Order maintenance management platform and method based on machine learning, and electronic equipment
Through the machine learning-based order maintenance and management platform, the order grouping and delivery solution optimization is used to use data matrix processing and clustering algorithms, the problem of insufficient intelligent processing capabilities in the insurance industry is solved, and accurate insurance products are realized Recommendations and resource optimization have improved customer satisfaction and market competitiveness.
Patent Information
- Application Number
- CN202411561700.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-05
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing order maintenance and management platform lacks intelligent processing capabilities in the insurance industry and cannot effectively utilize multi-dimensional information, resulting in insufficient accurate recommendation of insurance products, insufficient in-depth analysis of customer needs, and difficult to deal with market changes and emergencies, and unreasonable resource allocation , customer satisfaction declines.
Adopt the order maintenance and management platform based on machine learning, through data acquisition matrix processing, clustering grouping and comprehensive planning modules, and use clustering algorithms and dynamic programming algorithms to optimize order grouping and delivery solutions, and build a collaborative filtering matrix and attention model. Implement accurate grouping and dynamic optimization of orders.
Improve the intelligence level and efficiency of order processing, realize in-depth mining and comprehensive analysis, provide personalized insurance product services, balance cost, time and quality, enhance robustness and adaptability, reduce operational costs, and improve customer satisfaction and market competitiveness.
Smart Images

Figure CN119067789A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of order maintenance management, and more specifically, to an order maintenance management platform, method and electronic device based on machine learning. Background Art
[0002] The patent with application publication number CN105791417A discloses an intelligent deployment and process monitoring system based on a cloud management platform, including an order processing module, a message module, an order parsing module, an intelligent deployment module, a failure message routing processing module and a monitoring module. The order processing module is used for users to submit orders and review the orders; the message module is used to receive and process orders that have passed the review; the order parsing module is used to parse the orders and obtain the deployment tasks and deployment strategies in the orders; the intelligent deployment module is used to intelligently deploy the orders; the failure message routing processing module is used to analyze the failure reasons for the orders that failed to be deployed and feed back to the intelligent deployment module; the monitoring module is used to associate the order deployment log with the successfully deployed orders, so that the front desk can monitor and query the order deployment log; it can greatly reduce the manpower and material resources invested in the deployment, management, upgrade and maintenance of traditional business systems, and reduce operating costs.
[0003] However, in the current insurance industry, there are still many problems to be solved in the order maintenance management platform; traditional maintenance management platforms often lack the ability to intelligently process complex insurance order data, and it is difficult to effectively utilize the multi-dimensional information contained in the order, resulting in inaccurate insurance product recommendations and insufficient in-depth customer demand analysis; in terms of order classification, the existing methods are too simplistic and cannot accurately reflect the similarities and correlations between different insurance orders, affecting the effectiveness of cross-selling and precision marketing; the formulation of order delivery plans often relies on static rules and manual experience, lacks the ability of dynamic optimization and adaptive adjustment, and is difficult to cope with emergencies; in addition, existing methods have difficulties in balancing multiple goals such as underwriting costs, claims settlement timeliness and service quality, and often have problems of unreasonable resource allocation and reduced customer satisfaction; in the face of rapid changes in the insurance market and increasingly diversified customer needs, traditional platforms lack scalability and flexibility, making it difficult to support the rapid development and launch of new insurance products; at the same time, in terms of risk assessment and pricing models, they lack the ability of self-learning and continuous optimization, and are unable to adapt to market changes in a timely manner.
[0004] In view of this, the present invention proposes an order maintenance management platform, method and electronic device based on machine learning to solve the above problems. Summary of the invention
[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: an order maintenance management platform based on machine learning, comprising: a data acquisition matrix processing module, a clustering grouping module and a comprehensive planning module; each module is connected by wire and / or wireless means; The data acquisition matrix processing module includes a data acquisition unit and a comprehensive processing unit. The data acquisition unit is used to acquire the comprehensive data of the order to be maintained; the comprehensive processing unit is used to perform matrix processing on the comprehensive data of the order to be maintained to acquire the corresponding m order matrices; The clustering grouping module uses a clustering algorithm based on m order matrices to group maintenance orders and obtain several order demand clusters; The comprehensive planning module includes a preliminary planning unit and a terminal optimization unit. The preliminary planning unit is used to obtain a preliminary order delivery plan for each order demand cluster using a dynamic programming algorithm. The terminal optimization unit is used to optimize the preliminary order delivery plan to obtain a terminal order delivery plan.
[0006] Furthermore, the comprehensive data of the order to be maintained includes basic order information, customer information and insurance information; the basic order information includes order number, order time, order type and order status; customer information includes customer type and customer name; insurance information includes insurance type, insured object details, insurance amount, premium and insurance period.
[0007] Furthermore, the order matrix is obtained by: Clean the comprehensive data of the maintenance orders, encode the data with categorical attributes and convert them into numerical data; normalize the continuous numerical data to obtain standard order data; the standard order data includes several order data fields; Construct an order feature matrix X, with each order to be maintained as a row of the order feature matrix, and the order data field corresponding to the order to be maintained as a column of the order feature matrix; the element X(i, j) in the order feature matrix X corresponds to the value of the jth order data field of the i-th order to be maintained; construct a collaborative filtering matrix based on the order feature matrix X , the order feature matrix X and the collaborative filtering matrix Merge and obtain m order matrices.
[0008] Furthermore, the collaborative filtering matrix is constructed The methods include: Extract different types of entities from standard order data; construct heterogeneous information networks based on different types of entities, and use each type of entity as a type of node in the heterogeneous information network; construct edges between nodes based on the relationship between entities; heterogeneous information networks contain several types of nodes and edges; Define the order network path. For each order network path, record any two nodes as a node pair. Calculate the comprehensive similarity score between any node pairs along the order network path. Construct a collaborative filtering matrix , whose elements Indicates The node and The comprehensive similarity score between nodes.
[0009] Furthermore, the method for obtaining the comprehensive similarity score includes: Perform several random walks in the heterogeneous information network to generate a series of node sequences. Each random walk starts from a random node and randomly selects the next node according to the connection relationship of the edge until the preset maximum step length is reached; define a context window with a size of , for each node in the node sequence The set of context nodes that define it ;in, is the index of the node; For each node and its context node ; Construct objective function ; ;in, For Node The index of the context node; is a randomly sampled negative example node; is the index of the negative example node; For Node For context nodes The weight of For the node Based on context node The conditional probability of For the node Based on negative example nodes The conditional probability of is the sigmoid function; The node The node vector is represented as , the context node The node vector is represented as ; then the conditional probability ;in, For Node The node vector of To traverse the index of all nodes; Optimize the objective function using gradient descent or its variants , learn the node vector and ; Each iteration updates the node vector of a node and its context node; The maximum number of training times is preset. After the maximum number of training iterations, the node vector corresponding to each node is obtained. For any node pair, the cosine similarity between the two corresponding node vectors is calculated as the similarity score between the two. For all node pairs in the heterogeneous information network, their similarity scores are calculated and a similarity matrix is constructed. , whose elements Indicates The node and The similarity score between nodes; the similarity matrix corresponding to each order network path obtained; For nodes, extract their eigenvalues ;in, For the The number of edges that actually exist between the neighbors of a node, For the The number of nodes connected to the node; the eigenvalue Converted into vector representation, we get The feature vector corresponding to the node ; Similarly Extract the eigenvalue of each node and convert it into a vector representation to get the The feature vector corresponding to the node ; Define an attention model, and As input, the output is the attention weight of each order network path; the attention model is a feedforward neural network or a recurrent neural network; For order network path, and its attention weight is recorded as , and the sum of the attention weights of all order network paths is 1; for the Order network paths, computing node pairs Similarity score under the corresponding order network path ; Weighted sum of similarity scores of all order network paths to get node pair The weight of the weighted sum is the attention weight output by the attention model.
[0010] Furthermore, the method of grouping the maintenance orders includes: Initialize clustering parameters, including clustering threshold T, leaf node constant factor number and cluster center; define an initialized clustering feature tree, scan the order matrix, and insert each order to be maintained as a clustering feature node into the clustering feature tree; if the current order to be maintained cannot be inserted into the clustering feature tree during scanning, generate a new clustering feature node; repeat until all orders to be maintained are inserted into the clustering feature tree to obtain a preliminary clustering feature tree; and ensure that the number of child nodes of each clustering feature node does not exceed the number of leaf node constant factors; the preliminary clustering feature tree contains several leaf nodes; Scan all leaf nodes of the preliminary clustering feature tree, cluster the clustering feature entries in the leaf nodes, and generate initial clusters; the clustering feature entries include the total number of data objects in the clustering feature entries, the linear sum of all data objects in each dimension, and the square sum of all data objects in each dimension; The clustering methods for clustering feature entries in leaf nodes include: Traverse each leaf node of the preliminary clustering feature tree, extract all clustering feature items in the leaf node, and store them in a list. Initially, select K clustering feature items from the list as the center of the cluster; K is a positive integer greater than 1; for each clustering feature item in the list, calculate its distance to the center of the K clusters; For each cluster feature item, assign it to the initial cluster that is closest to it, then the K initial clusters are filled, calculate the centroid of all cluster feature items in the cluster at this time, and use the centroid as the new center of the cluster; repeat until the center of the cluster no longer changes, and generate the initial cluster; According to the clustering threshold T, the initial clusters are merged to obtain the final order demand cluster. For any two initial clusters and , calculate the comprehensive distance between them ; ;in, for and The mean of the comprehensive similarity scores of the data objects in ; for and The distance between the farthest cluster feature entries in ; for and The mean distance between all cluster feature entries in , for and The distance between the centers of According to the calculated comprehensive distance, a clustering hierarchical tree is constructed. Initially, each initial cluster is a separate tree node. At each level, the two clusters with the closest comprehensive distance are merged into a new tree node. The node height of the new tree node is equal to the comprehensive distance at the time of merging. Starting from the root node of the clustering hierarchy tree, traverse from top to bottom. When encountering a node with a height greater than T, stop traversing downward and output all child nodes of the tree node as the final order demand cluster.
[0011] Furthermore, the solution to the preliminary order delivery solution includes: Construct a mathematical model for order delivery and define decision variables , indicating the The first order demand cluster Is the maintenance order assigned to service resource, the value is 0 or 1; define the delivery objective function and constraints; the constraints are the available amount of service resources; ;in, For the The number of orders to be maintained in the order demand cluster, is the total number of available service resources; To be the first The first order demand cluster Maintenance orders are assigned to The cost of each service resource; To be the first The first order demand cluster Maintenance orders are assigned to The delivery time of each service resource; To be the first The first order demand cluster Maintenance orders are assigned to The service quality score of each service resource; and is the weight coefficient; To be the first The first order demand cluster Maintenance orders are assigned to The geographical distance of each service resource; Define the state transfer equation, Indicates that the The front of the order demand cluster The optimal solution for allocating maintenance orders to service resources; The formula of the state transfer equation is: ;in, Indicates that the The front of the order demand cluster The optimal solution for allocating maintenance orders to service resources; Indicates that the The front of the order demand cluster The optimal solution for allocating maintenance orders to service resources; recursively solve the optimal solution through the state transition equation , that is, for each order demand cluster, in the calculation When , enumerate all service resource allocation schemes and replace the one with the smallest price as The value of The allocation plan of service resources at the time of delivery; from First, according to the recorded service resource allocation plan, the specific allocation plan for each order to be maintained is backtracked to solve; the specific allocation plans for each order to be maintained in all order demand clusters are merged to obtain the overall preliminary order delivery plan.
[0012] Furthermore, the method of optimizing the preliminary order delivery plan includes: Build a job assembly line model, treat each maintenance order as a job, and the service resources required by the order as the machines required to perform the job. For each job, determine the type and quantity of machines required for its execution; based on the preliminary order delivery plan, preliminarily schedule the job, generate an initial scheduling plan, and optimize it based on the initial scheduling plan. The goal of the optimization is to shorten the total completion time of the job; Based on the initial scheduling scheme, an initial job scheduling sequence is generated; for each job, its priority value is defined as the remaining execution time of the job, and the jobs are sorted in ascending order according to the priority value; thus, a job queue sorted by priority is formed; Take the job with the highest priority value from the job queue and assign the required machine to the job. If several machines are available, select a machine adaptively. Update the machine status and job queue. Repeat until all jobs are assigned machines. Decode the job sequence sorted by priority value and assigned to machines into the actual scheduling plan, output the machine, start time and completion time assigned to each job, and calculate the total completion time corresponding to the actual scheduling plan; The evaluation threshold interval is preset, and the weighted sum of the total completion time and service resource utilization rate of the job obtained by the actual scheduling plan is calculated; if it is within the evaluation threshold interval, the current actual scheduling plan is determined as the terminal order delivery plan; if it is not within the evaluation threshold interval, the actual scheduling plan is further optimized to obtain the terminal order delivery plan; Ways to further optimize the actual scheduling plan include: Define the optimization function value as the weighted sum of the total completion time of the job and the utilization rate of the service resources; construct the job neighborhood space, and for any two jobs in the job queue, construct a new neighborhood solution, which is obtained by exchanging the positions of the two jobs in the job queue; for each job w, and all positions w' before and after it, construct a new neighborhood solution, and insert the job w into the position w' of the job queue; merge all the new neighborhood solutions generated to form a complete job neighborhood space; Based on the optimization function value, the simulated annealing algorithm is used in the job neighborhood space to optimize the actual scheduling plan and obtain the terminal order delivery plan; The cooling strategy of the simulated annealing algorithm is defined as The cooling formula used in the iteration is: ;in, is the temperature parameter of the next round of simulated annealing algorithm, is a normal number, is the current iteration number, is the difference between the optimization function value of the new neighborhood solution and the current solution, is the temperature parameter of the current round of simulated annealing algorithm; In the simulated annealing algorithm The cooling formula used in the iteration is: ;in, is the exponential decay factor, is the preset maximum number of iterations; if the maximum number of iterations is reached, the current optimal solution is output as the optimized scheduling plan, and the optimized scheduling plan is decoded into the terminal order delivery plan.
[0013] A method for order maintenance management based on machine learning, which is implemented based on the order maintenance management platform based on machine learning, comprises: S1, obtaining comprehensive data of orders to be maintained; matrixing the comprehensive data of the orders to be maintained to obtain corresponding m order matrices; S2, using a clustering algorithm based on m order matrices to group the maintenance orders to obtain several order demand clusters; S3. For each order demand cluster, a dynamic programming algorithm is used to solve and obtain a preliminary order delivery plan, and the preliminary order delivery plan is optimized to obtain a terminal order delivery plan.
[0014] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the order maintenance management method based on machine learning when executing the computer program.
[0015] Technical effects and advantages of the order maintenance management platform, method and electronic device based on machine learning of the present invention: The present invention improves the intelligence level and efficiency of order processing, and realizes in-depth mining and comprehensive analysis of order data; through more accurate order grouping and classification, it can better understand customer demand patterns and provide personalized insurance product service recommendations; the dynamic optimization and adaptive adjustment capabilities of insurance order delivery solutions are significantly enhanced, which not only effectively balances multiple goals such as cost, time and quality, but also can flexibly adjust strategies according to real-time conditions, has stronger robustness and adaptability, and can calmly deal with various uncertainties in insurance business, such as sudden large claims, concentrated compensation caused by natural disasters, etc.; in addition, it significantly improves resource utilization, reduces unnecessary manual intervention, greatly reduces operating costs, and significantly improves scalability and flexibility. It can seamlessly adapt to business needs of different scales and types, providing strong support for the rapid development of enterprises and a smarter, more efficient and more comprehensive order management solution, which not only significantly improves customer satisfaction, but also enhances the market competitiveness of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 A schematic diagram of an order maintenance management platform based on machine learning of the present invention; Figure 2 A schematic diagram of an order maintenance management method based on machine learning of the present invention; Figure 3 It is a schematic diagram of an electronic device of the present invention. DETAILED DESCRIPTION
[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0018] Example 1 See also Figure 1 As shown, the order maintenance management platform based on machine learning described in this embodiment includes: a data acquisition matrix processing module, a clustering grouping module and a comprehensive planning module; each module is connected by wired and / or wireless means to realize data transmission between modules; The data acquisition matrix processing module includes a data acquisition unit and a comprehensive processing unit. The data acquisition unit is used to acquire the comprehensive data of the orders to be maintained; the comprehensive processing unit is used to perform matrix processing on the comprehensive data of the orders to be maintained to acquire corresponding m order matrices, where m is a positive integer greater than 1; The clustering grouping module uses a clustering algorithm based on m order matrices to group maintenance orders and obtain several order demand clusters; The comprehensive planning module includes a preliminary planning unit and a terminal optimization unit. The preliminary planning unit is used to obtain a preliminary order delivery plan for each order demand cluster using a dynamic programming algorithm. The terminal optimization unit is used to optimize the preliminary order delivery plan to obtain a terminal order delivery plan, thereby realizing dynamic adjustment and optimization of orders.
[0019] The comprehensive data of the order to be maintained includes basic order information, customer information and insurance information; Basic order information includes order number, order time, order type (insurance, claim settlement or renewal) and order status (accepted, in progress or completed); Customer information includes customer type (company or individual) and customer name (company name or individual name); Insurance information includes insurance type (such as industrial equipment insurance, vehicle insurance, etc.), details of the insured object (such as industrial equipment model, vehicle information, etc.), insurance amount, premium and insurance period.
[0020] Furthermore, the order matrix is obtained by: The comprehensive data of the maintenance orders is cleaned to eliminate missing and erroneous data, and the data with categorical attributes (such as order type, service type, etc.) is encoded and converted into numerical data; continuous numerical data is normalized to make their numerical ranges consistent to obtain standard order data; standard order data includes several order data fields (such as insurance amount, insurance type, etc.).
[0021] Construct an order feature matrix X, with each order to be maintained as a row of the order feature matrix, and the order data fields corresponding to the order to be maintained as columns of the order feature matrix; the element X(i, j) in the order feature matrix X corresponds to the value of the jth order data field of the i-th order to be maintained.
[0022] A collaborative filtering matrix is constructed based on the order feature matrix X. Specifically, different types of entities in the standard order data are extracted, such as customers, insurance types, orders, etc.
[0023] A heterogeneous information network is constructed based on different types of entities, and each type of entity is used as a type of node in the heterogeneous information network; edges between nodes are constructed based on the relationship between entities (such as a customer purchases a certain insurance product, an order corresponds to a certain insurance product, etc.); the heterogeneous information network contains several types of nodes and edges.
[0024] Define the order network path. The order network path is a semantic path pattern in a heterogeneous information network, which describes the compound relationship between nodes of different types; for example, customer-purchase-insurance product-corresponding-order, which means that the customer has generated an order to be maintained by purchasing an insurance product.
[0025] For each order network path, any two nodes are recorded as node pairs, and the comprehensive similarity score between any node pairs along the order network path is calculated; the collaborative filtering matrix is constructed , whose elements Indicates The node and The comprehensive similarity score between nodes.
[0026] Several random walks are performed in heterogeneous information networks to generate a series of node sequences. Each random walk starts from a random node and randomly selects the next node according to the connection relationship of the edge until the preset maximum step size is reached. The generated node sequence captures the proximity relationship and structural context information between nodes in heterogeneous information networks.
[0027] Define a context window whose size is , used to determine the context scope of the central node; for each node in the node sequence The set of context nodes that define it ;in, The index of the node.
[0028] The node learning objective is defined as maximizing the conditional probability of the node based on the context node; that is, for each node and its context node ; Construct objective function ; ;in, For Node The index of the context node; is a randomly sampled negative node (other nodes unrelated to the current node); is the index of the negative example node; For Node For context nodes The weights of (the weights can be automatically learned through a feedforward neural network or RNN, which takes the features of the node and the context node as input and outputs the attention weights of each context node); For the node Based on context node The conditional probability of For the node Based on negative example nodes The conditional probability of is a sigmoid function, which is used to map the conditional probability to the range of [0, 1]. By minimizing the conditional probability of the node based on the negative node, the relationship between the central node and the context node can be better learned.
[0029] The node The node vector is represented as , the context node The node vector is represented as ; then the conditional probability ;in, For Node The node vector of The index of all nodes to traverse.
[0030] Optimize the objective function using gradient descent or its variants (such as SGD, Adam, etc.) , learn the node vector and ; Each iteration updates the node vector of a node and its context node.
[0031] The maximum number of training times is preset, and after the maximum number of training iterations, the node vectors corresponding to each node are obtained; these node vector representations capture the structural information and semantic information of the nodes in the network.
[0032] For any node pair, calculate the cosine similarity between the two corresponding node vectors; as the similarity score between the two; for all node pairs in the heterogeneous information network, calculate their similarity scores and construct a similarity matrix , whose elements Indicates The node and The similarity score between nodes is obtained; the similarity matrix corresponding to each order network path reflects the similarity of the node pairs under the path; different order network paths reflect different semantic information of the order to be maintained.
[0033] For nodes, extract their eigenvalues ;in, For the The number of edges that actually exist between the neighbors of a node, For the The degree of the node The number of nodes connected to the node; the eigenvalue Converted into vector representation, we get The feature vector corresponding to the node ; Similarly Extract the eigenvalue of each node and convert it into a vector representation to get the The feature vector corresponding to the node .
[0034] Define an attention model, and As input, the output is the attention weight of each order network path; the attention model is a feedforward neural network or a recurrent neural network.
[0035] For order network path, and its attention weight is recorded as , and the sum of the attention weights of all order network paths is 1; for the Order network paths, computing node pairs Similarity score under the corresponding order network path ; Weighted sum of similarity scores of all order network paths to get node pair The weight of the weighted sum is the attention weight output by the attention model.
[0036] The order feature matrix X and the collaborative filtering matrix Merge to obtain m order matrices; each order matrix contains the feature information of the order to be maintained and the similarity information with other orders.
[0037] Ways to group pending maintenance orders include: Initialize clustering parameters, including clustering threshold T, leaf node constant factor number and cluster center; define an initialized clustering feature tree, scan the order matrix, and insert each order to be maintained as a clustering feature node into the clustering feature tree; if the current order to be maintained cannot be inserted into the clustering feature tree during scanning, generate a new clustering feature node; repeat until all orders to be maintained are inserted into the clustering feature tree to obtain a preliminary clustering feature tree; ensure that the number of child nodes of each clustering feature node does not exceed the constant factor number of leaf nodes; the preliminary clustering feature tree contains several leaf nodes.
[0038] Scan all leaf nodes of the preliminary clustering feature tree, cluster the clustering feature entries in the leaf nodes, and generate initial clusters. The leaf nodes are used to store actual data objects (orders to be maintained) and corresponding clustering feature information. The basic units of these storages are clustering feature entries; clustering feature entries include the total number of data objects in the clustering feature entries, the linear sum of all data objects in each dimension (the sum of all fields), and the square sum of all data objects in each dimension (the sum of the squares of all fields).
[0039] The clustering methods for clustering feature entries in leaf nodes include: Traverse each leaf node of the preliminary clustering feature tree, extract all clustering feature items in the leaf node, and store them in a list. Initially, select K clustering feature items from the list as the center of the cluster; K is a positive integer greater than 1; for each clustering feature item in the list, calculate its distance to the center of the K clusters (Manhattan distance).
[0040] For each cluster feature item, assign it to the initial cluster that is closest to it, then the K initial clusters are filled, calculate the centroid of all cluster feature items in the cluster at this time, and use the centroid as the new center of the cluster; repeat until the center of the cluster no longer changes, and generate the initial cluster.
[0041] According to the clustering threshold T, the initial clusters are merged to obtain the final order demand cluster. Specifically, for any two initial clusters and , calculate the comprehensive distance between them ; ;in, for and The mean of the comprehensive similarity scores of the data objects in ; for and The distance between the farthest cluster feature entries in ; for and The mean distance between all cluster feature entries in , for and The distance between the centers of
[0042] According to the calculated comprehensive distance, a clustering hierarchical tree is constructed. Initially, each initial cluster is a separate tree node. At each level, the two clusters with the closest comprehensive distance are merged into a new tree node. The node height of the new tree node is equal to the comprehensive distance at the time of merging.
[0043] Starting from the root node of the clustering hierarchy tree, traverse from top to bottom. When encountering a node height (comprehensive distance) greater than T, stop traversing downward and output all child nodes of the tree node as the final order demand cluster; For each output order demand cluster, the cluster feature items in all initial clusters within the order demand cluster are merged, and the sum of the merged cluster feature vectors is calculated to obtain a comprehensive cluster feature item representing the order demand cluster.
[0044] Furthermore, the solution to the preliminary order delivery plan includes: Construct a mathematical model for order delivery, specifically, define the decision variables , indicating the The first order demand cluster Is the maintenance order assigned to service resource, the value is 0 or 1; define the delivery objective function and constraints; the constraints are the available amount of service resources.
[0045] ;in, For the The number of orders to be maintained in the order demand cluster, is the total number of available service resources; To be the first The first order demand cluster Maintenance orders are assigned to The cost of each service resource (different cost calculation methods are set according to the situation); To be the first The first order demand cluster Maintenance orders are assigned to The delivery time of each service resource; To be the first The first order demand cluster Maintenance orders are assigned to The service quality score of each service resource (obtained through a service quality scoring mechanism designed based on skill level, historical scores, and order complexity); and is the weight coefficient, which is used to balance the importance of the three factors of cost, time and quality; To be the first The first order demand cluster Maintenance orders are assigned to The geographical distance of the service resources.
[0046] Define the state transfer equation, Indicates that the The front of the order demand cluster The optimal solution for allocating maintenance orders to service resources.
[0047] The formula of the state transfer equation is: ;in, Indicates that the The front of the order demand cluster The optimal solution for allocating maintenance orders to service resources; Indicates that the The front of the order demand cluster The optimal solution for allocating maintenance orders to service resources; recursively solve the optimal solution through the state transition equation ,in It is The number of orders to be maintained for an order demand cluster; specifically, initialize , indicating that the cost is 0 when no maintenance order is allocated.
[0048] For each order demand cluster, calculate When , enumerate all service resource allocation schemes and replace the one with the smallest price as The value of The allocation plan of service resources at the time of from First, according to the recorded service resource allocation plan, the specific allocation plan for each order to be maintained is backtracked to solve; the specific allocation plans for each order to be maintained in all order demand clusters are merged to obtain the overall preliminary order delivery plan.
[0049] Further, the methods for optimizing the preliminary order delivery plan include: Build a job pipeline model, treat each maintenance order as a job, and the service resources required by the order as the machines required to perform the job. For each job (maintenance order), determine the type and quantity of machines required for its execution; model the priority and dependency relationships between jobs. For example, there may be priority constraints between multiple orders from the same customer; service resources are used to process and deliver various resources required for orders, including human resources and other resources (such as customer service personnel, insurance claims specialists, insurance underwriters, technical support personnel, etc.).
[0050] Based on the preliminary order delivery plan, the jobs (orders) are preliminarily scheduled to generate an initial scheduling plan, which is then optimized based on the initial scheduling plan. The goal of the optimization is defined as shortening the total completion time of the job.
[0051] Based on the initial scheduling scheme, an initial job scheduling sequence is generated, and it is ensured that the initial job scheduling sequence satisfies the job constraints, which include machine availability and job priority.
[0052] It needs to be explained that machines refer to various service resources required to process orders, including human resources, other resources, etc. Machine availability indicates whether these resources can be allocated for use within a certain period of time; job priority indicates the processing order or importance between different orders. The two constraints work together to ensure that orders are only allocated to available resources, the processing order of orders meets business needs and customer expectations, and resources are used efficiently, while meeting various business rules.
[0053] For each job, define its priority value as the remaining execution time of the job, and sort the jobs in ascending order (from small to large) according to the priority value; thus forming a job queue sorted by priority.
[0054] Take the job with the highest priority value from the job queue and assign the required machine to the job. If several machines are available, select the machine adaptively. Update the machine status (available time) and the job queue. That is, update the status of the machine assigned to the job and remove the job with the assigned machine from the job queue. Repeat until all jobs are assigned machines.
[0055] Decode the job sequence sorted by priority value and assigned to machines into the actual scheduling plan, output the machine, start time and completion time assigned to each job, and calculate the total completion time corresponding to the actual scheduling plan.
[0056] The adaptive machine selection methods include: Maintain a list of available machines and store the machines that can currently be assigned to the job. For each machine, record its available start time and end time. For each machine in the available machine list, calculate its remaining available time = machine available end time - current time; obtain the execution time of the current job; for each machine, add the job execution time to the machine's available start time to get the completion time of the job on the machine, and among all machines, find the machine that makes the job complete the earliest; if several machines have the same completion time, randomly select one.
[0057] An evaluation threshold interval is preset, and the weighted sum of the total completion time and service resource utilization of the job obtained by the actual scheduling plan is calculated; if it is within the evaluation threshold interval, the current actual scheduling plan is determined as the terminal order delivery plan; if it is not within the evaluation threshold interval, the actual scheduling plan is further optimized to obtain the terminal order delivery plan.
[0058] Ways to further optimize the actual scheduling plan include: The optimization function value is defined as the weighted sum of the total completion time of the job and the service resource utilization rate; the job neighborhood space is constructed. Specifically, for any two jobs in the job queue, a new neighborhood solution is constructed. The new neighborhood solution is obtained by exchanging the positions of two jobs in the job queue. For example, the original sequence is [1, 2, 3, 4, 5]. By exchanging jobs 2 and 4, a new neighborhood solution [1, 4, 3, 2, 5] is obtained.
[0059] For each job w, and all positions w' before and after it, a new neighborhood solution is constructed, and job w is inserted into position w' of the job queue; all the new neighborhood solutions generated are merged to form a complete job neighborhood space.
[0060] Based on the optimization function value, the simulated annealing algorithm is used to optimize the actual scheduling plan in the job neighborhood space to obtain the terminal order delivery plan.
[0061] The cooling strategy of the simulated annealing algorithm is defined as The cooling formula used in the iteration is: ;in, is the temperature parameter of the next round of simulated annealing algorithm, is a positive constant, is the current iteration number, is the difference between the optimization function value of the new neighborhood solution and the current solution, The temperature parameter for the current round of simulated annealing algorithm.
[0062] In the simulated annealing algorithm The cooling formula used in the iteration is: ;in, is an exponential decay factor, and , is the preset maximum number of iterations; if the maximum number of iterations is reached, the current optimal solution is output as the optimized scheduling plan, and the optimized scheduling plan is decoded into the terminal order delivery plan.
[0063] It should be noted that in the above process, uncertain factors such as machine failure and dynamic changes in order arrival need to be considered, and the order delivery plan should be adjusted dynamically at this time; at the same time, machine learning is combined to model job characteristics, machine status, etc. to improve the intelligence level of scheduling decisions.
[0064] This embodiment improves the intelligence level and efficiency of order processing, and realizes in-depth mining and comprehensive analysis of order data; through more accurate order grouping and classification, it can better understand customer demand patterns and provide personalized insurance product service recommendations; the dynamic optimization and adaptive adjustment capabilities of insurance order delivery solutions are significantly enhanced, which not only effectively balances multiple goals such as cost, time and quality, but also can flexibly adjust strategies according to real-time conditions, has stronger robustness and adaptability, and can calmly deal with various uncertainties in insurance business, such as sudden large claims, concentrated compensation caused by natural disasters, etc.; in addition, it significantly improves resource utilization, reduces unnecessary manual intervention, greatly reduces operating costs, and significantly improves scalability and flexibility. It can seamlessly adapt to business needs of different scales and types, providing strong support for the rapid development of enterprises and a smarter, more efficient and more comprehensive order management solution, which not only significantly improves customer satisfaction, but also enhances the market competitiveness of enterprises.
[0065] Example 2 See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides an order maintenance management method based on machine learning, including: S1. Obtain comprehensive data of the orders to be maintained; perform matrix processing on the comprehensive data of the orders to be maintained to obtain corresponding m order matrices; S2, using a clustering algorithm based on m order matrices to group the maintenance orders to obtain several order demand clusters; S3. For each order demand cluster, a dynamic programming algorithm is used to solve and obtain a preliminary order delivery plan, and the preliminary order delivery plan is optimized to obtain a terminal order delivery plan, thereby realizing dynamic adjustment and optimization of orders.
[0066] Example 3 See also Figure 3 As shown, this embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the above-mentioned order maintenance management method based on machine learning is implemented.
[0067] Since the electronic device introduced in this embodiment is an electronic device used to implement an order maintenance management method based on machine learning in the embodiment of the present application, based on the order maintenance management method based on machine learning introduced in the embodiment of the present application, the technical personnel of the field can understand the specific implementation of the electronic device of the present embodiment and its various variations, so how the electronic device implements the method in the embodiment of the present application is not described in detail here. As long as the technical personnel of the field implement the electronic device used in the order maintenance management method based on machine learning in the embodiment of the present application, it belongs to the scope of protection of this application.
[0068] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.
[0069] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.
Claims
1. An order maintenance management platform based on machine learning, characterized in that: include: Data acquisition matrix processing module, clustering grouping module and comprehensive planning module; each module is connected by wired and / or wireless means; The data acquisition matrix processing module includes a data acquisition unit and a comprehensive processing unit. The data acquisition unit is used to acquire the comprehensive data of the order to be maintained; The comprehensive processing unit is used to perform matrix processing on the comprehensive data of the orders to be maintained to obtain corresponding m order matrices; The clustering grouping module uses a clustering algorithm based on m order matrices to group maintenance orders and obtain several order demand clusters; The comprehensive planning module includes a preliminary planning unit and a terminal optimization unit. The preliminary planning unit is used to obtain a preliminary order delivery plan for each order demand cluster using a dynamic programming algorithm. The terminal optimization unit is used to optimize the preliminary order delivery plan to obtain a terminal order delivery plan.
2. The order maintenance management platform based on machine learning according to claim 1, characterized in that: The comprehensive data of the order to be maintained includes basic order information, customer information and insurance information; the basic order information includes order number, order time, order type and order status; customer information includes customer type and customer name; insurance information includes insurance type, insured object details, insurance amount, premium and insurance period.
3. The order maintenance management platform based on machine learning according to claim 2, characterized in that: The order matrix is obtained by: Clean the comprehensive data of the maintenance orders, encode the data with categorical attributes and convert them into numerical data; normalize the continuous numerical data to obtain standard order data; the standard order data includes several order data fields; Construct an order feature matrix X, with each order to be maintained as a row of the order feature matrix, and the order data field corresponding to the order to be maintained as a column of the order feature matrix; the element X(i, j) in the order feature matrix X corresponds to the value of the jth order data field of the i-th order to be maintained; construct a collaborative filtering matrix based on the order feature matrix X , the order feature matrix X and the collaborative filtering matrix Merge and obtain m order matrices.
4. The order maintenance management platform based on machine learning according to claim 3, characterized in that: Constructing a collaborative filtering matrix The methods include: Extract different types of entities from standard order data; construct heterogeneous information networks based on different types of entities, and use each type of entity as a type of node in the heterogeneous information network; construct edges between nodes based on the relationship between entities; heterogeneous information networks contain several types of nodes and edges; Define the order network path. For each order network path, record any two nodes as a node pair. Calculate the comprehensive similarity score between any node pairs along the order network path. Construct a collaborative filtering matrix , whose elements Indicates The node and The comprehensive similarity score between nodes.
5. The order maintenance management platform based on machine learning according to claim 4, characterized in that: The method for obtaining the comprehensive similarity score includes: Perform several random walks in the heterogeneous information network to generate a series of node sequences. Each random walk starts from a random node and randomly selects the next node according to the connection relationship of the edge until the preset maximum step length is reached; define a context window with a size of , for each node in the node sequence The set of context nodes that define it ;in, is the index of the node; For each node and its context node ; Construct objective function ; ;in, For Node The index of the context node; is a randomly sampled negative example node; is the index of the negative example node; For Node For context nodes The weight of For the node Based on context node The conditional probability of For the node Based on negative example nodes The conditional probability of is the sigmoid function; The node The node vector is represented as , the context node The node vector is represented as ; then the conditional probability ;in, For Node The node vector of To traverse the index of all nodes; Optimize the objective function using gradient descent or its variants , learn the node vector and ; Each iteration updates the node vector of a node and its context node; The maximum number of training times is preset. After the maximum number of training iterations, the node vector corresponding to each node is obtained. For any node pair, the cosine similarity between the two corresponding node vectors is calculated as the similarity score between the two. For all node pairs in the heterogeneous information network, their similarity scores are calculated and a similarity matrix is constructed. , whose elements Indicates The node and The similarity score between nodes; the similarity matrix corresponding to each order network path obtained; For nodes, extract their eigenvalues ;in, For the The number of edges that actually exist between the neighbors of a node, For the The number of nodes connected to the node; the eigenvalue Converted into vector representation, we get The feature vector corresponding to the node ; Similarly Extract the eigenvalue of each node and convert it into a vector representation to get the The feature vector corresponding to the node ; Define an attention model, and As input, the output is the attention weight of each order network path; the attention model is a feedforward neural network or a recurrent neural network; For order network path, and its attention weight is recorded as , and the sum of the attention weights of all order network paths is 1; for the Order network paths, computing node pairs Similarity score under the corresponding order network path ; Weighted sum of similarity scores of all order network paths to get node pair The weight of the weighted sum is the attention weight output by the attention model.
6. The order maintenance management platform based on machine learning according to claim 5, characterized in that: The method of grouping the maintenance orders includes: Initialize clustering parameters, including clustering threshold T, leaf node constant factor number and cluster center; define an initialized clustering feature tree, scan the order matrix, and insert each order to be maintained as a clustering feature node into the clustering feature tree; if the current order to be maintained cannot be inserted into the clustering feature tree during scanning, generate a new clustering feature node; repeat until all orders to be maintained are inserted into the clustering feature tree to obtain a preliminary clustering feature tree; and ensure that the number of child nodes of each clustering feature node does not exceed the number of leaf node constant factors; the preliminary clustering feature tree contains several leaf nodes; Scan all leaf nodes of the preliminary clustering feature tree, cluster the clustering feature entries in the leaf nodes, and generate initial clusters; the clustering feature entries include the total number of data objects in the clustering feature entries, the linear sum of all data objects in each dimension, and the square sum of all data objects in each dimension; The clustering methods for clustering feature entries in leaf nodes include: Traverse each leaf node of the preliminary clustering feature tree, extract all clustering feature items in the leaf node, and store them in a list. Initially, select K clustering feature items from the list as the center of the cluster; K is a positive integer greater than 1; for each clustering feature item in the list, calculate its distance to the center of the K clusters; For each cluster feature item, assign it to the initial cluster that is closest to it, then the K initial clusters are filled, calculate the centroid of all cluster feature items in the cluster at this time, and use the centroid as the new center of the cluster; repeat until the center of the cluster no longer changes, and generate the initial cluster; According to the clustering threshold T, the initial clusters are merged to obtain the final order demand cluster. For any two initial clusters and , calculate the comprehensive distance between them ; ;in, for and The mean of the comprehensive similarity scores of the data objects in ; for and The distance between the farthest cluster feature entries in ; for and The mean distance between all cluster feature entries in , for and The distance between the centers of According to the calculated comprehensive distance, a clustering hierarchical tree is constructed. Initially, each initial cluster is a separate tree node. At each level, the two clusters with the closest comprehensive distance are merged into a new tree node. The node height of the new tree node is equal to the comprehensive distance at the time of merging. Starting from the root node of the clustering hierarchy tree, traverse from top to bottom. When encountering a node with a height greater than T, stop traversing downward and output all child nodes of the tree node as the final order demand cluster.
7. The order maintenance management platform based on machine learning according to claim 6, characterized in that: The solution method of the preliminary order delivery plan includes: Construct a mathematical model for order delivery and define decision variables , indicating the The first order demand cluster Is the maintenance order assigned to service resource, the value is 0 or 1; define the delivery objective function and constraints; the constraints are the available amount of service resources; ;in, For the The number of orders to be maintained in the order demand cluster, is the total number of available service resources; To be the first The first order demand cluster Maintenance orders are assigned to The cost of each service resource; To be the first The first order demand cluster Maintenance orders are assigned to The delivery time of each service resource; To be the first The first order demand cluster Maintenance orders are assigned to The service quality score of each service resource; and is the weight coefficient; To be the first The first order demand cluster Maintenance orders are assigned to The geographical distance of each service resource; Define the state transfer equation, Indicates that the The front of the order demand cluster The optimal solution for allocating maintenance orders to service resources; The formula of the state transfer equation is: ;in, Indicates that the The front of the order demand cluster The optimal solution for allocating maintenance orders to service resources; Indicates that the The front of the order demand cluster The optimal solution for allocating maintenance orders to service resources; recursively solve the optimal solution through the state transition equation , that is, for each order demand cluster, in the calculation When , enumerate all service resource allocation schemes and replace the one with the smallest price as The value of The allocation plan of service resources at the time of from First, according to the recorded service resource allocation plan, the specific allocation plan for each order to be maintained is backtracked to solve; the specific allocation plans for each order to be maintained in all order demand clusters are merged to obtain the overall preliminary order delivery plan.
8. The order maintenance management platform based on machine learning according to claim 7, characterized in that: The method of optimizing the preliminary order delivery plan includes: Build a job assembly line model, treat each maintenance order as a job, and the service resources required by the order as the machines required to perform the job. For each job, determine the type and quantity of machines required for its execution; based on the preliminary order delivery plan, preliminarily schedule the job, generate an initial scheduling plan, and optimize it based on the initial scheduling plan. The goal of the optimization is to shorten the total completion time of the job; Based on the initial scheduling scheme, an initial job scheduling sequence is generated; for each job, its priority value is defined as the remaining execution time of the job, and the jobs are sorted in ascending order according to the priority value; thus, a job queue sorted by priority is formed; Take the job with the highest priority value from the job queue and assign the required machine to the job. If several machines are available, select a machine adaptively. Update the machine status and job queue. Repeat until all jobs are assigned machines. Decode the job sequence sorted by priority value and assigned to machines into the actual scheduling plan, output the machine, start time and completion time assigned to each job, and calculate the total completion time corresponding to the actual scheduling plan; The evaluation threshold interval is preset, and the weighted sum of the total completion time and service resource utilization rate of the job obtained by the actual scheduling plan is calculated; if it is within the evaluation threshold interval, the current actual scheduling plan is determined as the terminal order delivery plan; if it is not within the evaluation threshold interval, the actual scheduling plan is further optimized to obtain the terminal order delivery plan; Ways to further optimize the actual scheduling plan include: Define the optimization function value as the weighted sum of the total completion time of the job and the utilization rate of the service resources; construct the job neighborhood space, and for any two jobs in the job queue, construct a new neighborhood solution, which is obtained by exchanging the positions of the two jobs in the job queue; for each job w, and all positions w' before and after it, construct a new neighborhood solution, and insert the job w into the position w' of the job queue; merge all the new neighborhood solutions generated to form a complete job neighborhood space; Based on the optimization function value, the simulated annealing algorithm is used in the job neighborhood space to optimize the actual scheduling plan and obtain the terminal order delivery plan; The cooling strategy of the simulated annealing algorithm is defined as The cooling formula used in the iteration is: ;in, is the temperature parameter of the next round of simulated annealing algorithm, is a normal number, is the current iteration number, is the difference between the optimization function value of the new neighborhood solution and the current solution, is the temperature parameter of the current round of simulated annealing algorithm; In the simulated annealing algorithm The cooling formula used in the iteration is: ;in, is the exponential decay factor, is the preset maximum number of iterations; if the maximum number of iterations is reached, the current optimal solution is output as the optimized scheduling plan, and the optimized scheduling plan is decoded into the terminal order delivery plan.
9. A method for order maintenance management based on machine learning, which is implemented based on an order maintenance management platform based on machine learning according to any one of claims 1 to 8, characterized in that: include: S1. Obtain comprehensive data of orders to be maintained; Matrix the comprehensive data of the orders to be maintained to obtain the corresponding m order matrices; S2, using a clustering algorithm based on the m order matrices to group the maintenance orders to obtain several order demand clusters; S3. For each order demand cluster, a dynamic programming algorithm is used to solve and obtain a preliminary order delivery plan, and the preliminary order delivery plan is optimized to obtain a terminal order delivery plan.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the order maintenance management method based on machine learning described in claim 9 is implemented.
Citation Information
Patent Citations
Intelligent disposition and process monitoring system and method based on cloud management platform
CN105791417A
Large-scale order form matching method in community commerce cloud
CN103489108A
Multidimensional individualized recommendation method in heterogeneous network
CN106354862A
Real-time receiving and delivering integrated path planning system
CN114511126A
Electronic product CTO order clustering method based on improved vector space model
CN115293253A
Cited By
Member order processing method and device, equipment and storage medium
CN119295188A
Member order processing method, device, equipment and storage medium
CN119295188B
Internet medical platform asynchronous order docking processing method
CN120088042A
