenomenon level, topology location level, and temporal sequence to infer causal direction, and outputting the root cause work order. This invention combines probabilistic soft scoring with physical hard blocking verification, improving the robustness of work order correlation and the accuracy of root cause tracing under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of power system distribution network operation and maintenance management technology, and in particular to a method for identifying duplicate call orders based on customer demand correlation analysis. Background Technology

[0002] As a power infrastructure directly connecting end users, the operating status of the distribution network directly affects the reliability of power supply. When an anomaly or fault occurs in the power grid, it triggers a large number of customer complaint work orders stemming from the same root cause. Accurately identifying recurring customer complaints with inherent connections from massive amounts of discrete work order data is of significant engineering and technical importance for quickly locating faulty sections of the power grid, characterizing the scope of the fault's impact, guiding emergency repair scheduling, and improving the overall perception resolution of abnormal states in the distribution network.

[0003] Currently, the identification of associations in customer complaint work orders mainly relies on time-window-based clustering statistics or text feature similarity matching methods. These technical solutions typically set fixed time thresholds and two-dimensional geographic coordinate ranges. When a new work order and a historical work order fall within the set time-space window, or when the work order description text contains high-frequency overlapping words, it is determined to be a duplicate work order. In practical applications, due to the complexity of the distribution network topology and the differences in physical characteristics, the transmission patterns of different types of anomalies in the network structure vary significantly. Simple two-dimensional geographic distance is insufficient to accurately reflect the electrical coupling attenuation characteristics along the power grid path, and static association rules cannot adapt to the dynamic blocking effects caused by local equipment actions. This can easily lead to over-association or under-identification when handling complex abnormal operating conditions.

[0004] Currently, when processing multi-source heterogeneous operation and maintenance data of distribution networks, it is difficult to effectively balance the evolution of fault phenomena with the dynamic constraints of the power grid's physical state. This results in insufficient robustness and adaptability of existing methods in work order correlation under complex operating conditions. Therefore, it is necessary to study an identification method that can deeply fit the physical operating characteristics of distribution networks and improve the cross-dimensional data fusion processing capabilities to solve the problem of accurate source tracing under massive abnormal signals. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and defects of existing technologies and provide a method for identifying repeat call work orders based on customer demand correlation analysis. This invention combines probabilistic soft scoring with physical hard blocking verification, improving the robustness of work order correlation and the accuracy of root cause tracing under complex operating conditions. It also solves the problem of excessively high false positives caused by traditional clustering methods relying on single text similarity and ignoring the physical topological barriers of the power grid.

[0006] A method for identifying repeat call tickets based on customer request correlation analysis includes:

[0007] Obtain customer complaint work order records, power grid topology impedance weighted diagrams, and protection device action logs;

[0008] Based on the semantic vector extracted from customer complaint work order records and the topological location features extracted from the power grid topology impedance weighted map, the fault mode probability distribution corresponding to the customer complaint work order records is mapped to obtain the probability distribution of the fault mode.

[0009] Based on the probability distribution of fault modes, the spatiotemporal similarity of work order pairs determined by fault type is fused and weighted to obtain the joint correlation degree between work order pairs.

[0010] Candidate associated work order pairs are selected based on joint correlation degree, and the ordered sequence of protection devices corresponding to the candidate associated work order pairs in the power grid topology impedance weighted graph is extracted.

[0011] The physical reachability of the ordered sequence of protection devices is calculated using the action logs of the protection devices to determine the valid associated work order pairs;

[0012] Effectively associated work order pairs are clustered into associated work order groups. Multi-dimensional judgment signals are integrated within the associated work order groups to infer the causal propagation direction and determine the root work order.

[0013] Preferably, based on the semantic vector extracted from customer complaint work order records and the topological location features extracted from the power grid topology impedance weighted map, the fault mode probability distribution corresponding to the customer complaint work order records is mapped, including:

[0014] Extract semantic vectors from customer complaint work order records;

[0015] Extract the topological location features of the access nodes from the customer complaint work order records from the power grid topology impedance weighted map;

[0016] By concatenating semantic vectors with topological location features and mapping them to a predefined set of fault types, the probability distribution of fault modes corresponding to customer complaint work order records is obtained.

[0017] Preferably, the topological location features corresponding to the access nodes in the customer complaint work order records are extracted from the power grid topology impedance-weighted map, including:

[0018] Based on the impedance weighted graph of the power grid topology, the impedance electrical distance from the access node to the main power source point is extracted;

[0019] Extract the branch order of the distribution network branch line where the access node is located;

[0020] Count the total number of users downstream of the access node;

[0021] The impedance electrical distance, branch order, and total number of users are encoded and fused to form the topological location features.

[0022] Preferably, the joint correlation degree between the work order pairs is calculated using the following formula:

[0023] ;

[0024] in, Work orders in customer complaint work order records With work orders The degree of joint correlation between them; A predefined set of fault types; The fault type is one of the fault types in the fault type set. and These are work orders in the failure mode probability distribution. With work orders By fault type The probability of triggering; Based on fault type Work orders for calculating control parameters With work orders The spatiotemporal similarity between them.

[0025] Preferably, the spatiotemporal similarity is determined by the spatiotemporal attenuation basis function, based on the spatial attenuation factor characterizing the impedance voltage division attenuation effect under the physical constraints of the distribution network, and the time alignment factor characterizing the phenomenological evolution of faults and the perceived delay effect constructed using a Gaussian function, including:

[0026] ;

[0027] in, For the corresponding fault type The pre-configured attenuation impedance, Work orders based on impedance parameters extracted from the impedance weighted graph of the power grid topology. With work orders The positive sequence impedance magnitude between access nodes; Work orders extracted based on the time attribute in customer complaint work order records With work orders The time difference of incoming calls; For the corresponding fault type The pre-configured phenomenological delay coefficient; For the corresponding fault type Pre-configured time tolerance parameters; This is the spatial decay factor; This is the time alignment factor.

[0028] Preferably, the ordered sequence of protection devices corresponding to candidate associated work order pairs in the power grid topology impedance weighted graph is extracted, including:

[0029] In the power grid topology impedance weighted graph, based on the two access nodes corresponding to the candidate associated work order pair, the shortest path search algorithm is executed to obtain the shortest topology path connecting the two access nodes.

[0030] Extract nodes or edges with protected device attributes along the shortest topological path.

[0031] The extracted nodes or edges are arranged according to the spatial topology connection order to obtain an ordered sequence of protection devices.

[0032] Preferably, the physical reachability of the ordered sequence of protection devices is calculated using the protection device action log to determine valid associated work order pairs, including:

[0033] Based on customer complaint work order records, extract the call time of work order pairs, use the later call time of the two work order call times plus a pre-configured communication delay margin as the right endpoint, and use the pre-configured maximum propagation delay back from the right endpoint as the left endpoint to construct a log query window;

[0034] Based on the log query window, extract the set of action statuses of protection devices from the ordered sequence of protection device action logs;

[0035] Based on the action state set, the blocking status of the protection device in the ordered sequence of protection devices is determined. Combined with the blocking status, physical reachability is calculated, and work order pairs that fail to pass the physical reachability test are eliminated, while valid related work order pairs are retained.

[0036] Preferably, the blocking state of the protection devices in an ordered sequence of protection devices is determined based on the action state set, including:

[0037] If there is a successful tripping action and no successful reclosing record in the log query window, it is determined to be in an effective isolation state; if there is a successful reclosing record in the log query window or if the protection coordination logic determines that there is a failure to operate or an over-level trip, it is determined to be in an ineffective isolation state; if there is no action record in the log query window, it is determined to be in a state of no over-limit / no impact.

[0038] As a preferred method, physical reachability is calculated based on the blocking status, and work order pairs that fail to achieve physical reachability are removed, while valid associated work order pairs are retained, including:

[0039] Map the effective isolation state to the value 1, and map the ineffective isolation state and the non-limit-crossing / non-affected state to the value 0 to obtain the blocking coefficient corresponding to the protection device in the ordered sequence of protection devices;

[0040] Based on the blocking coefficient, physical reachability is calculated using a product function, expressed as follows: = ( - ), For physical accessibility, To protect the total number of protected devices in the ordered sequence of protected devices, To protect the first in the ordered sequence of equipment The blocking coefficient corresponding to each protection device; when When =0, the corresponding candidate associated work order pair is removed; When =1, the corresponding candidate associated work order pair will be retained as a valid associated work order pair.

[0041] As a preferred method, multi-dimensional judgment signals are integrated within the associated work order group to infer the causal propagation direction and determine the root work order, including:

[0042] Map the valid associated work order pairs to a graph structure, perform connected component detection, and obtain the associated work order group;

[0043] Extract the fault phenomenon level signal, topological location level signal, and time sequence signal of the work orders within the associated work order group to form a multi-dimensional judgment signal;

[0044] By integrating multi-dimensional decision signals, the comprehensive causal confidence score of work orders within the associated work order group is calculated, and the work order with the highest comprehensive causal confidence score is identified as the root work order.

[0045] This application maps the semantics of customer complaint texts to the topological features of the power grid into a probability distribution of fault modes, and uses the fault type as a control parameter. It combines the positive sequence impedance of the power grid and the time difference of power supply to calculate the spatiotemporal correlation, breaking the limitation of the traditional fixed physical distance window. It transforms the coarse geometric correlation into a probabilistic correlation that conforms to the electrical characteristics and phenomenological evolution law, reduces the correlation misjudgment rate in complex fault scenarios, and solves the problem that existing methods cannot effectively take into account the evolution law of fault phenomena.

[0046] This application extracts the ordered sequence of protection devices on the power grid topology path and verifies the physical reachability between work orders by combining the device action log. It incorporates the physical isolation effect caused by actions such as switch tripping into the calculation, effectively avoiding the merging of invalid work orders that cross electrical disconnection points, enhancing the robustness of topology association under dynamic power grid conditions, and solving the problem of existing methods ignoring the dynamic constraints of the power grid physical state.

[0047] After constructing a group of related work orders, this application integrates multi-dimensional judgment signals such as fault phenomenon level, upstream and downstream topology location, and time sequence to make comprehensive causal inference. This overcomes the blind spot in tracing caused by relying solely on time sequence judgment, and accurately converges the root cause location granularity of discrete customer complaints from fuzzy macro segments to specific single root cause work orders. This provides high-confidence decision support for on-site emergency repair scheduling and solves the problem of difficulty in accurate tracing under massive abnormal signals. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method for identifying repeat call work orders based on customer request association analysis provided in the embodiments of this application.

[0049] Figure 2 This is a flowchart illustrating the construction process of the power grid topology impedance weighted graph provided in this application embodiment.

[0050] Figure 3 This is a flowchart illustrating the probability distribution of fault modes corresponding to customer complaint work order records obtained based on semantic vectors and topological location feature mapping, provided in an embodiment of this application.

[0051] Figure 4 This is a flowchart provided in this application embodiment for extracting the topological location features corresponding to the access node in the customer complaint work order record from the power grid topology impedance weighted map.

[0052] Figure 5 This is a flowchart of extracting the ordered sequence of protection devices corresponding to candidate associated work order pairs in the power grid topology impedance weighted diagram, provided in an embodiment of this application.

[0053] Figure 6 This is a flowchart provided in an embodiment of the present application for calculating the physical reachability of an ordered sequence of protection devices using the action log of the protection device, and determining a valid associated work order pair.

[0054] Figure 7 This is a flowchart provided in this application embodiment that clusters effective associated work order pairs into associated work order groups, integrates multi-dimensional judgment signals within the associated work order groups to infer the causal propagation direction, and determines the root work order. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0056] In an exemplary embodiment of this application, the method for identifying repeat call tickets based on customer request correlation analysis, such as... Figure 1 As shown, the processing steps include the following:

[0057] S1. Obtain customer complaint work order records, power grid topology impedance weighted graph, and protection device action logs. The customer complaint work order records can be from real-time call records of the power customer service system, including complaint text, repair time, and customer access node information. The power grid topology impedance weighted graph is a pre-constructed graphical model reflecting the spatial structure of the distribution network, where nodes correspond to power grid equipment or customer access nodes, edges correspond to physical lines, and each edge is assigned an impedance parameter as a weight. Furthermore, the pre-constructed power grid topology impedance weighted graph can also include attribute labels for each protection device node to facilitate subsequent path search and state mapping. The protection device action logs are derived from the switch tripping and closing history information recorded by the distribution automation system or substation. By obtaining the above multi-source heterogeneous data, a data foundation is provided for subsequent soft scoring and hard physical isolation verification. In an optional implementation, the customer complaint work order records can be pre-standardized and cleaned to filter out invalid repair text.

[0058] S2. Based on the semantic vector extracted from customer complaint work order records and the topological location features extracted from the power grid topology impedance weighted map, the fault mode probability distribution corresponding to the customer complaint work order records is mapped to obtain the probability distribution. Specifically, for the semantic vector features, a pre-trained language model can be used to vectorize the text in the customer complaint work order records; for the topological location features, based on the position of the access node in the power grid topology impedance weighted map, numerical features such as distance and level are extracted and mapped to high-dimensional latent vectors through a multilayer perceptron. Finally, the semantic vector features and the high-dimensional latent vectors corresponding to the topological location features are concatenated bit-by-bit and input into a predefined fault classification model. This fault classification model outputs a normalized probability vector for each work order, where each element corresponds to a fault type in a predefined fault type set, and the sum of the values ​​of all elements is 1. The probability vectors of all work orders are arranged row-wise to form a probability distribution matrix. For example, the output probability vector can represent that a work order has an 80% probability of being caused by overload, a 10% probability of being caused by short circuit, and a remaining 10% probability of being caused by grounding. This probability mapping mechanism transforms apparent complaint text into a mathematical distribution reflecting underlying physical causes. By mapping apparent complaints to multidimensional probability distributions, the problem of single semantic matching failure is overcome. Even if two work order complaints are completely different, as long as they are likely to be caused by the same underlying physical fault, the mapping result can provide objective numerical basis for subsequent joint scoring.

[0059] S3. Based on the fault mode probability distribution, the spatiotemporal similarity of work order pairs determined by fault type is fused and weighted to obtain the joint correlation degree between work order pairs. Different fault types exhibit significant differences in propagation characteristics. To evaluate the dual effects of spatial attenuation and time delay, this application proposes using various fault types covered in the fault mode probability distribution as control parameters, and pre-configured parameters of spatial attenuation factor and time alignment factor to calculate the spatiotemporal similarity of multiple work order pairs. Then, based on a variant of the total probability formula, the probability product of each work order for the same fault type is used as a weight to perform a weighted summation of the corresponding spatiotemporal similarities, marginalizing the uncertainty of fault types. The result of this weighted summation is the joint correlation degree, whose numerical distribution is typically between 0 and 1. For a predetermined fault type, its spatial attenuation law can be approximated based on the voltage division effect of the radiation network impedance, while the time delay law can be calculated using a Gaussian function for alignment. Through the above fusion and weighting process, the propagation speed and influence radius of different faults can be adaptively adapted, accurately capturing hidden correlations with large spans but conforming to physical laws. The pre-configured parameters can be determined through experiments on the validation set based on the need to balance association accuracy and recall in actual application scenarios.

[0060] S4. Based on the joint correlation degree, candidate related work order pairs are selected, and the ordered sequence of protection devices corresponding to the candidate related work order pairs in the power grid topology impedance weighted diagram is extracted. In specific implementation, when selecting candidate related work order pairs, the calculated joint correlation degree is compared with a preset threshold; when the joint correlation degree is greater than or equal to the set threshold, the corresponding two work orders are considered to have high physical correlation feasibility in terms of probability and spatiotemporal attenuation dimensions, and are included in the set of candidate related work order pairs, thereby selecting candidate related work order pairs. Optionally, the preset threshold can be set to 0.5, or adjusted according to historical statistical data of actual power grid operation. When extracting the ordered sequence of protection devices corresponding to candidate associated work order pairs in the power grid topology impedance weighted graph, a shortest path search algorithm, such as Dijkstra's algorithm or A* algorithm, is used. On the power grid topology impedance weighted graph, based on the two access nodes corresponding to the candidate associated work order pairs, the shortest topological path between the corresponding access nodes is searched. Then, the path is traversed along the shortest topological path to extract all device nodes or edges with protection attributes. These extracted nodes or edges are then arranged according to the spatial topology connection order, finally forming the ordered sequence of protection devices. Extracting the ordered sequence of protection devices is crucial to accurately pinpointing the subsequent verification scope to physical barriers that may isolate fault propagation, avoiding blind global search calculations.

[0061] S5. Calculate the physical reachability of the ordered sequence of protection devices using the protection device action logs to determine valid associated work order pairs. For each protection device in the ordered sequence, its tripping or closing status can be retrieved from the protection device action log according to the set time query window. Combined with the pre-configured relay protection coordination logic, the complex action log is transformed into a binary blocking status. If the device successfully blocks the fault from spreading outward along the path, a truncation effect occurs, and the corresponding value is 0; if it does not act, reclosing is successful, or a cascading trip occurs, it is determined that the truncation is not effective, and the corresponding value is 1. Physical reachability can be calculated by constructing a product function to achieve a veto-type hard blocking verification. If there is at least one protection device on the path that produces a truncation effect, the product result of the physical reachability calculation is 0, and the association relationship of the work order pair is determined to be physically isolated and cut off, and it is removed as a phantom association; conversely, if the product result is 1, it indicates that the fault propagation path is unobstructed, and it is retained as a valid associated work order pair. By designing a multiplication verification mechanism based on physical reachability, the problem of previous clustering algorithms neglecting the dynamic isolation function of power grid protection devices is effectively overcome. Through a dual-drive filtering mechanism combining soft evaluation and hard blocking, the compliance and accuracy of work order association results at the physical level are improved.

[0062] S6. Cluster valid related work order pairs into related work order groups. Within these groups, integrate multi-dimensional decision signals to infer the causal propagation direction and determine the root work order. In practice, all remaining valid related work order pairs can be considered as edges in an undirected graph, with the work order itself as a node, constructing a work order related undirected graph. Use a depth-first search algorithm or a density-based graph clustering algorithm to extract connected regions from the work order related undirected graph; each connected region constitutes a related work order group. After forming the related work order groups, extract the fault phenomenon level signal, topological location level signal, and temporal sequence signal for each work order to form multi-dimensional decision signals. By assigning different priority weights to heterogeneous signals, comprehensively calculate the causal confidence of each work order inducing other work orders, ultimately locking down the node with the highest causal confidence value as the root work order output, and outputting the related work order group to which the root work order belongs.

[0063] In optional implementations, priority levels can be predetermined for inferring the direction of causal propagation. For example, the confidence level of the root cause level representing the abnormality of the distribution transformer can be higher than that of the conduction level representing the abnormality of the terminal voltage, and the confidence level of the upstream topology location can be higher than that of the downstream topology location. By integrating a multi-dimensional judgment signal discrimination mechanism, the problem of root cause location error caused by the difference in customer repair delay due to the reliance on a single time parameter is overcome, providing a clear dispatching guide for emergency repair of distribution network faults.

[0064] In one embodiment, before obtaining customer complaint work order records, power grid topology impedance weighted diagrams, and protection device action logs, the method further includes a step of constructing the power grid topology impedance weighted diagram, such as... Figure 2 The steps include:

[0065] S11, Obtain the original physical topology and complex impedance parameters of the distribution network. The original physical topology is specifically represented by the set of nodes and connections derived from the distribution network's geographic information system (GIS). Nodes represent distribution transformers, branch boxes, or customer access points, while connections represent actual overhead lines or cables. Complex impedance parameters refer to the electrical parameters corresponding to each line, typically including resistance and reactance values. Obtaining the actual complex impedance parameters provides an accurate physical spatial measurement basis for subsequent calculations, overcoming the shortcomings of traditional methods that rely solely on geographical straight-line distances while ignoring the actual routing of the power grid.

[0066] S12 determines the scalar edge weights based on the complex impedance parameters using a preset method. The preset method includes using either the positive-sequence impedance magnitude or the positive-sequence reactance of the complex impedance parameters as the scalar weights. Since complex impedance is difficult to directly use as a distance metric in graph calculations, it needs to be scalarized. For example, the formula Z=(R) can be used. 2 +X 2 ) 0.5 The complex modulus of the positive-sequence impedance is calculated as the scalar edge weight, where R is the resistance parameter and X is the reactance parameter. For high-voltage distribution network scenarios, since the R / X ratio (resistance / reactance ratio) of the lines is extremely small, the voltage drop effect of the resistive component is negligible; therefore, the positive-sequence reactance value is extracted as the scalar edge weight. Impedance scalarization reduces the storage and computational dimensionality of the underlying graph structure.

[0067] S13: Assign scalar edge weights to the corresponding edges of the original physical topology, and label the protection device attributes of each node based on the device type information of each node in the original physical topology, thus obtaining the power grid topology impedance weighted graph. After obtaining the scalar edge weights of each connection, map the corresponding values ​​back to the topology network to form an undirected weighted graph that can be directly used for graph traversal algorithms. In addition, traverse all nodes and mark the protection device attributes of nodes with circuit breaker, load switch, or fuse functions. After completing the scalarization and attribute labeling, it can be solidified into a pre-constructed power grid topology impedance weighted graph, serving as the static physical base for subsequent identification of associated work orders.

[0068] In one embodiment, based on the semantic vector extracted from the customer complaint work order record and the topological location features extracted from the power grid topology impedance weighted map, the fault mode probability distribution corresponding to the customer complaint work order record is mapped to obtain the following, such as... Figure 3 As shown, it includes the following steps:

[0069] S21. Extract the semantic vector from the complaint text in customer complaint work order records. Since customer complaint work order records typically contain text records left by customers when calling customer service hotlines or reporting repairs online, the complaint text is often an unstructured natural language description. This step transforms the unstructured natural language text into a high-dimensional feature representation that can be processed by a computer. In practice, a pre-trained language model is used to encode the long and short texts in the customer service records, outputting a fixed-dimensional floating-point array as a semantic vector. For example, a 768-dimensional dense vector can be output to represent key information such as power outages, flickering, and odors in the customer's description. This method captures potential common semantic features under different customer expression habits. Alternatively, a pre-trained language model can be used to vectorize the complaint text. The pre-trained language model can be implemented using a bidirectional encoder representation model or a robustly optimized bidirectional encoder representation model, outputting a fixed-dimensional semantic vector. This effectively captures the deep semantic information in the customer's repair description, such as distinguishing superficially similar but substantially different descriptions of power anomalies, providing high-dimensional data-driven features for subsequent fault mode classification.

[0070] S22. Extract the topological location features corresponding to the access nodes in customer complaint work order records from the power grid topology impedance weighted map. In addition to the textual semantic vector, it is also necessary to extract the spatial dimension of objective physical information. Since the customer complaint work order records are accompanied by the corresponding customer account number, the customer's corresponding access node can be located in the power grid topology impedance weighted map based on the customer account number. Then, the spatial representation can be extracted using the network structure to obtain the topological location features.

[0071] S23. The semantic vector and topological location features are concatenated and mapped to a predefined set of fault types through a classification model to obtain the fault mode probability distribution. This mainly involves concatenating the semantic vector and topological location features, mapping them to a predefined set of fault types, generating a probability distribution matrix, and obtaining the fault mode probability distribution. In specific implementation, after obtaining the semantic vector and topological location features, a bit-by-bit concatenation operation is performed to form a joint feature vector containing multimodal information. The concatenated fusion vector simultaneously possesses the joint representation capability of subjective complaint information and objective physical constraints. Using a fault classification model composed of a fully connected layer and a classification activation function (i.e., a normalized exponential function), the joint feature vector is projected onto a predefined set of fault types covering common typical fault types in specific distribution networks, including short-circuit faults, single-phase ground faults, and overload faults. The output probability distribution matrix is ​​a normalized one-dimensional array, where the sum of all elements in the array is 1. The output fault mode probability distribution is represented as a normalized probability vector for each work order. The values ​​of each element in this probability vector range from 0 to 1, and the sum of all elements equals 1. Arranging the probability vectors of all work orders in rows forms the probability distribution matrix. For example, if the predefined fault mode set includes short circuit, grounding, and overload, the output probability distribution matrix for a specific customer complaint work order record can be a one-dimensional array containing three elements. The normalized probability vector values ​​are 0.1, 0.1, and 0.8, respectively, indicating that the probability of the work order being caused by a short circuit is 10%, the probability of it being caused by a grounding fault is 10%, and the probability of it being caused by an overload fault is as high as 80%. This probabilistic soft output mechanism achieves data-driven soft classification, providing fault-tolerant redundancy information for subsequent calculations of spatiotemporal correlations under similar symptoms, and providing non-deterministic latent variable conditions to support the subsequent spatiotemporal scoring decay mechanism. This effectively avoids the amplification of classification errors caused by directly classifying work orders into a single fault type.

[0072] The fault classification model described in this application requires offline training to obtain parameters. Its training data consists of historical customer complaint work order records whose actual fault types have been confirmed through on-site maintenance. These records are derived from complaint work orders in the power system's historical operation and maintenance records that have been manually diagnosed and labeled with fault types. Each training sample includes the complaint text of the work order, the location information of the corresponding access node in the power grid topology impedance weighted diagram, and a manually labeled fault type as a monitoring signal. The training data can be obtained from the historical operation and maintenance records of the power customer service system. The specific scale and distribution of the training data can be adjusted according to the fault type coverage requirements in the actual application scenario. During training, the cross-entropy loss function can be used to measure the difference between the probability distribution output by the model and the true labels. Those skilled in the art can use conventional model training methods in the field to train the fault classification model, and the hyperparameters during the training process can be determined through conventional hyperparameter tuning methods.

[0073] In one embodiment, the topological location features corresponding to the access nodes in the customer complaint work order records are extracted from the power grid topology impedance-weighted map, such as... Figure 4 As shown, the process includes the following:

[0074] S31. Based on the impedance weighted graph of the power grid topology, the impedance electrical distance from the access node to the main power source is extracted. The impact of faults at different locations in the distribution network varies significantly on downstream and even upstream nodes. To accurately characterize the spatial attenuation properties of physical impacts, this application abandons the traditional geographical straight-line distance and adopts the electrical distance from the access node to the main power source. This impedance electrical distance truly reflects the electrical depth of the node in the radial power grid. The main power source typically refers to the bus or feeder outlet node of the substation where the distribution network is located. A shortest path search algorithm can be used to calculate the sum of the weights of all scalar edges along the path from the customer access node where the work order occurred to the main power source. This sum is used as the impedance electrical distance to quantify the absolute depth of the customer in the entire distribution network. Specifically, a graph traversal algorithm is used to trace from the access node along the network topology to the main power source of the distribution network, and the weights of the scalar edges corresponding to each branch line along the way are accumulated. The accumulated value is the impedance electrical distance from the access node to the main power source.

[0075] S32. Based on the power grid topology impedance weighted map, extract the branch order of the distribution network branch lines where the access node is located. After extracting the impedance electrical distance, it is also necessary to identify the network level of the access node, which is achieved by extracting the branch order. The branch order refers to the number of branch nodes traversed from the main line to the customer access node. For example, the branch order of a node directly connected to the main line is 0, while the branch order of a node connected after one branch is 1, that is, the branch order increases by 1 for each branch operation. Since high-order branches often correspond to smaller diameter terminal lines, the probability of local overload or poor contact faults is higher than that of the main line. Therefore, the branch order feature has important a priori discriminative value and can reflect the complexity of the network structure at the customer's location.

[0076] S33. Based on the impedance-weighted graph of the power grid topology, count the total number of users downstream of the access node. This can be done by traversing the power flow direction along the topology and calculating the total number of low-voltage customers in the subtree rooted at the access node. The total number of users is used to characterize the potential outage range after a fault at this node, or to measure the potential scope of the outage. The wider the impact of a node, the more likely it is to trigger a large number of complaints from the same source in a short period of time if it fails. Alternatively, a depth-first search or breadth-first search algorithm can be used to traverse all lower-level topology branches extending from the access node to the load side and calculate the sum of the effective customer counts at all end connections to obtain the total number of users.

[0077] S34. The impedance electrical distance, branch order, and total number of users are encoded and fused to form the topological location features. After obtaining the above three numerical features, their physical dimensions and numerical ranges are different, and direct use may affect the training stability of subsequent models. Therefore, numerical unification is required. To unify the above numerical features with different physical dimensions, a multilayer perceptron model can be used to perform nonlinear mapping on the above three numerical features, transforming them into high-dimensional latent vectors compatible with the semantic vector dimension, thereby obtaining the topological location features. Specifically, firstly, the impedance electrical distance, branch order, and total number of users are normalized, and then a multilayer perceptron model is used to map the above three one-dimensional numerical features into high-dimensional latent vectors, which constitute the topological location features. Through the encoding fusion mechanism, the topological constraints of the physical world are effectively transformed into a structured representation that can be read by the deep learning model and matched with semantic information.

[0078] In one embodiment, the joint correlation degree between work order pairs is calculated, for example, using the following formula:

[0079] ;

[0080] in, Work orders in customer complaint work order records With work orders The degree of joint correlation between them; A predefined set of fault types; The fault type is one of the fault types in the fault type set. and These are work orders in the failure mode probability distribution. With work orders By fault type The probability of triggering; Based on fault type Work orders for calculating control parameters With work orders The spatiotemporal similarity between them.

[0081] The calculation method of the above joint correlation formula is a variant application of the total probability formula. In actual distribution network customer service, there are phenomena such as power outages reported near the source and voltage flickering reported remotely. Although the two complaint symptoms are different, their underlying root causes may be the same. This can be achieved by traversing a predefined set of fault types. All fault types By treating each type of pre-defined fault as a latent variable and calculating its marginal expectation, the uncertainty brought about by apparent complaints can be effectively mitigated. Specifically, work orders... The probability of a fault being caused by a certain predetermined fault type and the work order The probabilities caused by the same fault type are multiplied together, and then multiplied by the spatiotemporal correlation degree for that predetermined fault type. The product results under all fault types are summed. The weighted scoring mechanism can ensure that when the system faces work order pairs with different apparent complaints but the same underlying cause, as long as the joint probability is high and conforms to the spatiotemporal decay law, it can still identify and output a high joint correlation degree.

[0082] In one embodiment, the spatiotemporal correlation is determined by fault type. The spatiotemporal attenuation basis function for the control parameters is calculated. The spatiotemporal attenuation basis function includes a spatial attenuation factor characterizing the impedance voltage division attenuation effect under the physical constraints of the distribution network, and a time alignment factor characterizing the phenomenological evolution of faults and the customer perception delay effect. The spatiotemporal correlation is determined by calculating the spatial attenuation factor and the time alignment factor.

[0083] Existing fault identification methods typically employ a fixed spatial distance threshold and a fixed time window threshold for hard matching across all fault types. However, short-circuit faults exhibit rapid propagation and wide-ranging impact, while overload faults show slow propagation and limited impact. Fixed threshold methods struggle to adapt to the varying physical evolution of different underlying faults. Therefore, this application abandons the fixed threshold method and introduces a spatiotemporal attenuation basis function with a predetermined fault type as the control parameter. This basis function integrates electrical principles with phenomenological laws. Specifically, the spatial attenuation factor is used to simulate the first-order approximation of voltage drop impedance division in a radial distribution network. The time alignment factor quantifies the time delay characteristics in the complete evolution chain of a fault, from its occurrence and the exceeding of physical parameter limits to customer awareness and the subsequent repair call. Through the combination of these two factors, the spatiotemporal correlation is transformed into a soft scoring index that can accommodate nondeterminism.

[0084] Specifically, the calculation formula for the spatiotemporal decay basis function in this application is as follows:

[0085] ;

[0086] in, For the corresponding fault type The pre-configured attenuation impedance, Work orders based on impedance parameters extracted from the impedance weighted graph of the power grid topology. With work orders The positive sequence impedance magnitude between access nodes; Work orders extracted based on the time attribute in customer complaint work order records With work orders The time difference of incoming calls; For the corresponding fault type The pre-configured phenomenological delay coefficient; For the corresponding fault type Pre-configured time tolerance parameters; This is the spatial decay factor; This is the time alignment factor.

[0087] In practical implementation, the positive sequence impedance magnitude Time difference with incoming call The formula can be normalized before being substituted into it, allowing the formula to operate in a dimensionless space; alternatively, a phenomenological lag coefficient can be pre-configured. Self-absorption dimension conversion, i.e. The physical meaning is the phenomenological delay time corresponding to a unit impedance distance, and its dimensions are... and Dimensionality is compatible. Specifically, the pre-configured attenuation impedance, pre-configured phenomenological delay coefficient, and pre-configured time tolerance parameter are obtained or determined through the following methods: obtaining pre-stored historical confirmed work order samples; using the historical confirmed work order samples, constructing maximum likelihood estimation functions or heuristic optimization fitness functions according to fault types; and extracting model parameters that maximize the spatiotemporal distribution probability of historical associated work orders under different fault types through offline fitting, which are then used as the pre-configured attenuation impedance, pre-configured phenomenological delay coefficient, and pre-configured time tolerance parameter for the corresponding fault types.

[0088] In one embodiment, the pre-configured attenuation impedance Pre-configured phenomenological delay coefficient and pre-configured time tolerance parameters The following steps can be used to obtain it:

[0089] Retrieve pre-stored historical confirmed related work order samples: These samples are derived from a set of related work order pairs confirmed through fault waveform analysis or on-site investigation in the power system operation and maintenance records. Each sample pair contains the positive sequence impedance magnitude between the access nodes of the two related work orders, the time difference of power arrival, and the confirmed fault type label. For each fault type... The impedance magnitude and time difference of each work order pair in the corresponding historical sample set are extracted as observation pairs. Based on the analytical form of the spatiotemporal attenuation basis function, a pre-configured attenuation impedance is constructed. Pre-configured phenomenological delay coefficient and pre-configured time tolerance parameters The maximum likelihood estimation function is used to estimate the parameters to be determined. The optimization objective of this function is to maximize the probability of the observed spatiotemporal distribution of historically related work orders. This function is solved offline using heuristic optimization methods such as nonlinear least squares or genetic algorithms. The parameter combinations that optimize the objective function are extracted and used as the pre-configured attenuation impedance, pre-configured phenomenological delay coefficient, and pre-configured time tolerance parameter for the corresponding fault type. Those skilled in the art can select appropriate optimization methods based on the scale and distribution characteristics of the actual historical data. The above optimization process is a conventional parameter estimation operation in this field.

[0090] The pre-configured attenuation impedance of this application Pre-configured phenomenological delay coefficient and pre-configured time tolerance parameters The acquisition process can also be:

[0091] Retrieve pre-stored historical confirmed related work order samples: These samples refer to a set of historical work order pairs confirmed by on-site maintenance personnel, possessing clear correlations and known fault types. Each sample record contains the positive sequence impedance magnitude between the access nodes corresponding to the two related work orders, as well as the time difference of incoming power. Using these historical confirmed related work order samples, group the samples according to fault type. For each fault type... Using the spatiotemporal decay basis function As the likelihood kernel function, the impedance magnitude and time difference of incoming calls corresponding to all samples within the fault type group are substituted to construct the maximum likelihood estimation function. The optimization objective of the maximum likelihood estimation function is to maximize the joint log-likelihood of all historically diagnosed work order samples under this fault type, i.e., to find the value that makes the maximum likelihood kernel function the most likely value. The parameter combination that achieves the maximum value; when the sample size is insufficient to support the stable convergence of the gradient optimization method, a heuristic optimization fitness function can also be constructed and solved using a genetic algorithm or particle swarm optimization algorithm. The maximum likelihood estimation function or heuristic optimization fitness function is solved offline, and the model parameters that maximize the spatiotemporal distribution probability of historically associated work orders under different fault types are extracted and used as the pre-configured attenuation impedance for the corresponding fault type. Pre-configured phenomenological delay coefficient and pre-configured time tolerance parameters Those skilled in the art can use conventional numerical optimization tools to complete the above-mentioned offline fitting process.

[0092] To illustrate the spatiotemporal soft decay principle and joint correlation calculation logic described above, a calculation example based on normalized values ​​is provided. Let's assume the extracted work order... With work orders After being mapped using a pre-configured semantic model, the probabilities of both being overload faults are 0.9 and 0.8, respectively. To simplify calculations, the predefined fault type set is set to include only overload. The normalized positive-sequence impedance magnitude between the two nodes is extracted through the aforementioned steps. =2.0, and the normalized call time difference between the two work orders. =4.0. Pre-configured attenuation impedance for overload faults. The pre-configured phenomenological delay coefficient is 8.0. Version 2.0, pre-configured time tolerance parameters The value is 1.0. Under the aforementioned parameter configuration, the calculated spatiotemporal correlation degree for overload faults is 0.8, and the joint correlation degree is 0.576, which is higher than the preset correlation threshold of 0.5. Therefore, this work order pair passes the soft scoring stage. The results show that when the time difference between the incoming power of two work orders exactly matches the expected propagation delay of their electrical distance, the time alignment factor reaches its maximum value, and the joint correlation degree is significantly higher than zero. The aforementioned processing adapts to the physical characteristics of slow overload fault propagation and effectively captures hidden correlated work orders with large spans but conforming to the physical laws of the distribution network.

[0093] For different power distribution network operating environments, dynamic pre-configured phenomenological delay coefficients and time tolerance parameters can be set for different seasons or peak load periods. For example, during the peak summer season, high temperatures accelerate the heat accumulation process of power distribution equipment, leading to a faster evolution of overload faults. In this case, adaptive coefficients can be used to proportionally reduce the pre-configured phenomenological delay coefficients, making the time alignment factor more consistent with the actual fault propagation speed under the predetermined physical environment.

[0094] Because relay protection actions in power systems have strict spatiotemporal coordination logic and hierarchical relationships, only by arranging protection devices according to the topological series sequence in the actual power grid can the physical process of fault propagation being successfully blocked by one or more circuit breakers along the way be accurately simulated in the subsequent physical reachability product calculation. This provides a rigorous basis for eliminating phantom-related work order pairs that do not meet the connectivity conditions. Therefore, this application proposes to extract the ordered sequence of protection devices corresponding to candidate related work order pairs in the power grid topology impedance weighted diagram. For example, in one embodiment, the ordered sequence of protection devices corresponding to candidate related work order pairs in the power grid topology impedance weighted diagram is extracted, such as... Figure 5 As shown, it includes the following steps:

[0095] S41. In the power grid topology impedance-weighted graph, a shortest path search algorithm is executed based on the two access nodes corresponding to the candidate associated work order pairs to obtain the shortest topological path connecting the two access nodes. As mentioned earlier, the power grid topology impedance-weighted graph has access nodes as vertices and power lines as edges, with positive-sequence impedance magnitudes representing electrical distance attached to the edges as scalar edge weights. Using a shortest path search algorithm (such as Dijkstra's algorithm or A* algorithm), starting from the access node corresponding to the first work order in the candidate associated work order pair and ending at the access node corresponding to the second work order, the algorithm traverses the graph along the direction of minimum impedance weights. By executing the shortest path search algorithm, a unique connected path with the minimum impedance electrical distance is determined in the radial or ring-shaped distribution network structure; this is the shortest topological path. Obtaining the shortest topological path is to eliminate branch lines unrelated to the actual fault propagation between the two access nodes, accurately limiting the subsequent analysis scope to the main lines where fault current or voltage sag is most likely to propagate.

[0096] For distribution network topologies containing a large number of nodes, a bidirectional breadth-first search algorithm can be used to advance the search from two access nodes to the intermediate node simultaneously, reducing the computational complexity and memory overhead of the pathfinding process.

[0097] S42. Following the shortest topology path, extract nodes or edges with protection device attributes along the shortest topology path. In the original physical topology of the distribution network, substation outgoing circuit breakers, line sectionalizing switches, pole-mounted circuit breakers, and fuses on the high-voltage side of distribution transformers are typically modeled as predetermined nodes or edges in the graph structure and pre-assigned with corresponding protection device attribute labels when constructing the power grid topology impedance weighted graph. In this step, the topology path obtained in the previous steps is traversed hop-by-hop, filtering out ordinary cable or overhead line segment nodes, retaining only elements carrying protection device attribute labels. Through targeted attribute filtering, physical protection devices with fault clearing and physical isolation capabilities can be accurately extracted from a large number of topology elements. The extracted devices constitute the physical barriers when faults propagate in space. Furthermore, the protection device attributes not only include the device type identifier but can also include the device's unique asset code in the power grid dispatch automation system, so that the corresponding tripping or reclosing state set can be directly matched and extracted from the protection device action log using this asset code.

[0098] S43. Arrange the extracted nodes or edges according to the spatial topology connection order to form an ordered sequence of protection devices. The spatial topology connection order specifically refers to the physical order in which nodes move along the topology path from the starting access node to the ending access node. Arranging the extracted nodes or edges with protection device attributes sequentially according to this movement direction generates a queue in the form of a one-dimensional array, i.e., the ordered sequence of protection devices. Assume that there are three protection devices sequentially on a certain topology path: a substation outgoing line switch, a main line sectionalizing switch, and a branch line fuse. According to the spatial topology connection order, these three protection devices will be recorded sequentially as the first, second, and third elements of the sequence, respectively.

[0099] In one embodiment, the physical reachability of the ordered sequence of protection devices is calculated using the protection device action log to determine valid associated work order pairs. This is mainly achieved by dynamically constructing a log query window and determining the polymorphic coordination logic of the protection devices. Candidate associated work order pairs are physically reachable verified, and work order pairs that fail the physical reachability verification are eliminated, while valid associated work order pairs are retained. Figure 6 As shown, the following steps are performed:

[0100] S51. Based on customer complaint work order records, extract the call times of the work order pairs. Use the later call time of the two work orders, with a pre-configured communication delay margin added, as the right endpoint. Use the right endpoint back to the pre-configured maximum propagation delay as the left endpoint to construct a log query window. The pre-configured communication delay margin is used to compensate for the time difference in the transmission of equipment action signals to the master station system, and can be determined by those skilled in the art based on the communication link delay characteristics of the actual power distribution automation system. The maximum propagation delay is an empirical parameter set according to different fault modes, and can be determined by those skilled in the art based on the physical propagation characteristics of different fault types. The maximum propagation delay for fast-propagating faults is less than that for slow-propagating faults. Using the later call time of the two work orders as the benchmark to construct the right endpoint can effectively cover later tripping events, avoiding the problem of missed action log checks due to customer repair delays or communication delays.

[0101] S52. In the log query window, extract the action status set of each protection device in the ordered sequence of protection devices from the protection device action log. The extracted action status set includes the tripping action, closing action, and protection over-limit alarm signal of each node device in the sequence within a predetermined time period, which constitutes the data basis for subsequent judgment on whether the power grid physical topology has been disconnected by protection action.

[0102] S53, based on the action state set, determine the blocking status of each protection device in the ordered sequence of protection devices, and calculate physical reachability in combination with the blocking status to complete the retention of valid associated work order pairs. In this application, the blocking status of each protection device in the ordered sequence of protection devices is divided into effective isolation status, ineffective isolation status, or no limit exceedance / no impact status, and the judgment rules are as follows:

[0103] When a successful tripping action is recorded in the log query window and no successful reclosing is recorded, the system is considered to be in a valid isolation state. A valid isolation state indicates that the fault has been completely isolated by the device, and the electrical connection is broken at this point. In business logic, this constitutes an insurmountable physical barrier, ensuring that power outages for users on both sides of the device cannot be directly caused by the same fault.

[0104] If a successful reclosing record exists in the log query window, or if a failure to operate and cascading trip is confirmed based on pre-configured protection coordination logic, the system is deemed to be in an ineffective isolation state. This application defines ineffective isolation as addressing two scenarios. The first is a successful reclosing caused by a transient fault; although a trip occurs, the power grid subsequently recovers. The second is a failure to operate by the local protection device, leading to a cascading trip by the upstream protection; in this case, although the fault current flows through the local device, the local device fails to disconnect. Neither of these scenarios constitutes a substantial blockage in the physical diagram. Specifically, when the local protection device has no tripping action record in the log query window, but its upstream protection device has a tripping action record, and the tripping time is within the reasonable time range within which the local device should have operated, the system is determined to have experienced a failure to operate and cascading trip based on the pre-configured protection coordination logic.

[0105] If no action records are found in the log query window, it is determined to be in a non-limit-crossing / non-impacted state. A non-limit-crossing / non-impacted state indicates that the device did not detect any fault characteristics during the corresponding time period, or the fault characteristics did not reach its action threshold, and therefore maintained a normal closed conduction state in the topology.

[0106] In one optional embodiment, based on protection action logic constraints or coordination relationships, the step-by-step action status of each protection device in the ordered sequence of protection devices is determined, and then the blocking status is judged: Specifically, according to the protection coordination logic, the actual blocking effect of each protection device in the ordered sequence of path protection devices on fault propagation is determined one by one, and it is marked as either effective isolation or ineffective isolation. The determination logic is as follows:

[0107] For any protection device in the protection device sequence, first check its action event record. If a tripping event exists in the record and no reclosing success event occurs after the trip, the device has performed a permanent trip, and the fault is determined to be isolated at this point. The device is then marked as effectively isolated. If a tripping event exists in the record and a reclosing success event subsequently occurs, it indicates that the device has performed reclosing and successfully restored power supply. The fault is transient, and the fault isolation state no longer lasts after successful reclosing. In this case, it is necessary to further determine whether the power arrival time of both work orders is earlier than the reclosing success time: if so, the fault is still in an isolated state at the time the work order was generated, and it is marked as effectively isolated; if the power arrival time of either work order is later than the reclosing success time, it is marked as not effectively isolated.

[0108] If the action event record in the window of the protection device is empty, meaning there are no action events in the query window, further determination is needed to determine whether the device has failed to operate. The determination method is as follows: search for the adjacent protection device upstream of this device in the ordered sequence of path protection devices, i.e., closer to the substation, and check whether there are any tripping events in its action event record. If the upstream protection device trips after the set time limit for the current protection device to operate, i.e., the upstream protection operates as backup protection, it can be inferred that the current protection device has failed to operate, and the fault has spread across the current protection interval to the upstream level. In this case, although the current protection device physically exists, it has not actually isolated the fault and should be marked as ineffective isolation.

[0109] If neither the local nor the upstream protection device has any recorded action, it indicates that no signs of fault propagation were detected in the protection zone. In this case, the device is marked as unreachable, meaning the fault most likely did not propagate to the protection zone, and the device does not need to take action. After each device is assessed individually, each device in the ordered sequence of path protection equipment receives a corresponding status: effective isolation, ineffective isolation, or unreachable propagation.

[0110] This application, combining the calculation of physical reachability with the blocking status, completes the retention of effective associated work order pairs. Physical reachability can be calculated and judged based on the calculation results as follows: Effective isolation status is mapped to a value of 1, and ineffective isolation status and non-limit-crossing or non-affected status are mapped to a value of 0, obtaining the blocking coefficient corresponding to each protection device in the ordered sequence of protection devices; physical reachability is calculated using a product function, the formula is as follows: = ( - ), For physical accessibility, To protect the total number of protected devices in the ordered sequence of protected devices, To protect the first in the ordered sequence of equipment The above formula represents a hard blocking model with a veto mechanism. If any blocking coefficient in the sequence has a value of 1, the overall physical reachability immediately becomes 0. When... When = 0, the corresponding candidate associated work order pair is removed; when When =1, the corresponding candidate associated work order pair will be retained as a valid associated work order pair.

[0111] To more clearly illustrate the effect of the hard blocking mechanism, a specific numerical calculation process is introduced as an example. Assume that between the two nodes extracted from two work orders, there are a total of three protection devices along the topological path. According to the decision logic, the blocking coefficient sets of the three protection devices are as follows: In the case of ineffective isolation, Corresponding to an effective isolation status, This corresponds to a state where the limit has not been exceeded or the impact has not occurred. Substituting this into the product formula, the physical reachability H = (1-0) × (1-1) × (1-0) = 0. It can be concluded that although this work order pair may have achieved a high score (e.g., 0.576) in the previous joint probability soft scoring stage, the system will still remove it through a hard blocking mechanism because there are effectively isolated devices on its topological path. This processing method eliminates the illusory associations caused by simply relying on probability calculations or spatial proximity, ensuring the physical rationality of the final retained work order pair.

[0112] In this embodiment, after filtering out isolated events that do not conform to physical laws, a graph algorithm is used to aggregate effective related work order pairs into a whole, and multi-dimensional heterogeneous signals are fused to infer the original root cause work order that caused a large number of repeated calls before outputting it. Effective related work order pairs are clustered into related work order groups, and multi-dimensional judgment signals are fused within these groups to infer the causal propagation direction, thus identifying and outputting the root cause work order, such as... Figure 7 As shown, it includes:

[0113] S61. Map valid associated work order pairs to a graph structure, perform connected component detection, and obtain associated work order groups. Specifically, treat each independent customer complaint work order record as a vertex in the graph structure, and treat valid associated work order pairs that pass physical reachability verification as undirected edges connecting the two vertices. By traversing all valid associated work order pairs, dynamically construct one or more undirected connected graphs. Use depth-first search / breadth-first search algorithms to perform connected component extraction operations on the graph structure. Each independent connected subgraph corresponds to an associated work order group, representing a set of concurrent complaints caused by the same physical root cause within a predetermined spatiotemporal range.

[0114] S62. Based on the fault mode probability distribution, the power grid topology impedance weighted map, and customer complaint work order records, extract the fault phenomenon level signal, topology location level signal, and temporal sequence signal from work orders within the associated work order group to form a multi-dimensional judgment signal. The fault phenomenon level signal refers to the physical representation level obtained based on the work order complaint text classification; the topology location level signal refers to the hierarchical relationship of the access node corresponding to the work order in the power flow direction of the distribution network; and the temporal sequence signal is the time sequence relationship based on the absolute timestamp of the customer calling customer service or submitting a repair request through a terminal. The fault phenomenon level signal can be obtained by classifying the complaint text based on a pre-trained language model or by mapping keywords in the complaint text. Those skilled in the art can choose an appropriate classification method according to the actual data volume and accuracy requirements. By simultaneously extracting the above signals from various dimensions and transforming them into structured feature vectors, a multi-dimensional judgment signal is formed, providing a comprehensive basis for subsequent causal relationship inference.

[0115] S63. Calculate the comprehensive causal confidence score of each work order within the associated work order group by integrating multi-dimensional judgment signals, and determine the work order with the highest comprehensive causal confidence score as the root cause work order. Specifically, in calculating the comprehensive causal confidence score of each work order within the associated work order group by integrating multi-dimensional judgment signals, priority weights are configured for each of the three dimensions according to preset priorities, and weighted judgment is performed. The preset priority order is as follows: the highest priority fault phenomenon level signal has the highest weight, with work orders belonging to the root cause level receiving a basic confidence score of 3, work orders belonging to the conduction level receiving a basic confidence score of 2, and work orders belonging to the terminal level receiving a basic confidence score of 1. When the fault phenomenon levels are the same, the contribution of the next lower priority topology location level signal is further superimposed: sorted by the impedance electrical distance from each work order access node to the main power supply point from smallest to largest, the work order with the highest ranking receiving an additional score of 0.3, decreasing by 0.1 for each subsequent ranking. When there are still work orders with the same score after the first two levels of judgment, the contribution of the lowest priority time sequence signal is finally added: the earliest call time receives an additional score of 0.03, and then decreases by 0.01 for each subsequent call; the sum of the three levels of scores is the comprehensive causal confidence of each work order.

[0116] In the highest priority fault phenomenon level signals, the confidence level of the root cause level is higher than that of the conduction level, and the confidence level of the conduction level is higher than that of the end level. Specifically, the root cause level includes work order types that directly reflect abnormalities in core equipment, such as high temperature in distribution transformers or abnormal noise in transformers; the conduction level includes work order types that reflect deterioration in line transmission conditions, such as low voltage or three-phase imbalance; and the end level includes work order types that reflect subjective perceptions of end customers, such as flickering lights or inability to start household appliances. In actual distribution network operation, due to the significant differences in the response time of customers discovering abnormalities and reporting repairs, relying solely on the repair time can easily lead to causal reversal. This application assigns the highest priority to phenomena reflecting essential equipment defects. As long as a work order is determined to be at the root cause level, even if its call time is late, its basic weighted score in the comprehensive causal confidence calculation is higher than that of end-level work orders. In the second priority topology location level signals, the confidence level of upstream locations is higher than that of downstream locations. When multiple work orders with the same fault level exist in a group of related work orders, a secondary priority determination will be introduced based on the radial topology of the power grid. Upstream locations are typically near substations or the start of branch lines, while downstream locations are near the end of lines. According to the power system fault evolution and propagation mechanism, anomalies in upstream equipment often trigger widespread customer awareness downstream; therefore, work orders located upstream are given a higher confidence weight. In the lowest priority category, which is the time sequence signal, the earlier the incoming call time, the higher the confidence score. If, after determining the highest and secondary priorities, multiple work orders still have the same confidence level, the time sequence signal will be used as the final determining factor; the work order with the earlier timestamp will be considered the event origin and assigned an additional confidence score.

[0117] Through rigorous recursive calculations using a three-tiered priority system, the overall causal confidence of each work order is quantified. A ranked list with specific numerical values ​​is then output, and the work order record with the highest score in the list is extracted and marked as the root cause work order. In actual work order scheduling and customer service feedback, not only the root cause work order but also its associated work order group is output, allowing maintenance personnel to comprehensively understand the distribution of all duplicate call work orders caused by the same root cause. The root cause work order is typically treated as the master work order and dispatched to the frontline repair team for on-site verification. The remaining work orders within the associated work order group are merged into sub-work orders for unified progress tracking. This multi-dimensional fusion inference mechanism overcomes the problem of incorrect root cause localization caused by the traditional system's reliance on timestamps, improving the accuracy of emergency repair resource scheduling.

[0118] To more clearly illustrate the multi-dimensional judgment signal fusion mechanism described above, a specific application scenario is given. Assume that the aforementioned steps result in a group of related work orders containing three work orders: Work order A comes from a customer near the distribution transformer, complaining of a humming and overheating transformer, with the call time being 10:45; Work order B comes from a customer in the middle section of the downstream main line of the transformer, complaining of low voltage at home, with the call time being 10:30; Work order C comes from a customer on the terminal branch line, complaining of flickering lights, with the call time being 10:35. First, the highest priority judgment is applied: Work order A's complaint belongs to the root cause level (transformer noise), Work order B belongs to the conduction level (low voltage), and Work order C belongs to the terminal level (flickering lights). Based on this, Work order A receives the highest basic confidence score. Since the fault phenomena of the three are at different levels, subsequent topology location and time order only produce minor additional scores and do not affect the sorting result. The system identifies Work order A as the root cause work order and outputs it. In this scenario, although work order A was the latest to arrive, it was still able to correctly locate the root cause because its fault symptom level had the highest priority. This avoids the problem of misjudging work order B as the root cause in traditional methods that simply rely on the order of time.

[0119] To verify the effectiveness of the method of this invention, tests were conducted using actual operation and maintenance data of a regional power distribution network. The test data covered customer complaint work order records generated by multiple feeders over several months. Experimental results show that, compared with traditional correlation methods based on fixed spatiotemporal windows, the method of this application improves both the accuracy of work order correlation and the correctness of root cause localization, while effectively reducing false correlation judgments caused by protection device actions. Those skilled in the art will understand that the specific performance improvement may vary depending on the scale of the power distribution network, topology characteristics, and data quality.

Claims

1. A method for identifying repeat call work orders based on customer request correlation analysis, characterized in that: include: Obtain customer complaint work order records, power grid topology impedance weighted diagrams, and protection device action logs; Based on the semantic vector extracted from customer complaint work order records and the topological location features extracted from the power grid topology impedance weighted map, the fault mode probability distribution corresponding to the customer complaint work order records is mapped to obtain the probability distribution of the fault mode. Based on the probability distribution of fault modes, the spatiotemporal similarity of work order pairs determined by fault type is fused and weighted to obtain the joint correlation degree between work order pairs. Candidate associated work order pairs are selected based on joint correlation degree, and the ordered sequence of protection devices corresponding to the candidate associated work order pairs in the power grid topology impedance weighted graph is extracted. The physical reachability of the ordered sequence of protection devices is calculated using the action logs of the protection devices to determine the valid associated work order pairs; Effectively associated work order pairs are clustered into associated work order groups. Multi-dimensional judgment signals are integrated within the associated work order groups to infer the causal propagation direction and determine the root work order.

2. The method for identifying repeat call tickets according to claim 1, characterized in that, Based on the semantic vector extracted from customer complaint work order records and the topological location features extracted from the power grid topology impedance weighted map, the fault mode probability distribution corresponding to the customer complaint work order records is mapped to obtain the following: Extract semantic vectors from customer complaint work order records; Extract the topological location features of the access nodes from the customer complaint work order records from the power grid topology impedance weighted map; By concatenating semantic vectors with topological location features and mapping them to a predefined set of fault types, the probability distribution of fault modes corresponding to customer complaint work order records is obtained.

3. The method for identifying repeat call tickets according to claim 2, characterized in that, Extract the topological location features corresponding to the access nodes in customer complaint work order records from the power grid topology impedance-weighted map, including: Based on the impedance weighted graph of the power grid topology, the impedance electrical distance from the access node to the main power source point is extracted; Extract the branch order of the distribution network branch line where the access node is located; Count the total number of users downstream of the access node; The impedance electrical distance, branch order, and total number of users are encoded and fused to form the topological location features.

4. The method for identifying repeat call tickets according to claim 1, characterized in that, The joint correlation degree between the work order pairs is calculated using the following formula: ； in, Work orders in customer complaint work order records With work orders The degree of joint correlation between them; A predefined set of fault types; The fault type is one of the fault types in the fault type set. and These are work orders in the failure mode probability distribution. With work orders By fault type The probability of triggering; Based on fault type Work orders for calculating control parameters With work orders The spatiotemporal similarity between them.

5. The method for identifying repeat call tickets according to claim 1, characterized in that, The spatiotemporal similarity is determined by the spatiotemporal attenuation basis function, based on the spatial attenuation factor characterizing the impedance voltage division attenuation effect under the physical constraints of the distribution network, and the time alignment factor constructed using a Gaussian function characterizing the phenomenological evolution of faults and the perceived delay effect by customers, including: ； in, For the corresponding fault type The pre-configured attenuation impedance, Work orders based on impedance parameters extracted from the impedance weighted graph of the power grid topology. With work orders The positive sequence impedance magnitude between access nodes; Work orders extracted based on the time attribute in customer complaint work order records With work orders The time difference of incoming calls; For the corresponding fault type The pre-configured phenomenological delay coefficient; For the corresponding fault type Pre-configured time tolerance parameters; This is the spatial decay factor; This is the time alignment factor.

6. The method for identifying repeat call tickets according to claim 1, characterized in that, Extract the ordered sequence of protection devices corresponding to candidate associated work order pairs in the power grid topology impedance weighted graph, including: In the power grid topology impedance weighted graph, based on the two access nodes corresponding to the candidate associated work order pair, the shortest path search algorithm is executed to obtain the shortest topology path connecting the two access nodes. Extract nodes or edges with protected device attributes along the shortest topological path. The extracted nodes or edges are arranged according to the spatial topology connection order to obtain an ordered sequence of protection devices.

7. The method for identifying repeat call tickets according to claim 1, characterized in that, The physical reachability of ordered sequences of protection devices is calculated using the protection device action logs to determine valid associated work order pairs, including: Based on customer complaint work order records, extract the call time of work order pairs, use the later call time of the two work order call times plus a pre-configured communication delay margin as the right endpoint, and use the pre-configured maximum propagation delay back from the right endpoint as the left endpoint to construct a log query window; Based on the log query window, extract the set of action statuses of protection devices from the ordered sequence of protection device action logs; Based on the action state set, the blocking status of the protection device in the ordered sequence of protection devices is determined. Combined with the blocking status, physical reachability is calculated, and work order pairs that fail to pass the physical reachability test are eliminated, while valid related work order pairs are retained.

8. The method for identifying repeat call tickets according to claim 7, characterized in that, The blocking status of protection devices in an ordered sequence of protection devices is determined based on the action state set, including: If there is a successful tripping action and no successful reclosing record in the log query window, it is determined to be in an effective isolation state; if there is a successful reclosing record in the log query window or if the protection coordination logic determines that there is a failure to operate or an over-level trip, it is determined to be in an ineffective isolation state; if there is no action record in the log query window, it is determined to be in a state of no over-limit / no impact.

9. The method for identifying repeated call tickets according to claim 8, characterized in that, Based on the blocking status, calculate physical reachability, remove work order pairs that fail physical reachability, and retain valid associated work order pairs, including: Map the effective isolation state to the value 1, and map the ineffective isolation state and the non-limit-crossing / non-affected state to the value 0 to obtain the blocking coefficient corresponding to the protection device in the ordered sequence of protection devices; Based on the blocking coefficient, physical reachability is calculated using a product function, expressed as follows: = ( - ), For physical accessibility, To protect the total number of protected devices in the ordered sequence of protected devices, To protect the first in the ordered sequence of equipment The blocking coefficient corresponding to each protection device; when When =0, the corresponding candidate associated work order pair is removed; When =1, the corresponding candidate associated work order pair will be retained as a valid associated work order pair.

10. The method for identifying repeat call tickets according to claim 1, characterized in that, By integrating multi-dimensional judgment signals within a group of related work orders, the direction of causal propagation is inferred, and the root work order is identified, including: Map the valid associated work order pairs to a graph structure, perform connected component detection, and obtain the associated work order group; Extract the fault phenomenon level signal, topological location level signal, and time sequence signal of the work orders within the associated work order group to form a multi-dimensional judgment signal; By integrating multi-dimensional decision signals, the comprehensive causal confidence score of work orders within the associated work order group is calculated, and the work order with the highest comprehensive causal confidence score is identified as the root work order.

Images