A machine learning based order matching method and system

Abstract

The application discloses a kind of order matching method and system based on machine learning, including obtaining maintenance order and maintenance worker data.Based on geographic location and maintenance type, generate geographic proximity and skill compatible order subset cluster, for each subcluster, according to maintenance order, predict the required minimum number of maintenance workers and initialize the distribution pool accordingly, build order-maintenance worker bipartite graph within the subcluster, apply graph neural network to generate joint embedding vector, perform multiple rounds of destruction-repair optimization cycle: remove part of order allocation randomly in each round, reinsert after forced to meet skill matching and work time constraints in the repair phase;If there is a conflict, dynamically expand the maintenance worker distribution pool, iterate until the solution converges, and output the final order allocation scheme.This application improves the spatial efficiency of order matching, resource allocation rationality and accuracy, ensures constraint compatibility and adaptability, optimizes service efficiency and satisfaction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an order matching method and system based on machine learning. Background Technology

[0002] In the repair service sector, the efficiency of matching orders with repair workers directly impacts service response speed, resource utilization, and user satisfaction. Current technologies often suffer from unreasonable spatial scheduling in order matching, failing to fully consider the geographical distribution of orders. This leads to excessively long travel times for repair workers across regions, reducing service efficiency. Regarding the allocation of repair worker numbers, estimations often rely on a single dimension (such as order quantity or estimated working hours), ignoring the diverse skill requirements of different repair types, easily resulting in staff shortages or redundancy. Furthermore, the lack of comprehensive consideration of skill matching and working hour allocation frequently leads to substandard service quality due to incomplete skill coverage, or service progress affected by unbalanced working hours, ultimately impacting the overall stability and reliability of the service. Summary of the Invention

[0003] To address at least one of the aforementioned technical problems, this invention provides a machine learning-based order matching method and system.

[0004] In a first aspect, the present invention provides a machine learning-based order matching method, the method comprising:

[0005] Obtain order data and repairman data. The order data includes geographical location, set of repair types, and estimated repair time. The repairman data includes skill set and maximum daily working hours.

[0006] Based on the geographical location of orders, order sub-clusters are generated using a density clustering algorithm;

[0007] Based on the total number of orders, total estimated working hours, and skill diversity index within each order sub-cluster, predict the lower limit of the number of maintenance workers required for each order sub-cluster.

[0008] Create a maintenance worker allocation pool with the minimum number of maintenance workers as the initial value;

[0009] Construct a bipartite graph of orders and repair workers within the sub-cluster, and generate a joint embedding vector that integrates location, repair type, and duration using a graph neural network;

[0010] Based on the joint embedding vector, a multi-round damage repair loop is executed. In each repair phase, skill matching and time constraints are forcibly satisfied. When the constraints cannot be satisfied, the repair worker allocation pool is dynamically expanded until the convergence condition is reached; the final order allocation scheme is output.

[0011] Preferably, the step of predicting the lower limit of the number of maintenance workers required for each order sub-cluster based on the total number of orders, total estimated working hours, and skill diversity index within the order sub-cluster includes:

[0012] The ratio of the total estimated working hours of orders within the sub-cluster to the average daily working hours of maintenance workers is used as the base quantity.

[0013] The basic quantity is weighted and corrected based on the skill diversity index, which is quantified by the uniformity of skill type distribution.

[0014] Divide the total number of orders by the average daily processing capacity of the repairman as a reference value for quantity;

[0015] The maximum value between the weighted adjusted base quantity and the quantity reference value is taken as the lower limit of the number of maintenance workers.

[0016] Preferably, the construction of the bipartite graph of orders and repair workers within the sub-cluster, and the generation of a joint embedding vector that fuses location, repair type, and duration using a graph neural network, includes:

[0017] Treat all orders within the sub-cluster and the repairmen in the repairman allocation pool as independent node sets in a bipartite graph, and establish fully connected edges between order nodes and repairman nodes.

[0018] Each order node is assigned a multi-dimensional feature vector, including normalized geographic location coordinates, a multi-hot encoded vector of the repair type set, and the deviation value between the expected repair time and the standard time.

[0019] Each maintenance worker node is assigned a multi-dimensional feature vector, including the coordinates of the center point of the historical service area, the overlap vector between the skill set and the current sub-cluster maintenance type requirements, and an efficiency factor calculated based on historical work orders.

[0020] A three-layer graph attention neural network is used to process the bipartite graph. By aggregating features such as geographic location correlation, repair type matching degree and time duration adaptability through message passing between nodes, the final output is a joint embedding vector with a dimension of 128, where each embedding vector represents the comprehensive matching degree of a specific order-repairman pairing in terms of spatial accessibility, skill coverage adequacy and work hour load balance.

[0021] Preferably, the step of executing multiple rounds of damage repair loops based on the joint embedding vector, forcibly satisfying skill matching and time constraints in each repair phase, and dynamically expanding the repair worker allocation pool when the constraints cannot be satisfied, until the convergence condition is reached, includes:

[0022] A weighted probability approach is used to remove a preset proportion of existing assignments, with pairings that are below a skill matching threshold or have excessive working hours being removed first.

[0023] Based on the redistribution constraints, repair workers are reassigned to the removed orders. The redistribution constraints include that the repair worker's skill set must include all repair types of the order and that the total time of a single repair worker's assigned orders is less than the product of the maximum daily working time and the utilization rate threshold.

[0024] When the number of consecutive failures during the repair phase reaches the preset limit, based on the missing skill type and the amount of work time shortfall, candidate repairmen who meet the requirements of skill coverage of more than 90% and historical work time utilization of less than 85% are transferred from the surrounding area to the repairman allocation pool.

[0025] When the global matching degree improves by less than 0.5% in three consecutive rounds of iterations, the convergence condition is considered met.

[0026] Secondly, the present invention also provides a machine learning-based order matching system, the system comprising:

[0027] The data acquisition module is used to acquire order data and repairman data. The order data includes geographical location, set of repair types and estimated repair time. The repairman data includes skill set and maximum daily working hours.

[0028] The order clustering module is used to generate order sub-clusters based on the geographical location of orders using a density clustering algorithm;

[0029] The maintenance worker demand forecasting module is used to predict the lower limit of the number of maintenance workers required for each order sub-cluster based on the total number of orders, total estimated working hours, and skill diversity index within the order sub-cluster.

[0030] The maintenance worker allocation pool initialization module is used to create a maintenance worker allocation pool with the minimum number of maintenance workers as the initial value.

[0031] The graph construction and embedding module is used to construct a bipartite graph of orders and repairmen within a sub-cluster, and to generate a joint embedding vector that integrates location, repair type and duration through a graph neural network.

[0032] The iterative optimization matching module is used to execute multiple rounds of damage repair loops based on the joint embedding vector. In each round of repair, skill matching and time constraints are forcibly satisfied. When the constraints cannot be satisfied, the repair worker allocation pool is dynamically expanded until the convergence condition is reached; the final order allocation scheme is output.

[0033] Preferably, the maintenance worker demand prediction module is further used for:

[0034] The ratio of the total estimated working hours of orders within the sub-cluster to the average daily working hours of maintenance workers is used as the base quantity.

[0035] The basic quantity is weighted and corrected based on the skill diversity index, which is quantified by the uniformity of skill type distribution.

[0036] Divide the total number of orders by the average daily processing capacity of the repairman as a reference value for quantity;

[0037] The maximum value between the weighted adjusted base quantity and the quantity reference value is taken as the lower limit of the number of maintenance workers.

[0038] Preferably, the graph construction and embedding module is further used for:

[0039] Treat all orders within the sub-cluster and the repairmen in the repairman allocation pool as independent node sets in a bipartite graph, and establish fully connected edges between order nodes and repairman nodes.

[0040] Each order node is assigned a multi-dimensional feature vector, including normalized geographic location coordinates, a multi-hot encoded vector of the repair type set, and the deviation value between the expected repair time and the standard time.

[0041] Each maintenance worker node is assigned a multi-dimensional feature vector, including the coordinates of the center point of the historical service area, the overlap vector between the skill set and the current sub-cluster maintenance type requirements, and an efficiency factor calculated based on historical work orders.

[0042] A three-layer graph attention neural network is used to process the bipartite graph. By aggregating features such as geographic location correlation, repair type matching degree and time duration adaptability through message passing between nodes, the final output is a joint embedding vector with a dimension of 128, where each embedding vector represents the comprehensive matching degree of a specific order-repairman pairing in terms of spatial accessibility, skill coverage adequacy and work hour load balance.

[0043] Thirdly, the present invention also provides an electronic device including a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs the method as described in the first aspect above and any possible implementation thereof.

[0044] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor of an electronic device, cause the processor to perform a method as described in the first aspect above and any possible implementation thereof.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] This invention constructs an intelligent order matching mechanism that balances spatial rationality, skill adaptability, and work hour balance by integrating spatial clustering, multi-dimensional resource demand prediction, graph neural network feature fusion, and dynamic optimization loops. Specifically, density clustering based on geographical location enables spatially intensive processing of orders, reducing the inefficiency of cross-regional scheduling; multi-dimensional parameters such as total order volume, total work hours, and skill diversity are used to predict the lower limit of the number of repair workers, providing a scientific basis for initial resource allocation; then, a graph neural network is used to fuse multi-modal features of orders and repair workers, such as geographical location, repair type, and duration, into a joint embedding vector to quantify the comprehensive matching degree; furthermore, a destruction-repair loop dynamic optimization allocation scheme is adopted, combined with the mandatory satisfaction of skill and work hour constraints and the dynamic expansion of the allocation pool, to achieve continuous iterative optimization of the matching scheme, ultimately achieving efficient, accurate, and constraint-compatible order-repair worker matching. Density clustering improves the spatial efficiency of order matching to reduce cross-regional commuting costs. Multi-dimensional parameters are used to predict the lower limit of the number of repair workers to optimize resource allocation and avoid redundancy or shortage. Graph neural networks are used to fuse multi-modal features to improve matching accuracy. The destruction-repair cycle and dynamic expansion of the allocation pool ensure the compatibility of skills and working hours constraints and the adaptability of the solution. Ultimately, this comprehensively improves service response efficiency, reliability, and the satisfaction of both users and repair workers.

[0047] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the background art, the accompanying drawings used in the embodiments of the present invention or the background art will be described below.

[0049] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the specification, serve to illustrate the technical solutions of this disclosure.

[0050] Figure 1 A flowchart illustrating a machine learning-based order matching method provided in an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the structure of an order matching system based on machine learning, provided in an embodiment of the present invention. Detailed Implementation

[0052] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0054] The current repair service order matching suffers from technical problems such as unreasonable spatial scheduling, a single dimension in the allocation of repairman numbers that ignores the need for skill diversity, a lack of comprehensive consideration in the allocation of skills and working hours, and insufficient dynamic adaptability.

[0055] Please see Figure 1 , Figure 1 This is a flowchart illustrating a machine learning-based order matching method provided in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] S100, acquire order data and repairman data. The order data includes geographical location, set of repair types and estimated repair time. The repairman data includes skill set and maximum daily working hours.

[0057] In this embodiment, in the steps of acquiring order data and repairman data, the repair order set can be collected through various channels such as repair requests submitted online by users, telephone appointment records, and demand information synchronized from third-party platforms. The geographical location information included in each order can specifically be latitude and longitude coordinates, a detailed address, or real-time location obtained through location services. The repair type set covers the category of equipment to be repaired (such as air conditioners, refrigerators, washing machines, etc.) and the type of fault (such as abnormal cooling, circuit faults, mechanical damage, etc.). The estimated repair time is estimated based on the processing time of similar historical faults, the complexity of the fault description, and the experience of the repair personnel. The repairman set comes from the system's registered repairman information database. The skill set includes the equipment repair qualifications possessed by the repairman, the types of faults they can handle, and their areas of expertise. The maximum daily working time is the total daily time that the repairman can pre-set for performing repair work, which can deduct necessary rest time and commuting time.

[0058] S200 generates order sub-clusters based on the geographical location of orders using a density clustering algorithm;

[0059] In this embodiment, during the process of generating order sub-clusters based on the geographical location of orders using a density clustering algorithm, the geographical locations of all orders are first standardized using spatial coordinates. A distance metric suitable for geographic space (such as actual surface distance based on latitude and longitude) is used to calculate the spatial distance between orders. Then, the density clustering algorithm is used to identify dense areas of order distribution. Taking each order as the core, the number of other orders within a certain spatial range around it is counted. When this number reaches a preset density threshold, these orders are grouped into the same initial cluster. The initial clusters are further merged and adjusted, and spatially adjacent initial clusters with similar densities are merged into a sub-cluster. At the same time, for isolated sparse orders, if their distance from a certain sub-cluster is within a set range, they are included in that sub-cluster, ultimately forming multiple order sub-clusters with relatively concentrated spatial distribution.

[0060] In one possible implementation, a neighborhood radius parameter is determined, taking into account regional traffic characteristics and maintenance skill distribution features. The basic spatial radius is determined based on the product of the historical average driving speed of the target area and the average repair time per order. Furthermore, the global maintenance type distribution dispersion is introduced as a compensation term; the more uneven the skill distribution, the larger the compensation value. During the density accessibility determination stage, not only is it required that the geographical distance between two orders be within the neighborhood radius, but the similarity of their maintenance type sets also exceeds a preset threshold to ensure skill compatibility among orders within the same sub-cluster. During clustering, orders that simultaneously meet the core point density condition (the number of orders within the neighborhood meets the standard) and the skill representativeness condition (their maintenance types cover high-frequency types within the neighborhood) are selected as initial core points. Geographically proximate and skill-matched order sub-clusters are formed through density propagation. Boundary points are assigned to the cluster containing the nearest core point that meets the skill similarity requirement. The final generated sub-clusters satisfy a dual constraint: the geographical distance between any two orders does not exceed the regional accessibility range, and there exists at least one maintenance worker skill within the cluster that can cover all order type requirements.

[0061] S300, based on the total number of orders, total estimated working hours and skill diversity index within each order sub-cluster, predicts the lower limit of the number of maintenance workers required for each order sub-cluster;

[0062] In this embodiment, in the step of predicting the lower limit of the number of repair workers based on the total number of orders, total estimated working hours, and skill diversity index within an order sub-cluster, the total estimated working hours are the cumulative value of the estimated repair time for all orders within the sub-cluster; the skill diversity index is calculated by statistically analyzing the number of repair types and the distribution of each type within the sub-cluster. A higher index indicates a greater variety of repair types with a balanced distribution, and vice versa; the basic number is obtained by dividing the total estimated working hours by the average daily effective working hours of the repair workers (i.e., the actual time used for repair operations within the maximum daily working time); the reference value is calculated by dividing the total number of orders by the average daily order volume processed by the repair workers (based on the average level statistically analyzed from historical data); when weighting the basic number, a higher skill diversity index results in a greater correction weight to reflect the demand for multi-skilled repair workers; finally, the weighted basic number is compared with the reference value, and the larger value is taken as the lower limit of the number of repair workers required for the sub-cluster, ensuring that both the total working hour requirement and the order volume are covered.

[0063] S400, create a maintenance worker allocation pool with the minimum number of maintenance workers as the initial value;

[0064] In this embodiment, when creating a maintenance worker allocation pool with a minimum number of maintenance workers as the initial value, maintenance workers whose service areas cover the sub-cluster's area or are relatively close are first selected based on the geographical location of the order sub-cluster. Secondly, maintenance workers whose skill sets can cover at least some of the maintenance types are further selected based on the maintenance type set of the sub-cluster. According to the above selection criteria, maintenance workers with a number not less than the minimum are selected from the maintenance worker set and included in the allocation pool. If the number of maintenance workers meeting the criteria is insufficient, the maintenance worker with the highest skill matching degree and the closest distance is selected first, and the gap information is recorded. At the same time, the allocation pool will reserve a certain amount of expansion space so that maintenance workers can be dynamically supplemented according to the constraint satisfaction situation in the future.

[0065] S500 constructs a bipartite graph of orders and repairmen within the sub-cluster, and generates a joint embedding vector that integrates location, repair type, and duration through a graph neural network;

[0066] In this embodiment, during the construction of a bipartite graph of orders and repairmen within a sub-cluster and the generation of a joint embedding vector using a graph neural network, the order node set of the bipartite graph contains all orders within the sub-cluster, and the repairman node set contains all repairmen in the allocation pool. Connection edges are established between order nodes and repairman nodes to represent potential allocation relationships. In the multi-dimensional feature vector assigned to order nodes, the normalized geographic location coordinates are obtained by mapping the order's geographic location to standardized values ​​within a specific spatial range; in the multi-hot encoded vector of the repair type set, each dimension corresponds to a possible repair type; if the order contains that type, the dimension value is 1, otherwise it is 0; the deviation between the expected repair time and the standard repair time is calculated by the difference between the expected repair time of the order and the average processing time for similar faults. In the multi-dimensional feature vector assigned to the repairman node, the coordinates of the center point of the historical service area are obtained based on the geographical weighted average of all past service orders of the repairman; in the overlap vector between the skill set and the current sub-cluster repair type demand, each dimension corresponds to a repair type within the sub-cluster, and the dimension value is a quantitative representation of whether the repairman possesses the skill of that type; the efficiency factor is calculated comprehensively based on the ratio of actual processing time to expected time in the repairman's historical work orders, work order completion quality score, etc. When using a three-layer graph attention neural network to process the bipartite graph, the first layer learns the basic representation of the node's own features, the second layer aggregates the geographical location association features between the order and the repairman through message passing between nodes (e.g., the closer the distance, the higher the association), and the third layer integrates the repair type matching degree (skill overlap degree) and the time adaptability (the synergistic relationship between the efficiency factor and the time deviation). The final output joint embedding vector comprehensively represents the matching degree of the order-repairman pairing from three dimensions: spatial accessibility, skill coverage sufficiency, and time load balance.

[0067] S600: Based on the joint embedding vector, execute multiple rounds of damage repair loops. In each repair phase, skill matching and time constraints are forcibly satisfied. When the constraints cannot be satisfied, the repair worker allocation pool is dynamically expanded until the convergence condition is reached; output the final order allocation scheme.

[0068] In this embodiment, during the multi-round disruption-repair cycle based on the joint embedding vector, the disruption phase assigns weights to existing allocation pairs based on the matching degree reflected by the joint embedding vector. Pairs with low matching degrees (such as insufficient skill matching or unbalanced workload) are assigned a higher removal probability, and a certain proportion of existing allocation relationships are removed according to this probability. The repair phase reassigns maintenance workers to the removed orders, strictly adhering to the following redistribution constraints: first, the maintenance worker's skill set must completely include all maintenance types required by the order to ensure processing capacity; second, the total estimated duration of all orders assigned to the same maintenance worker does not exceed the product of their maximum daily working time and a set utilization threshold to avoid workload overload. When the repair phase fails to find a maintenance worker allocation scheme that meets the constraints for the removed orders multiple times consecutively, it is determined that the current allocation pool capacity is insufficient. At this time, based on the missing skill types and unmet workload gaps in the sub-cluster, candidate maintenance workers meeting the conditions are selected from the maintenance worker set in the neighboring region and added to the allocation pool. Candidate maintenance workers must possess the missing key skills, and their historical workload utilization rate must be at a low level to ensure sufficient idle time. During the loop, the global matching degree (the average matching degree of the joint embedding vector of all orders-repairman pairings) is continuously calculated. When the improvement of the global matching degree in multiple consecutive loops is lower than the set standard, the convergence condition is determined, the loop stops and the final order allocation scheme is output.

[0069] In this embodiment, an intelligent order matching mechanism that balances spatial rationality, skill adaptability, and work hour balance is constructed by integrating spatial clustering, multi-dimensional resource demand prediction, graph neural network feature fusion, and dynamic optimization loop. Specifically, density clustering based on geographical location enables spatially intensive processing of orders, reducing the inefficiency of cross-regional scheduling; multi-dimensional parameters such as total order volume, total work hours, and skill diversity are used to predict the lower limit of the number of repair workers, providing a scientific basis for initial resource allocation; then, a graph neural network is used to fuse the multi-modal features of orders and repair workers, such as geographical location, repair type, and duration, into a joint embedding vector to quantify the comprehensive matching degree; furthermore, a destruction-repair loop dynamic optimization allocation scheme is adopted, combined with the mandatory satisfaction of skill and work hour constraints and the dynamic expansion of the allocation pool, to achieve continuous iterative optimization of the matching scheme, ultimately achieving efficient, accurate, and constraint-compatible order-repair worker matching. Density clustering improves the spatial efficiency of order matching to reduce cross-regional commuting costs. Multi-dimensional parameters are used to predict the lower limit of the number of repair workers to optimize resource allocation and avoid redundancy or shortage. Graph neural networks are used to fuse multi-modal features to improve matching accuracy. The destruction-repair cycle and dynamic expansion of the allocation pool ensure the compatibility of skills and working hours constraints and the adaptability of the solution. Ultimately, this comprehensively improves service response efficiency, reliability, and the satisfaction of both users and repair workers.

[0070] Preferably, the step of predicting the lower limit of the number of maintenance workers required for each order sub-cluster based on the total number of orders, total estimated working hours, and skill diversity index within the order sub-cluster includes:

[0071] The ratio of the total estimated working hours of orders within the sub-cluster to the average daily working hours of maintenance workers is used as the base quantity.

[0072] The basic quantity is weighted and corrected based on the skill diversity index, which is quantified by the uniformity of skill type distribution.

[0073] Divide the total number of orders by the average daily processing capacity of the repairman as a reference value for quantity;

[0074] The maximum value between the weighted adjusted base quantity and the quantity reference value is taken as the lower limit of the number of maintenance workers.

[0075] In this embodiment, when calculating the ratio of the total estimated working hours of orders within a sub-cluster to the average daily working hours of maintenance workers as the basic quantity, the total estimated working hours of orders within a sub-cluster are the sum of the estimated repair times of all orders included in the sub-cluster. The estimated repair time of each order can be determined based on the actual processing time of similar repair orders in history, the complexity of the fault description, and the repair difficulty level. The average daily working hours of maintenance workers refer to the effective time that maintenance workers can actually use to perform repair work in a working day. This time needs to deduct necessary rest time, commuting time between orders, and other non-repair operation time. It is usually determined based on the average level of the average daily effective working hours of maintenance workers in historical statistics. Through the ratio of the total estimated working hours to the average daily working hours, the basic number of maintenance workers required to meet the working hours requirements of the sub-cluster is initially quantified.

[0076] In the step of weighted adjustment of the basic quantity based on the skill diversity index, the skill diversity index is quantified by statistically analyzing the distribution characteristics of repair types within the sub-cluster. Specifically, it can include the number of repair types included in the sub-cluster, the proportion of each repair type in the total orders, and the evenness of distribution among different types. When there are many repair types and the proportion of each type is small, it indicates that the skill diversity index is high, and vice versa. During the weighted adjustment, the skill diversity index is positively correlated with the adjustment weight, that is, the higher the skill diversity index, the greater the corresponding adjustment weight. This is used to adjust the basic quantity upward, thereby reflecting the additional demand for repair workers with multiple repair skills and ensuring that the adjusted quantity can adapt to the diverse repair type needs within the sub-cluster. In the process of dividing the total number of orders by the average daily processing capacity of maintenance workers as a quantity reference value, the total number of orders is the total number of all orders contained in the sub-cluster; the average daily processing capacity of maintenance workers refers to the average number of maintenance orders that a maintenance worker can complete in a working day. This number is based on historical maintenance data statistics, taking into account factors such as the processing time of different maintenance types, the operational proficiency of maintenance workers, and the efficiency of order connection. It is usually taken as the average level of the number of orders completed by maintenance workers per day over a certain period of time. By using the ratio of the total number of orders to the average daily processing capacity, the required reference number of maintenance workers is quantified from the perspective of order quantity scale, so as to make up for the order quantity intensity demand that may be ignored by calculating only working hours.

[0077] In the step of using the maximum value between the weighted adjusted base quantity and the quantity reference value as the lower limit of the number of maintenance workers, the weighted adjusted base quantity needs to be compared with the aforementioned quantity reference value, and the larger value is selected as the lower limit of the number of maintenance workers required for the sub-cluster of the order. This is to ensure that the lower limit can meet the maintenance workload requirements corresponding to the total estimated working hours within the sub-cluster, and also cover the processing scale requirements corresponding to the total number of orders. This avoids the problem of insufficient number of maintenance workers due to considering only a single dimension, and provides a scientific and comprehensive quantity benchmark for the initial creation of the subsequent maintenance worker allocation pool, ensuring the rationality of the resource base in the subsequent order matching process.

[0078] Preferably, the construction of the bipartite graph of orders and repair workers within the sub-cluster, and the generation of a joint embedding vector that fuses location, repair type, and duration using a graph neural network, includes:

[0079] Treat all orders within the sub-cluster and the repairmen in the repairman allocation pool as independent node sets in a bipartite graph, and establish fully connected edges between order nodes and repairman nodes.

[0080] Each order node is assigned a multi-dimensional feature vector, including normalized geographic location coordinates, a multi-hot encoded vector of the repair type set, and the deviation value between the expected repair time and the standard time.

[0081] Each maintenance worker node is assigned a multi-dimensional feature vector, including the coordinates of the center point of the historical service area, the overlap vector between the skill set and the current sub-cluster maintenance type requirements, and an efficiency factor calculated based on historical work orders.

[0082] A three-layer graph attention neural network is used to process the bipartite graph. By aggregating features such as geographic location correlation, repair type matching degree and time duration adaptability through message passing between nodes, the final output is a joint embedding vector with a dimension of 128, where each embedding vector represents the comprehensive matching degree of a specific order-repairman pairing in terms of spatial accessibility, skill coverage adequacy and work hour load balance.

[0083] In this embodiment, all orders within the sub-cluster and the repairmen in the repairman allocation pool are treated as independent node sets in a bipartite graph. During the process of establishing fully connected edges between order nodes and repairman nodes, the order node set contains all unassigned repair orders within the sub-cluster, covering repair needs for different equipment types and fault degrees. The repairman node set contains all repairmen selected from the repairman set who initially meet the basic requirements of the sub-cluster, covering repair personnel with different skill combinations and service areas. The establishment of fully connected edges aims to cover all possible order-repairman pairings, regardless of whether the current repairman possesses the required skills for the order or their location. The connection of edges provides a complete potential matching space for subsequent feature learning, ensuring that the model can comprehensively consider the possibilities of various pairings. In the step of assigning a multi-dimensional feature vector to each order node, including normalized geographic coordinates, a multi-hot encoded vector of the repair type set, and the deviation value between the estimated repair time and the standard time, the normalized geographic coordinates are converted into standardized data within a uniform numerical range from the original geographic location information of the order (such as latitude and longitude or regional coding) to eliminate the scale difference between different geographic location representations. In the multi-hot encoded vector of the repair type set, each dimension corresponds to a preset repair type (such as home appliances, water and electricity, etc.). If the order contains this type, the corresponding dimension in the vector is assigned a specific quantitative value; otherwise, it is assigned another quantitative value, which intuitively presents the combination of repair skills required for the order. The deviation value between the estimated repair time and the standard time is calculated by comparing the estimated repair time of the order with the average processing time of similar fault orders. A positive number indicates that the estimated time of the order is higher than the average level, and a negative number indicates that it is lower than the average level, thus reflecting the relative complexity of the order.

[0084] Each repairman node is assigned a multi-dimensional feature vector, including the coordinates of the center point of the historical service area, the overlap vector between the skill set and the current sub-cluster's repair type requirements, and an efficiency factor calculated based on historical work orders. The coordinates of the center point of the historical service area are coordinates calculated by weighted averaging (e.g., by setting weights based on the number of services or service duration) of all repair orders completed by the repairman in the past, and are used to characterize the spatial range of their main services. In the overlap vector between the skill set and the current sub-cluster's repair type requirements, each dimension corresponds to a repair type existing in the sub-cluster. The vector value is quantitatively represented according to the repairman's mastery of the skills for that type (e.g., whether they have basic repair capabilities or rich experience), intuitively reflecting the degree of skill matching. The efficiency factor calculated based on historical work orders is a quantitative indicator derived by comprehensively considering factors such as the ratio of the actual time to the expected time of the repairman's past order processing and user ratings of service efficiency, and is used to reflect their efficiency level in handling repair work.

[0085] A three-layer graph attention neural network is used to process the bipartite graph. Through message passing between nodes, features of geographical location correlation, repair type matching, and time-end adaptability are aggregated, ultimately outputting a 128-dimensional joint embedding vector. Each embedding vector represents the comprehensive matching degree of a specific order-repairman pairing in terms of spatial accessibility, skill coverage sufficiency, and work hour load balancing. In the first layer, the graph attention neural network initially encodes the multi-dimensional feature vectors of the order node and the repairman node, extracting the basic representation of each feature. The second layer calculates the correlation weight between the order and repairman nodes through an attention mechanism (e.g., the closer the geographical location, the higher the weight), and aggregates the geographical location features of both parties based on this weight to form a fused representation of geographical location correlation. The third layer focuses on repair type matching (combining the order repair type encoding with the repairman skill overlap vector) and time-end adaptability (associating the order time-end deviation value with the repairman efficiency factor), integrating the two types of features through message passing. Finally, the output features of the three layers are fused into a 128-dimensional joint embedding vector. Each dimension of this vector corresponds to a quantitative representation of spatial accessibility (such as the closeness of geographical location association), skill coverage sufficiency (such as the completeness of repair type matching), and work hour load balancing (such as the degree of work hour suitability for repair workers in handling the order), providing comprehensive feature basis for subsequent order matching optimization.

[0086] In this embodiment, spatial clustering is used to achieve regionalized processing of orders. A minimum number of repair workers is predicted based on multi-dimensional parameters including the total number of orders, total estimated working hours, and skill diversity to establish a reasonable initial resource base. A bipartite graph of orders and repair workers is then constructed. A graph neural network is used to fuse multi-modal features such as location, repair type, and duration to generate a joint embedding vector to quantify the overall matching degree. Finally, a multi-round destruction-repair cycle dynamically optimizes the allocation scheme, forcibly satisfying skill and working hour constraints and dynamically expanding the repair worker allocation pool, forming an intelligent order matching mechanism that balances spatial rationality, skill adaptability, and working hour balance. Density clustering improves the spatial efficiency of order matching to reduce cross-regional commuting costs. Predicting a minimum number of repair workers based on multi-dimensional parameters optimizes resource allocation to avoid redundancy or shortage. Graph neural networks fuse multi-modal features to improve matching accuracy. The destruction-repair cycle and dynamically expanded allocation pool ensure compatibility with skill and working hour constraints and scheme adaptability, ultimately comprehensively improving service response efficiency, reliability, and both user and repair worker satisfaction.

[0087] Preferably, the step of executing multiple rounds of damage repair loops based on the joint embedding vector, forcibly satisfying skill matching and time constraints in each repair phase, and dynamically expanding the repair worker allocation pool when the constraints cannot be satisfied, until the convergence condition is reached, includes:

[0088] A weighted probability approach is used to remove a preset proportion of existing assignments, with pairings that are below a skill matching threshold or have excessive working hours being removed first.

[0089] Based on the redistribution constraints, repair workers are reassigned to the removed orders. The redistribution constraints include that the repair worker's skill set must include all repair types of the order and that the total time of a single repair worker's assigned orders is less than the product of the maximum daily working time and the utilization rate threshold.

[0090] When the number of consecutive failures during the repair phase reaches the preset limit, based on the missing skill type and the amount of work time shortfall, candidate repairmen who meet the requirements of skill coverage of more than 90% and historical work time utilization of less than 85% are transferred from the surrounding area to the repairman allocation pool.

[0091] When the global matching degree improves by less than 0.5% in three consecutive rounds of iterations, the convergence condition is considered met.

[0092] In this embodiment, a weighted probability method is used to remove a preset proportion of existing assignments. In the step where pairings with skill matching below a threshold or exceeding the time limit are removed first, the weight of the weighted probability is determined based on the comprehensive matching degree represented by the joint embedding vector. The lower the matching degree of the existing assignment, the higher the removal probability. The skill matching degree is quantitatively evaluated by the overlap ratio between the repairman's skill set and the order repair type set. When this ratio does not reach the set standard, it is determined that the skill matching degree is insufficient. Exceeding the time limit refers to the total expected duration of all orders assigned to a repairman exceeding the reasonable load range corresponding to his / her maximum daily working hours. Such pairings are given higher priority removal rights because they may cause service delays or quality degradation. By selectively removing some assignments in the above manner, the aim is to break the local optimum and create optimization space for subsequent reallocation. In the process of reassigning repairmen to removed orders based on the reallocation constraints, the execution of these constraints is carried out throughout the matching and screening process: For the constraint that the repairman's skill set must include all repair types of the order, each repair type involved in the order must be compared one by one to ensure that the repairman's skills and qualifications cover all types, preventing insufficient service capacity due to skill mismatches; for the constraint that the total duration of orders assigned to a single repairman is less than the product of the maximum daily working time and the utilization threshold, the estimated duration of all currently assigned orders and the estimated duration of orders to be assigned to that repairman must be accumulated, ensuring that the accumulated value does not exceed the product of the maximum daily working time and the preset utilization threshold, to avoid overload of working hours affecting service efficiency; by strictly adhering to the above constraints, it is ensured that the reassigned orders are effectively processed while balancing the workload of repairmen.

[0093] When the number of consecutive failures during the repair phase reaches a preset limit, the step of adding candidate repairmen from surrounding areas to the repairman allocation pool is based on the missing skill type and the time gap. Consecutive failures refer to multiple unsuccessful attempts to match repairmen that meet the redistribution constraints for the removed order. The missing skill type is the repair type required for the order that is not covered by the skill set of repairmen in the current allocation pool. The time gap is the difference between the total effective working hours that repairmen in the current allocation pool can provide and the total estimated working hours of the sub-cluster orders. The surrounding area refers to the service area that is geographically close to the sub-cluster of the order and has low commuting costs. The selection of candidate repairmen requires that their skill set has a high coverage ratio of the missing skill type and that their historical working hour utilization rate is in a low range to ensure that they have sufficient idle time to take on new orders. By supplementing such repairmen, the resource gap in the current allocation pool can be effectively alleviated. In the step where the global matching degree improvement is less than the set standard after three consecutive iterations, the global matching degree is determined to be the average of the comprehensive matching degree represented by the joint embedding vector of all confirmed order-maintenance worker pairings within the order sub-cluster. This comprehensively reflects the overall optimization level of spatial accessibility, skill coverage sufficiency, and work hour load balancing. The improvement is calculated by the difference between the global matching degree of the next iteration and the global matching degree of the previous iteration. When the improvement is lower than the set standard for multiple consecutive iterations, it indicates that the current allocation scheme is close to the global optimum. Further iterative optimization has limited effect on improving the matching quality. Stopping the loop at this point can ensure the matching effect while avoiding unnecessary consumption of computational resources, ensuring that the final allocation scheme output is both efficient and stable.

[0094] This embodiment constructs a dynamic optimization loop of "destruction-repair," combining weighted probability removal of low-quality allocations to break local optima, strict adherence to skill and time constraints to ensure matching feasibility, dynamic expansion of the maintenance worker pool based on resource gaps to address local resource shortages, and a global matching degree convergence criterion to balance optimization effectiveness and efficiency. This forms a closed-loop optimization mechanism that considers matching quality, constraint compatibility, and resource adaptability, enabling continuous iterative upgrades of the order-maintenance worker allocation scheme. Weighted removal of low-matching-degree allocations breaks local optima, enhancing the global optimization potential of the matching scheme; strict skill and time constraints ensure the feasibility of the allocation scheme, avoiding service issues caused by skill mismatches or time overload; dynamic expansion of the allocation pool effectively addresses local resource shortages, improving the matching success rate; and the setting of convergence conditions ensures matching quality while avoiding ineffective iterations. Ultimately, this achieves a comprehensive improvement in the accuracy, feasibility, and resource utilization efficiency of the matching scheme, optimizing the overall service effect.

[0095] In summary, the method provided in this embodiment can achieve at least the following effects:

[0096] This invention constructs an intelligent order matching mechanism that balances spatial rationality, skill adaptability, and work hour balance by integrating spatial clustering, multi-dimensional resource demand prediction, graph neural network feature fusion, and dynamic optimization loops. Specifically, density clustering based on geographical location enables spatially intensive processing of orders, reducing the inefficiency of cross-regional scheduling; multi-dimensional parameters such as total order volume, total work hours, and skill diversity are used to predict the lower limit of the number of repair workers, providing a scientific basis for initial resource allocation; then, a graph neural network is used to fuse multi-modal features of orders and repair workers, such as geographical location, repair type, and duration, into a joint embedding vector to quantify the comprehensive matching degree; furthermore, a destruction-repair loop dynamic optimization allocation scheme is adopted, combined with the mandatory satisfaction of skill and work hour constraints and the dynamic expansion of the allocation pool, to achieve continuous iterative optimization of the matching scheme, ultimately achieving efficient, accurate, and constraint-compatible order-repair worker matching. Density clustering improves the spatial efficiency of order matching to reduce cross-regional commuting costs. Multi-dimensional parameters are used to predict the lower limit of the number of repair workers to optimize resource allocation and avoid redundancy or shortage. Graph neural networks are used to fuse multi-modal features to improve matching accuracy. The destruction-repair cycle and dynamic expansion of the allocation pool ensure the compatibility of skills and working hours constraints and the adaptability of the solution. Ultimately, this comprehensively improves service response efficiency, reliability, and the satisfaction of both users and repair workers.

[0097] See Figure 2 In one embodiment, a machine learning-based order matching system is also provided, the system comprising:

[0098] The data acquisition module 100 is used to acquire order data and repairman data. The order data includes geographical location, set of repair types and estimated repair time. The repairman data includes skill set and maximum daily working time.

[0099] The order clustering module 200 is used to generate order sub-clusters based on the geographical location of the orders using a density clustering algorithm;

[0100] The maintenance worker demand forecasting module 300 is used to predict the lower limit of the number of maintenance workers required for each order sub-cluster based on the total number of orders, total estimated working hours and skill diversity index within the order sub-cluster.

[0101] The maintenance worker allocation pool initialization module 400 is used to create a maintenance worker allocation pool with the minimum number of maintenance workers as the initial value.

[0102] The graph construction and embedding module 500 is used to construct a bipartite graph of orders and repairmen within a sub-cluster, and to generate a joint embedding vector that integrates location, repair type and duration through a graph neural network.

[0103] The iterative optimization matching module 600 is used to execute multiple rounds of damage repair loops based on the joint embedding vector. In each repair phase, skill matching and time constraints are forcibly satisfied. When the constraints cannot be satisfied, the repair worker allocation pool is dynamically expanded until the convergence condition is reached; and the final order allocation scheme is output.

[0104] Preferably, the maintenance worker demand prediction module 300 is further used for:

[0105] The ratio of the total estimated working hours of orders within the sub-cluster to the average daily working hours of maintenance workers is used as the base quantity.

[0106] The basic quantity is weighted and corrected based on the skill diversity index, which is quantified by the uniformity of skill type distribution.

[0107] Divide the total number of orders by the average daily processing capacity of the repairman as a reference value for quantity;

[0108] The maximum value between the weighted adjusted base quantity and the quantity reference value is taken as the lower limit of the number of maintenance workers.

[0109] Preferably, the graph construction and embedding module 500 is further configured to:

[0110] Treat all orders within the sub-cluster and the repairmen in the repairman allocation pool as independent node sets in a bipartite graph, and establish fully connected edges between order nodes and repairman nodes.

[0111] Each order node is assigned a multi-dimensional feature vector, including normalized geographic location coordinates, a multi-hot encoded vector of the repair type set, and the deviation value between the expected repair time and the standard time.

[0112] Each maintenance worker node is assigned a multi-dimensional feature vector, including the coordinates of the center point of the historical service area, the overlap vector between the skill set and the current sub-cluster maintenance type requirements, and an efficiency factor calculated based on historical work orders.

[0113] A three-layer graph attention neural network is used to process the bipartite graph. By aggregating features such as geographic location correlation, repair type matching degree and time duration adaptability through message passing between nodes, the final output is a joint embedding vector with a dimension of 128, where each embedding vector represents the comprehensive matching degree of a specific order-repairman pairing in terms of spatial accessibility, skill coverage adequacy and work hour load balance.

[0114] It is understood that the system provided in this embodiment has functions or includes modules that can be used to execute the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0115] The present invention also provides an electronic device including a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs a method as described in any of the above possible implementations.

[0116] The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0117] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0118] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. Those skilled in the art will also readily understand that the various embodiments of the present invention have different focuses, and for the sake of convenience and brevity, the same or similar parts may not be repeated in different embodiments. Therefore, parts not described or not described in detail in one embodiment can be referred to in other embodiments.

Claims

1. A machine learning based order matching method, characterized by, The method comprises: obtaining order data and repairman data, the order data comprising a geographic location, a set of repair types, and an estimated repair duration, and the repairman data comprising a set of skills, a maximum daily working time; generating order sub-clusters based on the geographic location of the orders by a density clustering algorithm; predicting a lower limit of the number of repairmen required for each order sub-cluster based on the total number of orders, the total estimated working hours, and a skill diversity index in the order sub-cluster; creating a repairman allocation pool with the lower limit of the number of repairmen as an initial value; constructing a bipartite graph of orders and repairmen in the sub-cluster, and generating a joint embedding vector that integrates location, repair type, and duration by a graph neural network; performing a multi-round destruction-repair cycle based on the joint embedding vector, and forcibly meeting skill matching and working hour constraints in each repair stage, dynamically expanding the repairman allocation pool when the constraints cannot be met, until a convergence condition is reached; and outputting a final order allocation scheme; the multi-round destruction-repair cycle based on the joint embedding vector, and forcibly meeting skill matching and working hour constraints in each repair stage, dynamically expanding the repairman allocation pool when the constraints cannot be met, until a convergence condition is reached, comprises: adopting a weighted probability to remove a preset proportion of existing allocations, and pairs with a skill matching degree below a threshold value or a working hour load exceeding a limit are preferentially removed; redistributing repairmen to the removed orders based on redistribution constraints, the redistribution constraints including that the skill set of a repairman must contain all repair types of an order, and the total duration of an order allocated to a single repairman must be less than the product of the maximum daily working time and a utilization threshold; when the number of consecutive failures in the repair stage reaches a preset upper limit, based on the missing skill types and the working hour gap, candidate repairmen that meet a skill coverage rate of more than 90% and a historical working hour utilization rate of less than 85% are transferred from surrounding areas to join the repairman allocation pool; when the global matching degree improves by less than 0.5% in three consecutive cycles, it is determined that the convergence condition is reached. 2.The machine learning based order matching method of claim 1, wherein, the prediction of the lower limit of the number of repairmen required for each order sub-cluster based on the total number of orders, the total estimated working hours, and the skill diversity index in the order sub-cluster comprises: calculating the ratio of the total estimated working hours of the orders in the sub-cluster to the daily working hours of a repairman as a basic number; weighting and correcting the basic number according to the skill diversity index, which is quantified by the uniformity of the skill type distribution; dividing the total number of orders by the daily processing capacity of a repairman as a quantity reference value; taking the maximum value between the weighted and corrected basic number and the quantity reference value as the lower limit of the number of repairmen. 3.The machine learning based order matching method of claim 1, wherein, the construction of a bipartite graph of orders and repairmen in the sub-cluster, and the generation of a joint embedding vector that integrates location, repair type, and duration by a graph neural network, comprises: treating all orders in the sub-cluster and repairmen in the repairman allocation pool as independent node sets of the bipartite graph, and establishing fully connected edges between the order nodes and the repairman nodes; assigning a multi-dimensional feature vector to each order node, including normalized geographic location coordinates, a multi-hot encoding vector of the set of repair types, and a deviation value of the estimated repair duration from a standard duration; assign a multi-dimensional feature vector to each repairman node, including the historical service area center point coordinates, the skill set and the current sub-cluster repair type demand overlap vector, and the efficiency factor calculated based on the historical work order; adopt a three-layer graph attention neural network to process the bipartite graph, aggregate the geographical location relevance, repair type matching degree and time length adaptability features through message passing between nodes, and finally output a joint embedding vector with a dimension of 128, wherein each embedding vector represents the comprehensive matching degree of a specific order-repairman pair in terms of spatial accessibility, skill coverage sufficiency and workload balance.

4. A machine learning based order matching system, characterized by, The system comprises: a data acquisition module for acquiring order data and repairman data, the order data including geographical location, repair type set and estimated repair time, and the repairman data including skill set and daily maximum working time; an order clustering module for generating order sub-clusters based on the geographical location of the orders through a density clustering algorithm; a repairman demand prediction module for predicting the lower limit of the number of repairmen required by each order sub-cluster based on the total number of orders, total estimated working hours and skill diversity index in the order sub-cluster; a repairman allocation pool initialization module for creating a repairman allocation pool with the lower limit of the number of repairmen as the initial value; a graph construction and embedding module for constructing a bipartite graph of orders and repairmen in the sub-cluster and generating a joint embedding vector that integrates location, repair type and time length through a graph neural network; an iterative optimization matching module for performing multiple rounds of destruction and repair cycles based on the joint embedding vector, forcibly meeting the skill matching and working hour constraints in each repair stage, dynamically expanding the repairman allocation pool when the constraints cannot be met, and outputting a final order allocation scheme until the convergence condition is reached; The multiple rounds of destruction and repair cycles based on the joint embedding vector, forcibly meeting the skill matching and working hour constraints in each repair stage, dynamically expanding the repairman allocation pool when the constraints cannot be met, and reaching the convergence condition, include: adopting a weighted probability to remove a preset proportion of existing allocations, wherein pairs with a skill matching degree below a threshold or a working hour load exceeding the limit are preferentially removed; redistributing repairmen to the removed orders based on redistribution constraints, including that the repairman skill set must contain all repair types of the order, and the total working hours of a single repairman allocated order must be less than the product of the daily maximum working time and the utilization threshold; when the number of consecutive failures in the repair stage reaches a preset upper limit, based on the missing skill type and the working hour gap, candidate repairmen that meet the skill coverage rate of more than 90% and the historical working hour utilization rate of less than 85% are adjusted from the surrounding area to join the repairman allocation pool; when the global matching degree improvement of three consecutive cycles is less than 0.5%, it is determined that the convergence condition is reached.

5. The machine learning based order matching system of claim 4, wherein, The repairman demand prediction module is also used to: calculate the ratio of the total estimated working hours of the orders in the sub-cluster to the daily working hours of the repairmen as the basic number; weight and correct the basic number according to the skill diversity index, which is quantified by the uniformity of the skill type distribution; divide the total number of orders by the daily processing capacity of the repairmen as the quantity reference value; The maximum value between the weighted corrected base number and the number reference value is taken as the lower limit of the number of repairmen.

6. The machine learning based order matching system of claim 4, wherein, The graph construction and embedding module is further configured to: all orders in the sub-cluster and repairmen in the repairman pool are respectively taken as independent node sets of a bipartite graph, and full connection edges are established between order nodes and repairman nodes; each order node is assigned a multi-dimensional feature vector, including normalized geographic location coordinates, a multi-hot encoding vector of a repair type set, and a deviation value of an estimated repair duration from a standard duration; each repairman node is assigned a multi-dimensional feature vector, including historical service area center point coordinates, a coincidence degree vector of a skill set and current sub-cluster repair type demand, and an efficiency factor calculated based on historical work orders; a three-layer graph attention neural network is used to process the bipartite graph, and geographic location relevance, repair type matching degree, and duration adaptability features are aggregated through message passing between nodes, and finally a joint embedding vector with a dimension of 128 is output, wherein each embedding vector represents the comprehensive matching degree of a specific order-repairman pair in terms of spatial accessibility, skill coverage sufficiency, and work time load balancing.

7. An electronic device, comprising: comprise: a processor and a memory, the memory being configured to store computer program code comprising computer instructions, when the processor executes the computer instructions, the electronic device executes the machine learning-based order matching method according to any one of claims 1 to 3.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program comprises program instructions, and the program instructions, when executed by a processor of an electronic device, cause the processor to execute the machine learning-based order matching method according to any one of claims 1 to 3.

Images