An intelligent management system for industrial product sales data

By constructing a product knowledge graph and a graph neural network to aggregate product status information, combining factor decomposition machine to analyze sales personnel capabilities, and using a dual-tower fit calculation network to generate a sales path allocation scheme, the problem of path mismatch caused by product status changes and personnel turnover in industrial product sales management is solved, thereby improving the sales conversion rate.

CN123118166APending Publication Date: 2026-11-17FUJIAN YUANLI PUMP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611620160.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing industrial product sales management systems cannot dynamically perceive changes in product status and the capabilities of sales personnel, resulting in mismatched sales path configurations, especially during product lifecycle transitions and personnel changes, making it difficult to achieve accurate resource allocation.

Method used

We construct a product knowledge graph and aggregate product status information using graph neural networks. We combine factor decomposition machines to analyze sales personnel capabilities, generate sales path allocation schemes through a dual-tower fit calculation network, and optimize decisions through a feedback learning module.

Benefits of technology

It enables dynamic matching between product status and sales personnel capabilities, generates a fit score, solves the problem of sales path mismatch caused by changes in product life cycle and personnel turnover, and improves the sales conversion rate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118166A_ABST
    Figure CN123118166A_ABST
Patent Text Reader

Abstract

This invention relates to the field of product sales management technology, specifically to an intelligent management system for industrial product sales data. The system includes a product status monitoring module, a product knowledge graph construction module, a sales personnel capability analysis module, an allocation decision module, and a feedback learning module. This invention encodes the product lifecycle stages and their associated relationships (such as accessories, replacements, and upgrades) into a unified product status embedding vector, solving the problem that existing systems cannot perceive the mutual influence between products. By using a factor decomposition machine to extract factors related to sales personnel's new product adaptability, accessory product sales capability, and legacy order handling capability from historical sales records, a quantitative characterization of segmented sales capabilities is achieved. By generating adaptability scores and outputting sales path allocation schemes, products at different lifecycle stages can be assigned to the most suitable sales personnel.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of product sales management technology, and more specifically, to an intelligent management system for industrial product sales data. Background Technology

[0002] In the process of industrial product sales management, the diversity of product types and the differences in individual sales personnel capabilities have long existed. The superposition of two dynamic factors, namely product life cycle evolution and personnel turnover, further increases the complexity of precise allocation of sales resources.

[0003] On the one hand, industrial products generally go through stages such as introduction, growth, maturity, decline, and even discontinuation, and may also generate ancillary products and upgraded substitutes during the sales process. The requirements for sales personnel differ significantly depending on the product's stage of life: the introduction stage requires sales personnel to have the ability to develop new markets, ancillary products require sales personnel to be skilled in cross-selling, and discontinued products require sales personnel to be able to clear existing inventory and guide customers to switch to alternative products. When a product undergoes a stage transition, the original sales personnel configuration plan may no longer be applicable, but the existing system lacks a holistic understanding of changes in product status and struggles to proactively adjust sales paths.

[0004] On the other hand, sales staff turnover or personnel changes can lead to orders being reassigned to other current sales staff. The new sales staff then have to deal with client relationships not developed by them, partially completed sales processes, and communication records left by the previous sales staff. Completing these legacy orders presents a completely different challenge than regular sales, demanding exceptional firefighting capabilities. Current sales management systems typically distribute legacy orders based solely on the sales staff's current availability, without considering individual differences in their ability to handle others' legacy orders, resulting in a significantly lower-than-expected conversion rate after assignment.

[0005] Both of these scenarios point to a deeper problem: the lack of a dynamic matching mechanism between the product's sales status and the sales personnel's capabilities. The assignment of sales personnel and the allocation of sales resources at different sales stages often rely on manual experience or simple rules, failing to adapt to changes in product status and personnel. Existing automated allocation solutions are mostly based on static rules or single performance rankings, unable to analyze sales personnel's capabilities in specific dimensions such as "adapting to new products," "selling ancillary products," and "taking over legacy orders," nor can they model the relationships between products and the impact of their lifecycle transitions on the sales path. Therefore, how to construct an intelligent management solution that can dynamically perceive product status, deeply characterize the multidimensional capabilities of sales personnel, and perform targeted allocation at different stages of product sales has become an urgent technical problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent management system for industrial product sales data, thereby solving the problems mentioned in the background art.

[0007] To achieve the above objectives, the intelligent management system for industrial product sales data includes a product status monitoring module, a product knowledge graph construction module, a sales personnel capability analysis module, an allocation decision module, and a feedback learning module.

[0008] The product status monitoring module is used to obtain the current life cycle stage of each product and the relationship between products;

[0009] The product knowledge graph construction module constructs a product knowledge graph based on the relationships between the products. The product knowledge graph uses products as nodes and the relationships as edges. It also uses a graph neural network to aggregate neighbor information for each product node and generate a product state embedding vector for that product node.

[0010] The sales personnel capability analysis module extracts a multi-dimensional implicit representation of each salesperson's capabilities from historical sales records;

[0011] The allocation decision module is used to receive the product state embedding vector of the target product involved in the sales task to be allocated and the multi-dimensional implicit representation of the candidate salesperson's capabilities. It uses the fit calculation network to generate a fit score for each candidate salesperson for the target product and generates a sales path allocation scheme based on the fit score. The sales path allocation scheme assigns the sales task of the target product to the selected salesperson.

[0012] The feedback learning module is used to obtain the actual sales results after the sales path allocation scheme is executed, calculate the reward value based on the actual sales results, and use reinforcement learning strategy to update the network parameters of the fitness calculation network and the multi-dimensional implicit representation of the sales personnel capability analysis module.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] This intelligent management system for industrial product sales data constructs a product knowledge graph and aggregates neighboring product status information using a graph neural network. It encodes the product lifecycle stages and related relationships such as accessories, replacements, and upgrades into a unified product status embedding vector, solving the problem of existing systems being unable to perceive the mutual influence between products. By using a factor decomposition machine to extract factors related to sales personnel's new product adaptability, accessory product sales ability, and legacy order handling ability from historical sales records, it achieves a quantitative characterization of segmented sales capabilities. Based on this, a dual-tower fit calculation network dynamically matches product status with the multi-dimensional capabilities of sales personnel in the latent space, generating fit scores and outputting sales path allocation schemes. This ensures that products at different lifecycle stages are assigned to the most suitable sales personnel. The feedback learning module calculates reward values ​​based on actual sales results and continuously optimizes allocation decisions and capability representations using reinforcement learning strategies, forming a closed-loop adaptive evolution that fundamentally solves the problem of sales path mismatch caused by product status changes and personnel turnover. Attached Figure Description

[0015] Figure 1 This is a block diagram of the overall system structure of the present invention;

[0016] Figure 2 This is an interface diagram of the product status monitoring system of the present invention;

[0017] Figure 3 This is a diagram of the product knowledge graph interface of the present invention;

[0018] Figure 4 This is a vector interface diagram embedded in the product status of the present invention. Detailed Implementation

[0019] The technical solutions in 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.

[0020] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0021] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] Please see Figure 1 As shown, an intelligent management system for industrial product sales data is provided, including a product status monitoring module, a product knowledge graph construction module, a sales personnel capability analysis module, an allocation decision module, and a feedback learning module;

[0023] The product status monitoring module is used to obtain the current life cycle stage of each product and the relationships between products;

[0024] The product knowledge graph construction module builds a product knowledge graph based on the relationships between products. The product knowledge graph uses products as nodes and relationships as edges. It also uses a graph neural network to aggregate neighbor information for each product node and generate a product state embedding vector for that product node.

[0025] The sales personnel capability analysis module extracts a multi-dimensional implicit representation of each salesperson's capabilities from historical sales records. This multi-dimensional implicit representation includes at least a new product adaptability factor, a supplementary product sales capability factor, and a legacy order takeover capability factor.

[0026] The allocation decision module receives the product state embedding vector of the target product involved in the sales task to be allocated and the multi-dimensional implicit representation of the candidate salesperson's capabilities. It uses the fit calculation network to generate a fit score for each candidate salesperson for the target product and generates a sales path allocation scheme based on the fit score. The sales path allocation scheme assigns the sales task of the target product to the selected salesperson.

[0027] The feedback learning module is used to obtain the actual sales results after the sales path allocation plan is implemented, calculate the reward value based on the actual sales results, and use reinforcement learning strategies to update the network parameters of the fitness calculation network and the multi-dimensional implicit representation of capabilities in the sales personnel capability analysis module.

[0028] The specific implementation method is as follows:

[0029] First, since the sales status of various industrial products will differ at different stages, and the adaptability of sales personnel will also differ, in order to facilitate the subsequent phased sales adaptation of industrial products, this solution uses a product status monitoring module to obtain the current life cycle stage of each product and the relationship between products.

[0030] The corresponding life cycle stages include the introduction stage, growth stage, maturity stage, decline stage, and shutdown stage, and the related relationships include subordinate relationships, substitution relationships, and upgrading relationships.

[0031] During the data collection process, at the data acquisition level, the product status monitoring module uses a standard data interface with the enterprise resource planning system and product data management system to periodically retrieve product master data, sales records, product configuration lists, and product status change logs at fixed update cycles (e.g., daily or weekly). The retrieved raw data is then processed through noise reduction, deduplication, and standardization to form structured product information that can be directly used by subsequent modules.

[0032] Meanwhile, in terms of lifecycle stage identification, each product is labeled with its current stage using preset judgment rules. The specific judgment rules are as follows:

[0033] When a product's sales have been continuously increasing within a preset observation window since its launch and its market penetration rate is below the first threshold, it is determined to be in the introduction phase; when the sales growth rate exceeds the preset growth rate threshold and there is no continuous decline, it is determined to be in the growth phase.

[0034] When sales volume tends to stabilize and the fluctuation range is less than the stability threshold, it is judged to be in the maturity stage.

[0035] When sales show a downward trend for several consecutive periods, it is determined to be a period of decline;

[0036] When a product is marked as ceasing production, it is directly determined to be in a production stoppage period.

[0037] In terms of relationship extraction, the product status monitoring module identifies and outputs three types of directed business relationships based on the product configuration list and product substitution relationship file. Among them:

[0038] A subordinate relationship indicates that a certain product is a necessary or optional accessory to another main product;

[0039] A substitution relationship indicates that a new product functionally covers and completely replaces an old product.

[0040] An upgrade relationship indicates that the product is a high-end iteration of the same series.

[0041] These relationships constitute the edge information of the product knowledge graph, which is the factual basis for subsequent neighbor aggregation in the graph neural network.

[0042] Finally, the product status monitoring module outputs a full product snapshot once per update cycle. The snapshot includes each product's unique identifier, lifecycle stage tag, and the identifiers and relationship types of other products associated with that product. This snapshot data is directly sent to the product knowledge graph construction module as input for graph construction and node feature initialization. Figure 2 The image shows the interface of the product status monitoring system. As shown, the system displays the currently monitored industrial products and their models through partitions. The lifecycle of each industrial product is obtained through judgment rules, and its corresponding business associations are obtained by combining the product configuration list and product substitution relationship file. For example, if precision bearings, hydraulic valve bodies, sensor groups, motor controllers, industrial robots, precision gears, CNC systems, and pneumatic components are selected as industrial products for this monitoring, after analysis by the product status monitoring module, the lifecycle stage label of precision bearings is "mature stage", and there are 5 associated products. All other monitored industrial products are marked with their corresponding lifecycle stage labels and association relationships.

[0043] Furthermore, a product knowledge graph is constructed based on the relationships between products using a product knowledge graph construction module. The product knowledge graph uses products as nodes and relationships as edges, and a graph neural network is used to aggregate neighbor information for each product node to generate a product state embedding vector for that node. The specific implementation scheme is as follows:

[0044] First, when constructing the product knowledge graph, the product knowledge graph construction module treats each product as an independent node. The initial attributes of a node are constructed by concatenating the one-hot encoding of the product's lifecycle stage and the standardized values ​​of the product attributes. The one-hot encoding of the lifecycle stage is constructed by mapping the five stages—introduction, growth, maturity, decline, and discontinuation—to a five-dimensional binary vector. After the product status monitoring module determines the current stage of a product, the corresponding dimension is set to 1, and the other four dimensions are set to 0. For example, when a product is currently determined to be in the growth stage, its one-hot encoding is... This approach avoids introducing erroneous numerical relationships between stages into the model, ensuring that the graph convolutional network can learn in a discrete category space.

[0045] The corresponding standardized values ​​for product attributes are constructed as follows: Multiple continuous attributes related to product sales characteristics are selected, including the product's standard price, historical average gross profit margin, product weight class, and standard delivery cycle. For each attribute, its mean and standard deviation are calculated within the entire product set. Then, the Z-score standardization method is used to transform the original attribute values ​​into standardized values ​​with a mean of 0 and a standard deviation of 1. That is, let the original value of a product on the k-th attribute be... The mean of this attribute across all products is The standard deviation is The standardized value is:

[0046] ;

[0047] All standardized attribute values ​​are arranged in a fixed order into a numerical vector.

[0048] The numerical vectors and one-hot encodings are then concatenated along the vector dimension to form the initial attribute vector of node v. The specific expression is as follows:

[0049] ;

[0050] in to The values ​​of each dimension of the one-hot encoding represent the lifecycle stage. to Let m be the standardized values ​​of product attributes.

[0051] Through the above concatenation, the initial attribute vector carries both discrete stage information and continuous attribute information of the product, providing a complete input basis for the subsequent neighbor aggregation of the graph convolutional network.

[0052] Subsequently, the relationships output by the product status monitoring module are used as directed edges, with the edge type label taken from one of the following: subordinate, replacement, or upgrade relationships. This generates a heterogeneous directed graph, denoted as . ,in Represents a set of product nodes. This represents a set of edges indicating relationships. For example... Figure 3 The diagram shows the product knowledge graph interface, using different colors to represent the life cycle stages of corresponding industrial products. For example, green represents growth-stage products, blue represents mature-stage products, orange represents new products, and red represents declining products. The corresponding industrial products are represented by model numbers. For example, A-1001 represents precision bearings, and D-4120 represents motor controllers. The two industrial products are connected by dashed lines, representing the corresponding relationships, namely, subordinate relationships, substitution relationships, and upgrade relationships. For example, the associated products of motor controller (D-4120) include precision bearings (A-1001), precision gears (F-6300), hydraulic valve bodies (B-2050), and pneumatic components (H-8500).

[0053] To generate product state embedding vectors with neighbor awareness capabilities, this module employs... Layered graph convolutional networks aggregate neighbor information for product nodes. For the first layer... The feature vector update formula for the product node v in layer 1 is:

[0054] ;

[0055] in, Indicates that node v is at the th The feature vector of the layer, the initial input This is the initial attribute vector for the product node. Indicates that node v is in the graph The set of neighboring nodes in the middle, This is a mean aggregation function that performs an element-wise mean operation on the feature vectors of neighboring nodes. Indicates the first The learnable weight matrix of the layer, for Activation function.

[0056] go through After the layered graph convolution operation, the feature vector of each product node is layer by layer incorporated with information from its direct and indirect neighbors, thereby capturing the topological context of the product in the relational network. The output of the last layer is then taken. As the product state embedding vector of product node v, denoted as This embedding vector encodes information from two levels simultaneously:

[0057] First, the product's own life cycle stage and basic attributes;

[0058] Secondly, the context of associated product states is implicitly expressed through multi-layered neighbor aggregation. For example, when a product has a subsidiary product in decline, its product state embedding vector reflects the ripple effects of the subsidiary product's decline through graph convolution aggregation. Similarly, when a substitute product is in a period of rapid growth, the original product's embedding vector is also modulated by this neighbor information. This vector representation method, which integrates the individual state of a product with its ecological relationships, provides a continuous and comparable numerical basis for the allocation decision module to perform accurate fit calculations.

[0059] like Figure 4 The diagram shows the product state embedding vector interface. After feature vector updates, the corresponding industrial product's product state embedding vector is formed by concatenating one-hot encoding with the standardized values ​​of the product attributes. For example, the product state embedding vector for a precision bearing (A-1001) is represented as follows: .

[0060] Furthermore, since different sales personnel have diverse sales capabilities and are suited to sell different industrial products, and the sales status (life cycle and related relationships) of industrial products changes over time, this solution uses a sales personnel capability analysis module to extract a multi-dimensional implicit representation of each salesperson's capabilities from historical sales records in order to perform adaptive sales processing.

[0061] The multidimensional capability implicitly represents at least three factors: new product adaptability, accessory product sales capability, and legacy order takeover capability. These factors correspond to the sales personnel's capability levels in the three sub-dimensions of new product adaptability, accessory product sales, and legacy order takeover, respectively. The specific extraction process is as follows:

[0062] First, retrieve each completed sales record from the sales record database and construct a training sample for each record. The feature domains of the sample include: salesperson identification S, product identification P, and whether it is a new product indicator. (The value is 1 if the product sold is in the introduction period at the time of the transaction, otherwise it is 0), whether the product sales mark is attached. (Value is 1 if the sale is accompanied by the sale of related products, otherwise it is 0), whether there is an order handover flag. (The value is 1 when the sale is converted from a legacy order transferred by another salesperson, and 0 otherwise), and the standardized sales amount y for this sale. The standardized sales amount is normalized based on the average transaction amount of the company's products to eliminate the dimensional differences caused by different product price ranges.

[0063] Subsequently, a factorization machine model was used to train the above samples. Factorization machines can effectively learn second-order interaction relationships between features under sparse feature combinations, and are particularly suitable for decomposing the potential ability factors of sales personnel from the co-occurrence records of sales personnel and various sales contextual markers. Its prediction formula is:

[0064] ;

[0065] in, This represents the input feature vector. Indicates the global bias term. This represents the first-order weight of the i-th feature. and These are the latent vectors corresponding to the i-th and j-th features, respectively. This represents the inner product operation of two vectors. The factorization machine learns its parameters through stochastic gradient descent to minimize the predicted sales. The mean square error between the actual standardized sales amount y and the actual standardized sales amount y.

[0066] After the factorization machine model completes training and reaches the preset convergence condition, the salesperson capability analysis module performs latent vector extraction, fixing the latent vectors corresponding to the salesperson's identity identifiers learned during training, and directly outputting them as the implicit representation of the salesperson's multidimensional capabilities. The convergence condition is determined using a dual standard: when the mean squared error on the validation set decreases by less than a preset convergence threshold in multiple consecutive training rounds, and the gradient norm of the parameter updates is lower than a preset gradient threshold, the model is considered to have converged. At this point, the latent vectors of each feature domain have stabilized near the optimal solution.

[0067] The module then extracts the latent vector corresponding to the salesperson from the embedding layer parameter matrix of the factorization machine, based on the index of the salesperson's identity identifier S, denoted as... In the mathematical framework of factorization machines, each discrete value in each feature domain is assigned an independent latent vector. These latent vectors are continuously updated during training by capturing the second-order interactions between features. Salesperson identification, as one of the feature domains, has a latent vector that encodes the latent behavioral pattern information condensed by the interaction of the salesperson with contextual features such as new product indicators, accessory product sales indicators, and legacy order takeover indicators across all historical sales records.

[0068] The dimension d of the latent vector is explicitly set to 3 during model initialization. The values ​​of the three dimensions are naturally differentiated through the fitting of the factorization machine to the sales forecast target during training. The three dimensions correspond to the new product adaptability factor, the accessory product sales capability factor, and the legacy order takeover capability factor in a fixed index order.

[0069] Specifically, the dimension value with an index of 1 corresponds to the new product adaptability factor, which reflects the strength of a salesperson's potential to close a deal when faced with an introductory product.

[0070] The dimension value with an index of 2 corresponds to the accompanying product sales capability factor, which reflects the sales personnel's tendency to simultaneously promote the sales of ancillary products during the main product's transaction process;

[0071] The dimension value with an index of 3 corresponds to the legacy order takeover capability factor, which reflects the salesperson's ability to take over unfinished orders handed over by others and successfully convert them into actual transactions.

[0072] Finally, the salesperson capability analysis module outputs a multi-dimensional implicit capability representation vector for each salesperson, which is then used by the allocation decision module to calculate the fit. This mechanism enables salespeople who have never handled legacy orders to obtain a reasonable estimate of their legacy order takeover capability factor through the collaborative filtering effect with their peers who have similar product structures and customer groups, thus avoiding cold start and capability blind spots.

[0073] Specifically, the allocation decision module receives the product state embedding vector of the target product involved in the sales task to be allocated and the multi-dimensional implicit representation of the candidate salesperson's capabilities. It then uses the fit calculation network to generate a fit score for each candidate salesperson for the target product. The specific calculation method is as follows:

[0074] First, the allocation decision module receives sales tasks to be allocated from the task scheduling system. Each task contains a unique identifier for the target product. The module calls the query interface of the product knowledge graph building module to obtain the target product. Corresponding product state embedding vector Simultaneously, the module determines the set of candidate sales personnel based on initial screening rules such as task type, sales region, and product line. And retrieve each candidate salesperson in batches from the salesperson capability analysis module. Implicit representation of multidimensional capabilities .

[0075] The module then uses a pre-trained fit calculation network to evaluate the fit of each "target product - candidate salesperson" pair. The fit calculation network employs a dual-tower neural network structure, consisting of two independent feedforward subnetworks: the product tower and the salesperson tower. Both subnetworks are encoded separately in the latent space and then generate a scalar score through a dot product operation. Specifically, the product tower receives the product state embedding vector of the target product. As input, it is mapped to a product latent vector through a fully connected layer. The mapping formula is:

[0076] ;

[0077] in This represents the weight matrix of the product tower. This represents the bias vector. This represents the activation function.

[0078] Next, the salesperson tower receives the implicit representation of the multidimensional capabilities of the candidate salesperson s. As input, it is also mapped to a salesperson latent vector through a fully connected layer. The corresponding mapping formula is:

[0079] ;

[0080] in, This represents the weight matrix of the salesperson pyramid. This represents the corresponding bias vector.

[0081] Subsequently, the sales personnel were hidden. With product latent vectors Mapped to a shared latent space of the same dimension, the corresponding fit score is obtained by calculating the dot product of the product latent vector and the salesperson latent vector. The specific calculation formula is as follows:

[0082] ;

[0083] It's worth noting that the dot product operation measures the directional consistency and magnitude of two vectors in the latent space. A higher value indicates a stronger fit between the required capabilities of the target product and the capabilities possessed by the salesperson. This is achieved by analyzing the set of candidate salespeople. Each salesperson performs the above calculations to obtain a suitability score list.

[0084] Subsequently, a sales path allocation plan is generated based on the suitability score. The sales path allocation plan assigns the sales tasks of the target product to the selected sales personnel. The specific plan is as follows:

[0085] First, the allocation decision module reads two preset allocation constraints from the system: the maximum number of tasks currently pending for each salesperson. Matching threshold with minimum capability For each candidate salesperson s, the allocation decision module obtains the number of sales tasks that have been assigned but not yet completed. The results are compared with the task volume limit; simultaneously, it is checked whether the suitability score reaches the minimum matching threshold. The filtering rule is to retain only those that simultaneously meet the criteria. and Candidates. Setting constraints can optimize matching while avoiding overloading sales staff and low-quality assignments.

[0086] The module then sorts the shortlisted candidates in descending order of their fit scores, selecting the salesperson with the highest score. If multiple candidates have the same highest score, the candidate with the highest current workload is selected first. A smaller number of sales personnel is used as the tiebreaker. Finally, the allocation decision module generates a sales path allocation plan, which clearly records the target product identifier, the selected sales personnel identifier, and the plan generation timestamp, and pushes it to the task scheduling system for execution.

[0087] Through a three-stage cascaded process of candidate set preparation, fit score calculation, constraint filtering, and optimal assignment, the allocation decision module transforms the results of product status perception and sales personnel multidimensional capability analysis into executable assignment instructions, realizing fixed-point allocation at different product lifecycle nodes and under different types of sales tasks.

[0088] Finally, the actual sales results after the sales path allocation plan was implemented are obtained through the feedback learning module. The reward value is calculated based on the actual sales results, and the specific calculation method is as follows:

[0089] First, after the sales path allocation plan generated and pushed by the allocation decision module is executed, the actual transaction status and transaction amount of this sale are written back to the database. The feedback learning module listens for change events in this database or polls it at fixed intervals to obtain the actual sales results corresponding to this allocation plan.

[0090] Actual sales results include two key fields:

[0091] First, there is a transaction indicator, which shows whether the sales target has ultimately been converted into a valid order;

[0092] Second, the actual sales amount is recorded when the transaction is completed; otherwise, the value of this field is zero.

[0093] The actual sales results are then converted into a scalar reward value. This is used to evaluate the quality of this allocation decision. The reward value is calculated with reference to a preset baseline set by the company:

[0094] If the actual sales exceed the preset baseline, the allocation decision is considered positive, the reward value is positive, and the greater the exceedance, the higher the reward value.

[0095] If the actual sales are lower than the preset baseline, the reward value will be negative, and the greater the gap, the greater the penalty.

[0096] If the sales amount is exactly equal to the baseline, the bonus value is zero.

[0097] The specific calculation formula is as follows:

[0098] ;

[0099] in, , This represents the actual sales amount for this transaction. Indicates the preset baseline. Represents a symbolic function. This represents the natural logarithm function. The purpose of logarithmic transformation is to smooth and compress abnormally high or low sales figures, preventing extreme values ​​from having an excessive impact on model parameters and maintaining the stability of the learning process.

[0100] Furthermore, the feedback learning module utilizes reinforcement learning strategies to update the network parameters of the fit calculation network and the multi-dimensional implicit capability representation in the salesperson capability analysis module. Specifically, it employs the REINFORCE reinforcement learning algorithm based on policy gradients to update all learnable parameters of the product pyramid and salesperson pyramid in the fit calculation network. This stage treats each sales task allocation as a decision-making action, with the allocation decision module selecting from the candidate salesperson set... The probability of selecting salesperson s is determined by the strategy under the current parameters. Given that the log-probability of an action being selected is... ,in To calculate the product tower in the fit calculation network , And the parameters of the sales personnel tower , A set. Strategies Specifically, this involves scoring the fit on the set of valid candidates after constraint filtering. The normalized probability distribution is calculated using the following formula:

[0101] ;

[0102] in, Indicates satisfaction and The set of valid candidates. The goal of parameter update is to maximize the expected reward, and the update formula is:

[0103] ;

[0104] in, Indicates the learning rate. This represents the reward value calculated in the previous step. Represents the logarithmic probability with respect to the parameter The gradient, when the reward value When the reward value is positive, the parameter shifts in the direction that increases the probability of the selected action, making the system more inclined to repeat the selection when encountering similar product and candidate combinations in the future; when the reward value is positive... When the value is negative, the parameter shifts in the direction that reduces the probability of the selected action, so that the system avoids similar choices in the future.

[0105] Finally, the implicit representation of capabilities needs to be fine-tuned. The feedback learning module uses the error between the actual sales revenue and the predicted sales revenue from the factorization machine in the salesperson capability analysis module to fine-tune the multidimensional implicit representation of the currently selected salesperson's capabilities online. In this step, all other parameters of the factorization machine (global bias, first-order weights, latent vectors of non-salespersons, etc.) are frozen, and the focus is solely on minimizing the mean squared error to refine the latent vector of the current salesperson s. Perform one step of gradient descent update, the specific expression is as follows:

[0106] ;

[0107] in This indicates fine-tuning the learning rate. This represents the predicted sales revenue made by the factorization machine based on the current parameters for this sale. This represents actual sales revenue. This update allows the three competency factors for salespeople to adaptively adjust as their recent actual performance fluctuates. For example, if a salesperson's legacy order takeover competency factor was originally low, but they have recently successfully converted multiple legacy orders, then this competency factor will gradually increase after several fine-tunings, making its corresponding dimension value in the multidimensional competency implicit representation more accurately reflect their current actual competency.

[0108] By sequentially operating four stages—sales result acquisition, reward value calculation, fit network parameter update, and capability implicit representation fine-tuning—the feedback learning module forms a complete online learning closed loop: every assignment made by the allocation decision module is tested in a real business environment, and the test results are fed back in the form of reward values ​​and prediction errors, respectively driving the strategy optimization of the fit calculation network and the instance-level correction of the capability implicit representation. This enables the system to automatically maintain the dynamic optimality of the sales path allocation scheme while the product status evolves and the sales personnel's capabilities change.

[0109] This invention constructs a product knowledge graph and aggregates neighboring product state information using a graph neural network. It encodes the product lifecycle stages and their relationships (such as affiliates, replacements, and upgrades) into a unified product state embedding vector, solving the problem of existing systems' inability to perceive the mutual influence between products. By using a factorization machine to extract factors related to sales personnel's new product adaptability, ancillary product sales ability, and legacy order handling ability from historical sales records, it achieves a quantitative characterization of segmented sales capabilities. Based on this, a dual-tower fit calculation network dynamically matches product states with sales personnel's multi-dimensional capabilities in the latent space, generating fit scores and outputting sales path allocation schemes. This ensures that products at different lifecycle stages are assigned to the most suitable sales personnel. The feedback learning module calculates reward values ​​based on actual sales results and continuously optimizes allocation decisions and capability representations using reinforcement learning strategies, forming a closed-loop adaptive evolution. This fundamentally solves the problem of sales path mismatch caused by product state changes and personnel turnover.

[0110] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent management system for industrial product sales data, characterized in that, It includes a product status monitoring module, a product knowledge graph construction module, a sales personnel capability analysis module, an allocation decision module, and a feedback learning module; The product status monitoring module is used to obtain the current life cycle stage of each product and the relationship between products; The product knowledge graph construction module constructs a product knowledge graph based on the relationships between the products. The product knowledge graph uses products as nodes and the relationships as edges. It also uses a graph neural network to aggregate neighbor information for each product node and generate a product state embedding vector for that product node. The sales personnel capability analysis module extracts a multi-dimensional implicit representation of each salesperson's capabilities from historical sales records; The allocation decision module is used to receive the product state embedding vector of the target product involved in the sales task to be allocated and the multi-dimensional implicit representation of the candidate salesperson's capabilities. It uses the fit calculation network to generate a fit score for each candidate salesperson for the target product and generates a sales path allocation scheme based on the fit score. The sales path allocation scheme assigns the sales task of the target product to the selected salesperson. The feedback learning module is used to obtain the actual sales results after the sales path allocation scheme is executed, calculate the reward value based on the actual sales results, and use reinforcement learning strategy to update the network parameters of the fitness calculation network and the multidimensional implicit representation of the sales personnel capability analysis module.

2. The intelligent management system for industrial product sales data according to claim 1, characterized in that, The product status monitoring module includes the following lifecycle stages: introduction, growth, maturity, decline, and discontinuation. Relationships include subordinate relationships, substitution relationships, and escalation relationships; The term "subsidiary relationship" indicates that a certain product is a mandatory or optional accessory to another main product. The substitution relationship indicates that the new product functionally covers and completely replaces the old product; The upgrade relationship indicates that the product is a high-end iteration of the same series.

3. The intelligent management system for industrial product sales data according to claim 2, characterized in that, The rules for determining the lifecycle stage are as follows: When a product's sales have been continuously increasing within a preset observation window since its launch and its market penetration rate is below the first threshold, it is determined to be in the introduction phase; when the sales growth rate exceeds the preset growth rate threshold and there is no continuous decline, it is determined to be in the growth phase. When sales volume tends to stabilize and the fluctuation range is less than the stability threshold, it is judged to be in the maturity stage. When sales show a downward trend for several consecutive periods, it is determined to be a period of decline; When a product is marked as ceasing production, it is directly determined to be in a production stoppage period.

4. The intelligent management system for industrial product sales data according to claim 1, characterized in that, The method for constructing a product knowledge graph in the product knowledge graph construction module includes the following steps: S1. Each product is treated as an independent node, and the initial attributes of the node are formed by concatenating the one-hot encoding of the product's life cycle stage and the standardized values ​​of the product attributes. S2. Using the relationships output by the product status monitoring module as directed edges, with edge type labels taken from one of the following: subordinate, substitution, and upgrade relationships, construct a heterogeneous directed graph, denoted as . ,in Represents a set of product nodes. Represents the set of edges with associated relationships; S3, Adopt Layered graph convolutional networks aggregate neighbor information for product nodes; S4, after After the layer graph convolution operation, the feature vector of each product node will be incorporated with the information of its direct and indirect neighbors layer by layer, capturing the topological context of the product in the relational network. S5. Take the output of the last layer as the product state embedding vector of the product node.

5. The intelligent management system for industrial product sales data according to claim 4, characterized in that, The one-hot encoding of the life cycle stages in S1 is a five-dimensional binary vector mapped from the five stages of introduction, growth, maturity, decline, and shutdown. The method for constructing the standardized values ​​of the product attributes is as follows: S1.1 Select multiple continuous attributes related to product sales characteristics, including but not limited to the product's standard pricing, historical average gross profit margin, product weight class, and standard delivery cycle; S1.2 For each attribute, calculate its mean and standard deviation within the entire product set, and then use the Z-score standardization method to transform the original attribute values ​​into standardized values ​​with a mean of 0 and a standard deviation of 1. S1.3 Arrange all standardized attribute values ​​into a numerical vector in a fixed order.

6. The intelligent management system for industrial product sales data according to claim 4, characterized in that, The product state embedding vector in S5 simultaneously encodes information at two levels: First, the product's own life cycle stage and basic attributes; Secondly, the associated product state context is implicitly expressed through multi-level neighbor aggregation.

7. The intelligent management system for industrial product sales data according to claim 1, characterized in that, The implicit representation of multidimensional capabilities in the sales personnel capability analysis module includes at least the new product adaptability factor, the accessory product sales capability factor, and the legacy order takeover capability factor.

8. The intelligent management system for industrial product sales data according to claim 7, characterized in that, The method for extracting the implicit representation of each salesperson's multidimensional capabilities in the salesperson capability parsing module includes the following steps: S10. Retrieve each completed sales record from the sales record database and construct a training sample for each record. The feature domains of the sample include: salesperson identification S, product identification P, and whether it is a new product indicator. Does it come with a product sales label? Whether the order receipt mark was left behind And the standardized sales amount y for this sale; S11. Train the sample from step S10 using a factorization machine model to calculate the predicted sales revenue. ; S12. Factorization machines learn parameters through stochastic gradient descent to minimize predicted sales. The mean squared error between the actual standardized sales revenue y and the true standardized sales revenue y; S13. Perform latent vector extraction operation to fix the latent vectors corresponding to the salesperson's identity identifiers learned during the training process, and directly output them as the implicit representation of the salesperson's multidimensional capabilities. S14. From the embedding layer parameter matrix of the factorization machine, extract the latent vector corresponding to the salesperson based on the index of the salesperson's identity identifier S, denoted as... ; S15. Initialize the dimension of the latent vector to 3, and process the latent vector... Index matching is performed based on the dimensions; The dimension value with index 1 corresponds to the new product adaptability factor; The dimension value with index 2 corresponds to the accompanying product sales capability factor; The dimension value with index 3 corresponds to the legacy order takeover capability factor.

9. The intelligent management system for industrial product sales data according to claim 1, characterized in that, The method for generating a suitability score for each candidate salesperson for the target product in the allocation decision module includes the following steps: S20. Receive sales tasks to be assigned from the task scheduling system. Each task contains a unique identifier for the target product. ; S21. Call the query interface of the product knowledge graph construction module to obtain the target product. Corresponding product state embedding vector ; S22. Determine the set of candidate sales personnel based on initial screening rules such as task type, sales region, and product line. And retrieve each candidate salesperson in batches from the salesperson capability analysis module. Implicit representation of multidimensional capabilities ; S23. A dual-tower neural network structure is adopted, where the product tower receives the product state embedding vector of the target product. As input, it is mapped to a product latent vector through a fully connected layer. ; S24. Implicit representation of the multidimensional capabilities of sales personnel receiving candidate sales personnel s. As input, it is mapped to a salesperson latent vector through a fully connected layer. ; S25. Implicitly targeting sales personnel With product latent vectors Mapped to a shared latent space of the same dimension, the corresponding fit score is obtained by calculating the dot product of the product latent vector and the salesperson latent vector.

10. The intelligent management system for industrial product sales data according to claim 9, characterized in that, The method for generating a sales path allocation scheme based on fit score in the allocation decision module includes the following steps: S200, Read the two preset allocation constraints of the system: the upper limit of the number of tasks to be processed by the salesperson. Matching threshold with minimum capability ; S201. For each candidate salesperson s, the allocation decision module obtains the number of sales tasks that have been allocated but not yet completed. And compare it with the upper limit of the task volume; S202, Filtering only retains those that simultaneously meet the following conditions. and The candidates are selected, and the set of candidates that pass the screening is obtained. S203. Sort the sales staff in descending order of their suitability scores, and select the sales staff with the highest scores. If multiple candidates have the same highest score, prioritize the sales staff with the highest number of pending tasks. A smaller number of salespeople is used as a tiebreaker to generate the final sales path allocation scheme.