AI-based robot inventory intelligent scheduling method and system

Abstract

The application relates to the technical field of robot scheduling, in particular to an AI-based robot inventory intelligent scheduling method and system, which comprises the following steps: acquiring technical evolution information and market transaction information of robots; constructing a knowledge graph comprising intergenerational evolution edges and demand substitution edges; generating state embedding vectors of each node through a graph neural network; constructing an adversarial generative network to generate candidate scenarios along the intergenerational evolution edges and iteratively optimize strategies to output a target scheduling strategy; and effectively reducing inventory costs and intergenerational depreciation costs.

Description

Technical Field

[0001] This invention relates to the field of robot scheduling technology, and in particular to an AI-based intelligent scheduling method and system for robot inventory management. Background Technology

[0002] In the field of robot inventory management, existing technologies typically rely on statistical forecasting models based on historical data, combined with classic inventory optimization strategies such as economic order quantity (EOQ) or safety stock formulas, to plan the procurement, storage, and sales of robots. During management, robots are often treated as homogeneous commodities, focusing only on their physical lifespan depreciation or fixed inventory holding costs, and inventory levels are adjusted by periodically updating demand forecasts. Some solutions introduce regression or time-series models from machine learning to capture the fluctuations in market demand, but still treat the value of robots as a constant that decreases linearly over time.

[0003] This conventional approach has significant flaws. Existing methods fail to effectively address the generational devaluation phenomenon caused by the rapid iteration of robotics technology—after the advent of a new generation of robots, the market value of the previous generation of robots may drop sharply and non-linearly. Traditional models cannot reflect this change in the rate of value decay caused by technological evolution, resulting in a serious underestimation of the actual losses of old robots remaining in inventory, and thus leading to suboptimal procurement and disposal decisions.

[0004] Furthermore, existing scheduling strategies generally lack a structured understanding of the technological substitution relationships between robots. In market transactions, new robot models often replace the demand for older models, but traditional forecasting models are based only on macro-time series and cannot distinguish the generational relationships and market competition between different robot models. This makes scheduling solutions lag behind in responding to technological upgrades and unable to proactively balance inventory holding costs and the risk of generational depreciation. Summary of the Invention

[0005] This invention provides an AI-based intelligent scheduling method and system for robot inventory management, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an AI-based intelligent scheduling method for robot inventory management, comprising:

[0007] Obtain information on the technological evolution and market transactions of inventory robots;

[0008] A robot inventory management knowledge graph is constructed, mapping each robot to a knowledge graph node. Based on the technological evolution information, generational evolution edges are established and assigned generational weights. The generational weights represent the value decay rate of previous generation robots. At the same time, demand substitution edges are established based on market transaction information.

[0009] For knowledge graphs, a state embedding vector for each node is generated based on the message passing mechanism of a graph neural network to integrate the graph topology position.

[0010] An adversarial generative network is constructed, comprising a scene generation module and a strategy evaluation module. The scene generation module generates candidate scenes containing subsequent generation technology activation events along the intergenerational evolution edge. The strategy evaluation module calculates a comprehensive loss value and optimizes the scheduling strategy based on the candidate scenes and the state embedding vectors of each node. The optimized strategy is fed back to the scene generation module to generate new scenes that drive further optimization of the strategy. The process is iteratively executed until convergence, and then the target scheduling strategy is output. The comprehensive loss value consists of an inventory cost term and an intergenerational depreciation cost term.

[0011] A robot inventory management knowledge graph is constructed, mapping each robot to a knowledge graph node. Based on the technological evolution information, intergenerational evolution edges are established and assigned intergenerational weights. These weights represent the value decay rate of previous generations of robots. Simultaneously, demand substitution edges are established based on market transaction information, including:

[0012] Map each robot's identifier to a knowledge graph node;

[0013] Based on the aforementioned technological evolution information, identify the previous generation robot nodes and the next generation robot nodes that have a technological iteration relationship, and establish an intergenerational evolution edge between the previous generation robot nodes and the next generation robot nodes.

[0014] The initial generation weight is calculated based on the differences in technical parameters between the previous generation and the next generation in the technology evolution information. The price change data of the previous generation robot after the release of the next generation technology is obtained from the historical market transaction information. The value decay coefficient is extracted based on the time decay characteristics of the price change data. The value decay coefficient is fused with the initial generation weight to generate the final generation weight and assigned the corresponding generation evolution edge.

[0015] Based on the market transaction information, robot node pairs with market competition relationships are identified, and demand substitution edges are established between the robot node pairs.

[0016] The value decay coefficient is extracted based on the time decay characteristics of the price change data, including:

[0017] The number of transaction records of previous generation robots in historical market transaction information is statistically analyzed in each time period after the release of the next generation robot technology. The length of the time window is dynamically determined based on the distribution density of the number of transaction records. The length of the time window is inversely proportional to the distribution density of the number of transaction records.

[0018] Identify abnormal data points where the price change exceeds a preset deviation threshold within each time window, determine a weighting coefficient based on the degree of deviation of the abnormal data points, and perform weighted processing or remove the abnormal data points.

[0019] The local decay rate is calculated for the processed price change data within each time window, and a moving average mechanism is used to smooth the local decay rate of adjacent time windows over time.

[0020] When the number of valid samples of price change data is lower than a preset sparsity threshold, the average decay rate of other robots of the same technology generation is obtained as the generational decay prior mode. The generational decay prior mode is then weighted and fused with the local decay rate of the previous generation robot in inverse proportion to the number of samples to generate a value decay coefficient.

[0021] For knowledge graphs, a state embedding vector for each node is generated based on the message passing mechanism of a graph neural network, including:

[0022] For each node in the knowledge graph, aggregate the features of its neighboring nodes on the generational evolution edge and the features of its neighboring nodes on the demand substitution edge.

[0023] The characteristics of adjacent nodes on the generational evolution edge are assigned generational propagation weights, so that the generational propagation weights are positively correlated with the generational weights of the corresponding generational evolution edge. The characteristics of adjacent nodes on the demand substitution edge are assigned competitive propagation weights.

[0024] The weighted features of adjacent nodes of the generational evolution edge and the features of adjacent nodes of the demand replacement edge are fused and aggregated to obtain the neighborhood aggregation features of the node.

[0025] The initial features of the node are nonlinearly transformed with the neighborhood aggregation features to generate a state embedding vector that integrates the node's topological position in the dimensions of generational evolution and demand competition.

[0026] The scene generation module generates candidate scenes containing subsequent technology activation events along the generational evolution edge, including:

[0027] Identify the previous generation robot nodes in the knowledge graph that have descendant robot nodes, and use the previous generation robot nodes as the starting nodes for generating candidate scenarios; for each starting node, trace along its generational evolution edge to the corresponding descendant robot node, and obtain the technical parameters and historical technology release cycle data of the descendant robot node.

[0028] An initial probability distribution of subsequent technology activation events is constructed based on the aforementioned technical parameters and historical technology release cycle data. The subsequent technology activation events include the predicted release time and technical performance indicators of the subsequent technologies.

[0029] Extract the generational weights of the generational evolution edges, and use the generational weights as a time adjustment factor to scale and transform the time axis of the initial probability distribution so that the predicted release time is negatively correlated with the generational weights.

[0030] The initial probability distribution after scaling transformation is sampled to generate the technology activation time point of the descendant, and the technical impact intensity corresponding to the activation time point is determined according to the technical performance index of the descendant robot node.

[0031] The time series containing the activation time points of the next-generation technology and the intensity of the technological impact are constructed as candidate scenarios, which are used to simulate the devaluation of the previous generation robot under the impact of the next-generation technology.

[0032] The strategy evaluation module calculates a comprehensive loss value based on the candidate scenarios and the state embedding vectors of each node, including:

[0033] Based on the current scheduling strategy, the inventory holding time of each robot node within the candidate scenario time range is obtained. The inventory cost of each node is calculated based on the inventory holding time and the unit time maintenance cost of each node. The inventory costs of all nodes are summed and normalized to obtain the inventory cost item.

[0034] For the descendant technology activation events contained in the candidate scenarios, identify the generational evolution edge corresponding to the descendant technology activation event, trace back along the generational evolution edge to the previous generation robot node affected by the technological impact, obtain the state embedding vector of the previous generation robot node, and extract the feature components of the generational evolution dimension from the state embedding vector.

[0035] The feature components are fused with the generational weights of the corresponding generational evolution edges to calculate the unit value decay of each previous generation robot node at the time of activation of the next generation technology.

[0036] The predicted inventory quantity of the previous generation robot node at the time of activation of the next generation technology is obtained according to the current scheduling strategy. The unit value decay rate is multiplied by the predicted inventory quantity and the node unit price to obtain the intergenerational depreciation cost of each previous generation robot node. The intergenerational depreciation costs of all previous generation robot nodes are summed and normalized to obtain the intergenerational depreciation cost item.

[0037] The weighted sum of the inventory cost item and the intergenerational depreciation cost item yields the comprehensive loss value.

[0038] The optimization strategy is fed back to the scene generation module to generate new scenes that drive further optimization of the strategy. This process is iterated until convergence, at which point the target scheduling strategy is output, including:

[0039] The strategy evaluation module performs gradient optimization on the parameters of the current scheduling strategy based on the comprehensive loss value, and generates an optimized scheduling strategy.

[0040] Extract the direction of the processing time adjustment of each robot node in the optimized scheduling strategy, and identify the robot nodes whose processing time is advanced as high-risk nodes;

[0041] The high-risk node identifier is fed back to the scene generation module. The scene generation module calculates the ratio of the number of high-risk nodes to the total number of nodes as the scene enhancement coefficient. For the generational evolution edge corresponding to the high-risk node, the original scene generation intensity is multiplied by the scene enhancement coefficient to obtain the enhanced scene generation intensity. According to the enhanced scene generation intensity, the descendant technology activation event for the high-risk node is generated to form a new scene.

[0042] The strategy evaluation module recalculates the comprehensive loss value based on the new scenario and continues to optimize the scheduling strategy, repeating the feedback and scenario generation process.

[0043] When the change in the comprehensive loss value is lower than the preset convergence threshold in multiple consecutive iterations, the optimization process is determined to have converged, and the current scheduling strategy is output as the target scheduling strategy.

[0044] A second aspect of this invention provides an AI-based intelligent scheduling system for robot inventory management, comprising:

[0045] The information acquisition unit is used to acquire information on the technological evolution and market transactions of the inventory robot.

[0046] The knowledge construction unit is used to construct a robot inventory knowledge graph, mapping each robot to a knowledge graph node, establishing generational evolution edges based on the technology evolution information and assigning generational weights, wherein the generational weights represent the value decay rate of the previous generation of robots, and establishing demand substitution edges based on market transaction information.

[0047] The vector generation unit is used to generate state embedding vectors for each node in the knowledge graph by means of the message passing mechanism of the graph neural network.

[0048] The strategy optimization unit is used to construct an adversarial generative network including a scenario generation module and a strategy evaluation module. The scenario generation module generates candidate scenarios containing subsequent generation technology activation events along the generational evolution edge. The strategy evaluation module calculates the comprehensive loss value and optimizes the scheduling strategy based on the candidate scenarios and the state embedding vectors of each node. The optimized strategy is fed back to the scenario generation module to generate new scenarios that drive further optimization of the strategy. The process is iteratively executed until convergence, and then the target scheduling strategy is output. The comprehensive loss value consists of an inventory cost item and a generational depreciation cost item.

[0049] A third aspect of the present invention provides an electronic device, comprising:

[0050] processor;

[0051] Memory used to store processor-executable instructions;

[0052] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0053] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0054] The knowledge graph constructed in this invention transforms information on the evolution of robotics technology and market transactions into generational evolution edges and demand substitution edges. Generational weights precisely quantify the rate of value decay of previous generations of robots, while demand substitution edges dynamically reflect shifts in market preferences. This two-dimensional knowledge structure enables the system to capture the complex relationship between technological iteration and market demand, providing a structured cognitive foundation for scheduling decisions.

[0055] The message passing mechanism of graph neural networks aggregates the state information of each node along the knowledge graph topology path, generating a state embedding vector that integrates intergenerational evolution relationships and demand substitution relationships. This vector simultaneously encodes the robot's intergenerational position, market status, and potential substitution relationships, enabling the agent to accurately assess the real-time value weights and supply and demand dynamics of different robot models, and improving the accuracy of differentiated regulation of heterogeneous robot groups.

[0056] Generative Adversarial Networks (GANs) simulate subsequent technology activation events along the generational evolution edge through a scenario generation module, generating a set of candidate scenarios that include uncertainties such as technology updates and demand changes. A strategy evaluation module calculates a comprehensive loss value consisting of inventory costs and generational depreciation costs based on state embedding vectors. This iterative game forces the scheduling strategy to simultaneously consider both current inventory holding costs and future technology depreciation risks, adaptively adjusting the pace of procurement, warehousing, and outbound shipments in a dynamic equilibrium, effectively suppressing large-scale asset impairment caused by technology iteration. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating an AI-based intelligent scheduling method for robot inventory management.

[0058] Figure 2 Flowchart for constructing a robot's inventory management knowledge graph. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0061] Figure 1 This is a flowchart illustrating the AI-based intelligent inventory management scheduling method for robots according to an embodiment of the present invention. The AI-based intelligent inventory management scheduling method for robots includes:

[0062] Obtain information on the technological evolution and market transactions of inventory robots;

[0063] A robot inventory management knowledge graph is constructed, mapping each robot to a knowledge graph node. Based on the technological evolution information, generational evolution edges are established and assigned generational weights. The generational weights represent the value decay rate of previous generation robots. At the same time, demand substitution edges are established based on market transaction information.

[0064] For knowledge graphs, a state embedding vector for each node is generated based on the message passing mechanism of a graph neural network to integrate the graph topology position.

[0065] An adversarial generative network is constructed, comprising a scene generation module and a strategy evaluation module. The scene generation module generates candidate scenes containing subsequent generation technology activation events along the intergenerational evolution edge. The strategy evaluation module calculates a comprehensive loss value and optimizes the scheduling strategy based on the candidate scenes and the state embedding vectors of each node. The optimized strategy is fed back to the scene generation module to generate new scenes that drive further optimization of the strategy. The process is iteratively executed until convergence, and then the target scheduling strategy is output. The comprehensive loss value consists of an inventory cost term and an intergenerational depreciation cost term.

[0066] In one optional implementation, a robot inventory management knowledge graph is constructed, mapping each robot to a knowledge graph node. Generational evolution edges are established based on the technological evolution information and assigned generational weights, where each generational weight represents the value decay rate of the previous generation of robots. Simultaneously, demand substitution edges are established based on market transaction information, including:

[0067] Map each robot's identifier to a knowledge graph node;

[0068] Based on the aforementioned technological evolution information, identify the previous generation robot nodes and the next generation robot nodes that have a technological iteration relationship, and establish an intergenerational evolution edge between the previous generation robot nodes and the next generation robot nodes.

[0069] The initial generation weight is calculated based on the differences in technical parameters between the previous generation and the next generation in the technology evolution information. The price change data of the previous generation robot after the release of the next generation technology is obtained from the historical market transaction information. The value decay coefficient is extracted based on the time decay characteristics of the price change data. The value decay coefficient is fused with the initial generation weight to generate the final generation weight and assigned the corresponding generation evolution edge.

[0070] Based on the market transaction information, robot node pairs with market competition relationships are identified, and demand substitution edges are established between the robot node pairs.

[0071] Combination Figure 2 The flowchart illustrating the construction process of the robot inventory knowledge graph explains that when mapping each robot's identifier to a knowledge graph node, all robot models involved in the inventory need to be uniformly coded, with each robot model corresponding to a unique node in the graph. In addition to storing the model identifier, the node attributes also need to include structured attribute fields such as a technical specification summary, launch date, current inventory level, and historical selling price. These attribute fields will serve as initial feature inputs for the node in subsequent edge weight calculations and graph neural network message passing. For robots of the same category but different generations, their nodes exist as independent entities in the graph and are not merged, in order to fully preserve the hierarchical structure information of generational evolution.

[0072] When identifying robot node pairs with a technological iteration relationship, the system uses product iteration records, technical white papers, and official upgrade announcements in the technology evolution information to determine whether two robots constitute an evolutionary relationship between a predecessor and a successor. The criteria for judgment include: whether the successor robot has systematically replaced the predecessor robot in core technical indicators (such as motion accuracy, perception capability, load capacity, and energy consumption), and whether the manufacturer has explicitly stated the inheritance relationship between the two. For node pairs that meet the above conditions, a directed generational evolution edge is established between the predecessor and successor robot nodes, with the direction pointing from the predecessor to the successor, representing the directionality of technological evolution. A predecessor node can establish evolution edges to multiple successor nodes to support scenarios with multiple models or branch iterations within a single generation.

[0073] When calculating the initial generational weights, the parameter vectors of previous and subsequent generations are extracted from the technological evolution information. Several key technological dimensions (such as positioning accuracy, maximum speed, endurance, sensor configuration level, etc.) are selected, and the parameter differences in each dimension are normalized and then weighted and summed to obtain a comprehensive measure of technological parameter differences. Let the previous generation robot be at the... The parameter values ​​for each technical dimension are The parameter values ​​for the corresponding dimension of the offspring robot are The normalized weights for this dimension are Then the technical parameter difference measurement value The calculation is as follows: ;in This represents the total number of technical dimensions involved in the calculation. The larger the value, the higher the degree to which the next generation technology replaces the previous generation technology, the stronger the pressure of value decay faced by the previous generation robot, and the higher the corresponding initial generation weight.

[0074] When acquiring historical market transaction data on price changes of previous-generation robots after the release of subsequent technologies, the official release date of the subsequent product is used as the time benchmark. The average market price sequence of the previous-generation robot over several subsequent statistical periods (e.g., monthly or quarterly) is extracted. This price sequence is compared with the benchmark price before release, and the price retention rate at each time point is calculated, i.e., the ratio of the current average price to the benchmark price. The decreasing trend of the price retention rate over time is the time decay characteristic. An exponential decay fit is applied to the price retention rate sequence, and the decay rate parameter is extracted as the value decay coefficient. Let the first... The price retention rate for each statistical period is The fitting model is: ;in This is the initial retention rate (usually close to 1). This is the value decay coefficient. This represents the number of statistical periods since the release of subsequent technologies. The above model is fitted to the observed price retention rate sequence using the least squares method, and the solution is obtained. The estimated value. The higher the value, the faster the price of the previous generation robot declines after the release of the next generation technology, and the faster the market's recognition of the value of the previous generation product diminishes.

[0075] Value depreciation coefficient with initial generation weight Perform fusion calculations to generate the final generation weights. The fusion operation uses a weighted linear combination method: ;in and These are the fusion weighting coefficients for the technical parameter difference component and the value decay component, respectively, with a sum of 1. The specific values ​​can be configured based on the business side's emphasis on the speed of technology iteration and market price sensitivity. Final generational weight. Assigning corresponding generational evolution edges as edge weight inputs in the message passing stage of the graph neural network directly affects the attenuation ratio when the previous generation node transmits information to the next generation node, thereby encoding the inventory depreciation risk of the previous generation robot due to the activation of the next generation technology at the topological level.

[0076] When identifying robot node pairs with market competition based on market transaction information, the following signals are emphasized: robot model pairs that are alternately purchased by customers in the same application scenarios (such as warehousing, precision assembly, and customer service); model pairs that show significant negative correlation in sales volume within the same procurement cycle; and model pairs that are explicitly listed as mutually substitutable alternatives by customers in market research data. These signals are comprehensively scored, and robot node pairs exceeding a set threshold are determined to have market competition. An undirected demand substitution edge is then established between the two nodes. The edge weight of the demand substitution edge can be assigned based on the absolute value of the negative correlation coefficient between the sales volumes of the two products in historical transaction data. The stronger the negative correlation, the higher the weight of the demand substitution edge, indicating a stronger mutual substitutability of the two robots at the market level.

[0077] Demand substitution edges and generational evolution edges together constitute the complete topological structure of the robot's inventory management knowledge graph. Generational evolution edges describe the vertical evolutionary relationships at the technological level, while demand substitution edges describe the horizontal competitive relationships at the market level. The coexistence of these two types of edges in the graph allows the graph neural network to simultaneously capture the transmission paths of inventory depreciation caused by technological iteration and the demand transfer paths caused by market competition during message transmission, providing more comprehensive semantic support for subsequent scheduling strategy optimization. In actual construction, for the same node pair that has both technological iteration relationships and market competition relationships, it is permissible to establish both generational evolution edges and demand substitution edges. The two types of edges carry different semantic information and are distinguished during the message transmission stage through different edge type embeddings to avoid information confusion.

[0078] In one optional implementation, extracting a value decay coefficient based on the time decay characteristics of the price change data includes:

[0079] The number of transaction records of previous generation robots in historical market transaction information is statistically analyzed in each time period after the release of the next generation robot technology. The length of the time window is dynamically determined based on the distribution density of the number of transaction records. The length of the time window is inversely proportional to the distribution density of the number of transaction records.

[0080] Identify abnormal data points where the price change exceeds a preset deviation threshold within each time window, determine a weighting coefficient based on the degree of deviation of the abnormal data points, and perform weighted processing or remove the abnormal data points.

[0081] The local decay rate is calculated for the processed price change data within each time window, and a moving average mechanism is used to smooth the local decay rate of adjacent time windows over time.

[0082] When the number of valid samples of price change data is lower than a preset sparsity threshold, the average decay rate of other robots of the same technology generation is obtained as the generational decay prior mode. The generational decay prior mode is then weighted and fused with the local decay rate of the previous generation robot in inverse proportion to the number of samples to generate a value decay coefficient.

[0083] For example, the value depreciation factor The extraction process relies on the refined processing of price change data in historical market transaction information. Since the market performance of previous generation robots after the release of subsequent robot technologies often exhibits non-uniform time distribution characteristics, directly modeling the entire historical data in a unified manner can easily obscure key price drop points. Therefore, it is necessary to introduce an adaptive time window segmentation mechanism.

[0084] We statistically analyze the number of transaction records of previous-generation robots in historical market transaction information within each time period after the release of subsequent-generation robot technologies, using this data to measure the distribution density of the data. Let's assume the... The number of transaction records within each statistical period is Time window length and They are inversely proportional, that is, when When it is large, Take a smaller value to improve the resolution of local price changes; when When smaller, Choose a larger value to ensure that there are sufficient samples within the window to support local statistics. Specifically, Based on Confirmed, among which As the baseline window length, The adjustment coefficient controls the sensitivity of the window length to changes in trading density. This adaptive window segmentation method can create fine-grained time slices in the early stages of a technology release (active trading, volatile prices) and coarse-grained time slices in the later stages (sparse trading, stable prices), thus capturing the dynamic patterns of price decay more accurately overall.

[0085] Within each time window, factors such as sudden market events, promotional activities, or data entry errors may cause some transaction prices to deviate from the normal decline trend, forming outlier data points. Outlier data points are identified by calculating the deviation of each transaction price within each time window from the median price within that window. When the absolute value of the deviation exceeds a preset deviation threshold... When this happens, the data point is marked as an outlier. For marked outlier data points, a weighting factor is determined based on the degree of deviation. ,in This is the ID of the outlier data point. The greater the deviation, the higher the ID number. The smaller the value, the lower the contribution of that data point to subsequent statistics. When the deviation exceeds an extreme threshold, Setting it to 0 is equivalent to directly removing the data point; when the deviation is between Between extreme thresholds, The weighting decreases from 1 to near 0 using a linear or exponential method, achieving a soft reduction. This approach is more robust than hard removal, preserving potentially useful information from outliers while suppressing their interference with the overall decay trend estimation.

[0086] For the processed price change data, calculate the local decay rate within each time window. Let the... The weighted average price after weight reduction within each time window is: The local attenuation rate between adjacent windows Defined as This reflects the percentage decrease in price of the previous generation robot during that time period. Since the local decay rate between adjacent time windows may fluctuate due to data volatility at the window boundaries, a moving average mechanism is used to... The sequence is time-smoothed. Let the smoothed decay rate be... The moving average window width is ,but The smoothed decay rate sequence eliminates short-term random disturbances, making the overall decay trend more consistent and providing a more reliable basis for subsequent fitting of the decay coefficient. It provides higher quality input data.

[0087] When the number of effective transaction samples of a previous generation robot is lower than a preset sparsity threshold In this case, relying solely on the robot's own historical price data is insufficient to reliably estimate its local attenuation rate; therefore, external prior information needs to be introduced to supplement it. The average attenuation rate of other robots in the same technological generation as the previous generation robot is obtained and used as a priori model for generational attenuation. The generational decay prior model reflects the general value decay pattern of contemporary technological products when facing the impact of subsequent generations of technology, and has strong statistical representativeness. This model is compared with the local decay rate of previous generation robots. The weighted fusion is performed inversely proportional to the sample size to obtain the estimated decay rate after fusion. Let the current number of valid samples be... Then the weights of the prior patterns in the fusion weights and Inversely proportional to the weight of its own local decay rate and Proportional, specifically expressed as .when When the value approaches 0, the fusion result is completely dominated by the prior pattern; when Much larger In this case, the fusion result is mainly based on its own local decay rate, and the influence of the prior mode is negligible. This Bayesian weighted fusion strategy effectively avoids estimation bias caused by insufficient samples in data-sparse scenarios, while ensuring the individual relevance of the estimation results when data is sufficient.

[0088] The fused decay rate sequence Based on this, the value decay coefficient is finally extracted by fitting its time series with an exponential decay model. Combined with the aforementioned price retention rate and , , The exponential relationship between them can be used to... The sequence is transformed into the corresponding retention rate sequence, and then the least squares method or maximum likelihood estimation method is used to... Perform parameter fitting to achieve It can accurately reflect the true rate of value decay of the previous generation robot in the current market environment. The results... The value will further participate in generational weighting. The calculation provides data support for assigning weights to the intergenerational evolution edges in the knowledge graph, ultimately affecting the optimization direction of the message passing process and scheduling strategy of the graph neural network.

[0089] In one optional implementation, for a knowledge graph, a state embedding vector for each node is generated using the message passing mechanism of a graph neural network, comprising:

[0090] For each node in the knowledge graph, aggregate the features of its neighboring nodes on the generational evolution edge and the features of its neighboring nodes on the demand substitution edge.

[0091] The characteristics of adjacent nodes on the generational evolution edge are assigned generational propagation weights, so that the generational propagation weights are positively correlated with the generational weights of the corresponding generational evolution edge. The characteristics of adjacent nodes on the demand substitution edge are assigned competitive propagation weights.

[0092] The weighted features of adjacent nodes of the generational evolution edge and the features of adjacent nodes of the demand replacement edge are fused and aggregated to obtain the neighborhood aggregation features of the node.

[0093] The initial features of the node are nonlinearly transformed with the neighborhood aggregation features to generate a state embedding vector that integrates the node's topological position in the dimensions of generational evolution and demand competition.

[0094] For example, after the knowledge graph is constructed, a state embedding vector that can simultaneously capture intergenerational evolutionary relationships and market competition relationships needs to be generated for each robot node in the graph. The message passing mechanism of graph neural networks provides a natural technical framework for this: each node updates its own feature representation by aggregating information from its neighboring nodes, and the semantic information carried by different types of edges needs to be differentiated before it can be effectively fused.

[0095] Knowledge graphs contain two types of structural edges: generational evolution edges depict the technological inheritance and value transfer relationships from previous generations of robots to subsequent generations, while demand substitution edges depict the competition and substitution relationships between different robot products at the market level. For any node in the graph... First, identify the set of neighboring nodes on the generational evolution edge. The set of neighboring nodes on the demand substitution edge Separating the two types of neighbor sets is a prerequisite for ensuring differentiated aggregation in the future. If the two types of neighbors are treated together, the intergenerational inheritance signal will be diluted by the market competition signal, causing the state embedding vector to fail to accurately reflect the position of the node in the technology evolution chain.

[0096] For the aggregation of neighbor node features on the generational evolution edge, an intergenerational propagation weight is introduced. To measure nodes For nodes The degree of information contribution. The intergenerational propagation weight is positively correlated with the intergenerational weight on the corresponding intergenerational evolution edge, that is, the larger the intergenerational weight (indicating that the value of the previous generation of robots decays faster and the technological impact of the next generation is stronger), the greater the influence weight of the information transmitted by the intergenerational evolution edge on the target node. Specifically, It can be obtained by normalizing the generational weights, so that the nodes The sum of all propagation weights along the generational evolution direction is 1, thus ensuring the numerical stability of the aggregation result. Neighborhood aggregation characteristics on generational evolution edges. This can be represented as a weighted sum of the initial feature vectors of the neighboring nodes: ;in For nodes The initial feature vector contains multi-dimensional original features such as the robot's technical specifications, historical inventory, and market price.

[0097] To aggregate the features of neighbor nodes on the demand substitution edge, a competitive propagation weight is introduced. To measure nodes With nodes The intensity of market competition between the two products. The calculation of the competition propagation weight can be comprehensively determined based on dimensions such as customer overlap rate, price range similarity, and functional coverage of the two robot products in market transaction information, and is also normalized. Neighborhood aggregation characteristics on demand substitution edges. Represented as: .

[0098] Because the demand substitution edge reflects horizontal competition rather than vertical inheritance. The information carried mainly reflects the nodes. The substitution pressures faced in the current market landscape, and The information about the generational technological impact they carry is semantically different, so the two need to be integrated through a fusion and aggregation process rather than simply superimposed.

[0099] The fusion and aggregation stage will incorporate neighborhood characteristics of the generational evolution dimension. Neighborhood characteristics in the dimension of competition with demand By splicing, we obtain nodes. Neighborhood aggregated feature vector : ;in This represents a vector concatenation operation. Compared to element-wise addition, concatenation preserves the independence of the two types of information, allowing subsequent nonlinear transformation layers to adaptively learn the fusion ratio of the two types of information, rather than artificially fixing the fusion weights during the aggregation stage. This improves the adaptability of the state embedding vector to different market environments.

[0100] During the stage of generating the final state embedding vector, the nodes initial feature vector Features of neighborhood aggregation Common input nonlinear transformation layer. Initial features. It preserves the inherent attribute information of the node itself, while This carries contextual information from the generational evolution chain and the market competition network. The combination of these two elements ensures that the final state embedding vector possesses both the individual distinctiveness of the node and reflects its positional characteristics within the overall graph topology. The nonlinear transformation utilizes a learnable weight matrix. and activation function (e.g., ReLU or ELU) Completed: ;in For bias vectors, That is, a node The final generated state embedding vector. The dimension is determined by The number of rows is determined by the input dimension of the policy evaluation module in the subsequent adversarial generative network, and is usually set to a fixed value that matches the input dimension of the policy evaluation module in the subsequent adversarial generative network, so as to ensure the parameter consistency of the overall architecture.

[0101] The aforementioned message passing process can undergo multiple iterations. Each iteration uses the embedding vector output from the previous iteration as the node feature input for the current iteration, allowing the node's state embedding vector to gradually aggregate structural information from more distant neighbors. In practical applications, the number of iterations is typically set to 2 to 3 to achieve a balance between receptive field coverage and computational overhead. Too many iterations can lead to oversmoothing, causing the embedding vectors of different nodes to become similar, thus weakening the discernibility of positional differences between robot nodes in the generational evolution chain.

[0102] It is worth noting that, because the evolutionary edge of generations has directionality (from the previous generation to the next generation), in aggregation... It is necessary to distinguish nodes. Is it acting as a message receiver (descendant node) or a message sender (previous generation node)? For nodes acting as descendants... ,That Includes all edges pointing through generational evolution The predecessor node; for nodes that are predecessor nodes Then, features of its descendant nodes can be aggregated to replace pressure with perception technology. This differentiated aggregation strategy on the directed graph enables the state embedding vector to accurately encode the hierarchical position of the node in the generational evolution dimension, providing an accurate node state description basis for the scene generation module in the subsequent adversarial generative network to generate descendant technology activation events along the generational evolution edge.

[0103] In one optional implementation, the scene generation module generates candidate scenes containing descendant technology activation events along the generational evolution edge, including:

[0104] Identify the predecessor robot nodes in the knowledge graph that have descendant robot nodes, and use the predecessor robot nodes as the starting nodes for generating candidate scenarios.

[0105] For each starting node, trace along its generational evolution edge to the corresponding descendant robot node, and obtain the technical parameters and historical technology release cycle data of the descendant robot node;

[0106] An initial probability distribution of subsequent technology activation events is constructed based on the aforementioned technical parameters and historical technology release cycle data. The subsequent technology activation events include the predicted release time and technical performance indicators of the subsequent technologies.

[0107] Extract the generational weights of the generational evolution edges, and use the generational weights as a time adjustment factor to scale and transform the time axis of the initial probability distribution so that the predicted release time is negatively correlated with the generational weights.

[0108] The initial probability distribution after scaling transformation is sampled to generate the technology activation time point of the descendant, and the technical impact intensity corresponding to the activation time point is determined according to the technical performance index of the descendant robot node.

[0109] The time series containing the activation time points of the next-generation technology and the intensity of the technological impact are constructed as candidate scenarios, which are used to simulate the devaluation of the previous generation robot under the impact of the next-generation technology.

[0110] For example, in the scenario of intelligent scheduling of robot inventory management, simulating the devaluation scenario of older generation robots facing technological challenges from younger generations is the core basis for formulating reasonable scheduling strategies. The primary task of the scenario generation module is to locate older generation robot nodes with younger generation robot nodes in the knowledge graph. Specifically, it traverses all nodes in the knowledge graph and checks whether each node has a starting edge for intergenerational evolution. If a node has at least one intergenerational evolution edge pointing to a younger generation robot node, then that node is marked as the starting node for candidate scenario generation. This identification process ensures that scenario generation only targets older generation robots that are truly facing intergenerational competition pressure, avoiding invalid scenario construction for robots without younger generation nodes, thereby improving overall computational efficiency.

[0111] For each identified starting node, path tracing is performed along its generational evolution edge. Upon reaching the corresponding descendant robot node, the set of technical parameters associated with that descendant node and historical technology release cycle data are extracted. The set of technical parameters covers the performance indicators of the descendant robot in various key technical dimensions, such as quantitative indicators like motion accuracy, load capacity, perception range, and energy consumption. The historical technology release cycle data comes from the historical release intervals of robots of the same generation or category recorded in market transaction information, including statistics on the time span from the release of the previous generation product to the official market launch of the descendant product. These data together form the input basis for subsequent probability distribution modeling, ensuring that the generated candidate scenarios are adequately supported by historical data.

[0112] Based on the acquired technical parameters and historical technology release cycle data, an initial probability distribution for subsequent technology activation events is constructed. Subsequent technology activation events comprise two key elements: predicted release time and technical performance metrics. The initial probability distribution for predicted release time is modeled using a log-normal distribution, assuming there are a total of [number missing] events in the historical technology release cycle sample set. The set of valid historical records and the corresponding publication period observations are: The mean is calculated by taking the natural logarithm of each observation. and standard deviation This allows us to determine the parameters of the initial log-normal distribution. The log-normal distribution is chosen because technology release cycles typically exhibit a right-skewed distribution, meaning that shorter release intervals are more likely, while extremely long intervals are less frequent—a pattern consistent with real-world technology iteration. Technical performance indicators are based on the technical parameter vectors of subsequent robot nodes. The overall performance improvement is calculated by comparing the difference vector between the technical parameters of the next generation robot and those of the previous generation, serving as the fundamental input for estimating the intensity of subsequent technological impacts.

[0113] Intergenerational weights are introduced here as time adjustment factors to perform a scaling transformation on the time axis of the initial probability distribution. Let the intergenerational weights of the extracted intergenerational evolution edges be... Time axis scaling factor and They are negatively correlated, specifically expressed as ,in To adjust the sensitivity coefficient, which controls the influence of generational weights on the time axis scaling, the mean parameter of the initial log-normal distribution is used. Multiply by scaling factor The scaled mean of the distribution is obtained. The standard deviation parameter remains unchanged to maintain the reasonableness of the distribution shape. The physical meaning of this scaling transformation is: intergenerational weights When the value of the previous generation of robots is relatively large, the rate of value decay of the previous generation of robots is faster, and the expected release time of the next generation of technology should be brought forward accordingly. That is, the predicted release time is negatively correlated with the generational weight. Conversely, when the generational weight is relatively small, the value of the previous generation of robots is relatively stable, and the predicted distribution of the release time of the next generation of technology shifts to a longer period, so that the time structure of the candidate scenario is consistent with the intensity of generational competition.

[0114] Monte Carlo sampling is performed on the scaled probability distribution to generate a sequence of activation time points for the progenitor technique. Let the sampling time span for a single candidate scene be... ,exist The activation time points of the next generation technology are obtained by sampling within the range according to a scaled log-normal distribution. If in If no activation time point is sampled, then... Setting it to a dummy value beyond the scheduling window indicates that descendant technologies in this candidate scenario will not be activated during the current scheduling period. Technological Impact Intensity Based on the technical performance indicators of the descendant robot nodes, let the performance difference vector between the descendant robot and the previous generation robot in each technical dimension be denoted as . ,in The The component represents the first... The performance improvement in each technical dimension, the intensity of the technological impact, is determined by... The result is obtained by applying a weighted norm, i.e. ,in This is a weighted matrix based on the importance of different technological dimensions, reflecting the differentiated impact of various technologies on market demand substitution. (Technological impact intensity) The higher the value, the stronger the market substitution ability of the next generation of technology for the previous generation of robots, and the faster the depreciation rate of the inventory of the previous generation of robots will be after the activation point.

[0115] Activation time of next generation technologies and technical impact strength Combining into event pairs The event pairs obtained from multiple samplings are arranged chronologically to construct candidate scenarios in time series form. A complete candidate scenario includes a timeline from the current moment to the end of the scheduling window, and discretely distributed technology activation events along this timeline. Each candidate scenario corresponds to a devaluation scenario hypothesis of the previous generation of robots under the impact of the next generation of technology: the earlier the technology activation and the higher the impact intensity, the greater the devaluation pressure on the inventory of the previous generation of robots, thus driving the scheduling strategy to optimize towards accelerated clearing; conversely, if the technology activation time in the candidate scenario is later or the impact intensity is lower, the scheduling strategy has more room to maintain a higher inventory level to cope with near-term demand.

[0116] By generating multiple candidate scenarios in batches, the scenario generation module provides the policy evaluation module with a diverse set of inputs covering different technological shock scenarios. The policy evaluation module calculates the comprehensive loss value based on these candidate scenarios and the state embedding vectors of each node, and feeds the optimized scheduling policy back to the scenario generation module. This drives the scenario generation module to focus on high-pressure scenario regions that result in greater policy loss in the next iteration, thus achieving a closed loop for adversarial training of the generative adversarial network. As the iterations progress, the candidate scenarios generated by the scenario generation module gradually concentrate on the weak areas of policy evaluation, prompting the scheduling policy to continuously optimize under the most challenging technological shock scenarios, ultimately converging to a target scheduling policy that is robust to the risk of intergenerational devaluation.

[0117] In one optional implementation, the policy evaluation module calculates a comprehensive loss value based on the candidate scenario and the state embedding vector of each node, including:

[0118] Based on the current scheduling strategy, the inventory holding time of each robot node within the candidate scenario time range is obtained. The inventory cost of each node is calculated based on the inventory holding time and the unit time maintenance cost of each node. The inventory costs of all nodes are summed and normalized to obtain the inventory cost item.

[0119] For the descendant technology activation events contained in the candidate scenarios, identify the generational evolution edge corresponding to the descendant technology activation event, trace back along the generational evolution edge to the previous generation robot node affected by the technological impact, obtain the state embedding vector of the previous generation robot node, and extract the feature components of the generational evolution dimension from the state embedding vector.

[0120] The feature components are fused with the generational weights of the corresponding generational evolution edges to calculate the unit value decay of each previous generation robot node at the time of activation of the next generation technology.

[0121] The predicted inventory quantity of the previous generation robot node at the time of activation of the next generation technology is obtained according to the current scheduling strategy. The unit value decay rate is multiplied by the predicted inventory quantity and the node unit price to obtain the intergenerational depreciation cost of each previous generation robot node. The intergenerational depreciation costs of all previous generation robot nodes are summed and normalized to obtain the intergenerational depreciation cost item.

[0122] The weighted sum of the inventory cost item and the intergenerational depreciation cost item yields the comprehensive loss value.

[0123] For example, after receiving the candidate scenarios output by the scenario generation module, the strategy evaluation module needs to quantitatively evaluate the inventory holding status of each robot node based on the time range of the candidate scenarios and the current scheduling strategy. Regarding the calculation of inventory costs, based on the inbound time and expected outbound time of each robot node in the current scheduling strategy, the actual inventory holding time of each node within the time range of the candidate scenarios is calculated. Let the nodes... The inventory holding duration within the candidate scenario time range is The corresponding unit time maintenance cost is Then the node The inventory cost is Maintenance cost per unit time The inventory cost includes multiple items such as storage occupancy fees, daily maintenance fees, and capital occupation costs, and is obtained from historical operational data of each robot model. After summing the inventory costs of all nodes, the total number of nodes is normalized to obtain the inventory cost item. The calculation method is as follows ,in This represents the total number of robot nodes participating in scheduling within the knowledge graph. The purpose of normalization is to eliminate the impact of differences in node size on the magnitude of the loss value, making the inventory cost item comparable to the intergenerational depreciation cost item, thereby ensuring the rationality of the subsequent weighted summation.

[0124] In calculating the intergenerational devaluation cost, candidate scenarios may include one or more descendant technology activation events. For each descendant technology activation event, the corresponding intergenerational evolution edge needs to be identified in the knowledge graph. The identification criteria for intergenerational evolution edges are: the descendant robot node identifier recorded in the activation event matches the terminating node of the intergenerational evolution edge. After identifying the target intergenerational evolution edge, the backtracking direction of the edge is traced to locate the set of predecessor robot nodes affected by the technological impact. The depth of the backtracking can be flexibly set according to the intergenerational hierarchy of the knowledge graph, usually tracing to the direct predecessor node. If there is a cross-generational impact, it can be extended to nodes of earlier generations, but the degree of technological impact should be reduced accordingly for each generation level traced upwards.

[0125] After obtaining the affected previous generation robot nodes, extract each previous generation node. State embedding vector State embedding vector Generated by the message passing mechanism of a graph neural network, it integrates node characteristics with graph topological location information, with each dimension corresponding to different semantic features. In calculating intergenerational depreciation costs, it is necessary to... The intergenerational evolution dimension feature components are extracted. These feature components are then processed using a predefined feature extraction matrix. Acting on We obtain the extracted feature component vector as follows: ,but ,in The intergenerational evolution feature extraction matrix is ​​obtained through backpropagation during the model training phase. Its row dimension corresponds to the feature dimension of the intergenerational evolution semantic space, and its column dimension is consistent with the dimension of the state embedding vector.

[0126] eigencomponents generational weights corresponding to the generational evolution edge Perform fusion calculations to compute the nodes of the previous generation robot. The unit value decay at the time of subsequent technology activation. The specific form of the fusion operation is: based on generational weights. For the feature component vector The weighted values ​​are then compressed into a scalar through a learnable linear mapping layer to obtain the unit value decay rate. ,Right now ,in These are the weight vectors of the linear mapping layer, jointly optimized with the overall network during the training phase. (Generational weights) The fusion result combines the technical parameter difference measurement value and the value decay coefficient. The larger the value, the more significant the value reduction of the previous generation robot after the activation of the next generation technology. Therefore, it plays a regulatory role in amplifying or reducing the feature components in the fusion calculation, so that the unit value decay can accurately reflect the degree of devaluation impact of the previous generation robot under different generational relationships.

[0127] After obtaining the unit value decay rate, further obtain the previous generation robot nodes. The predicted inventory quantity is based on the current scheduling strategy at the time of activation of the next-generation technology. And the unit price of the robot model corresponding to that node. Forecast inventory quantity The results, derived from the current scheduling strategy's projection over the candidate scenario timeline, reflect the number of previous-generation robots still in inventory at the activation time of the next-generation technology under this strategy. Multiplying the unit value decay rate, the predicted inventory quantity, and the node unit price yields the node... intergenerational devaluation costs ,Right now The physical meaning of this product form is clear: the unit value decay rate describes the proportion of value loss per robot under technological shock, the predicted inventory quantity reflects the number of affected robots, and the node unit price converts the value loss from a proportional dimension to a monetary dimension. The product of these three is the total depreciation loss suffered by the previous generation node at the time of technological shock.

[0128] Sum the intergenerational depreciation costs of all previous-generation robot nodes affected by the offspring technology activation event, and use the total number of affected nodes. After normalization, the intergenerational depreciation cost item is obtained. ,Right now ,in This represents the set of all previous-generation robot nodes affected by descendant technology activation events. When a candidate scenario contains multiple descendant technology activation events, the intergenerational depreciation cost corresponding to each event is calculated separately, summed, and then normalized to ensure that the intergenerational depreciation cost term maintains a consistent magnitude across different candidate scenario scales.

[0129] Overall loss value It is obtained by weighted summation of inventory cost and intergenerational depreciation cost, i.e. ,in and These are the combined weighting coefficients for inventory cost and intergenerational depreciation cost, respectively, which satisfy... And all of them are positive numbers. and The value is set based on the relative importance placed on inventory holding costs and the risk of depreciation due to technological iteration in the actual business scenario. It can be further adjusted during the training process after initialization through historical data fitting or expert experience. (Comprehensive loss value) As the core output of the policy evaluation module, it drives the current scheduling policy to update the gradient along the loss descent direction. On the other hand, it serves as an evaluation signal to feed back to the scene generation module, guiding the scene generation module to prioritize the generation of new candidate scenes that can expose the weaknesses of the current policy, thereby driving the entire adversarial generative network to continuously converge to the optimal scheduling policy during the iteration process.

[0130] In one optional implementation, the optimization strategy is fed back to the scene generation module to generate a new scene that drives further optimization of the strategy. This process is iteratively executed until convergence, after which the target scheduling strategy is output, including:

[0131] The strategy evaluation module performs gradient optimization on the parameters of the current scheduling strategy based on the comprehensive loss value, and generates an optimized scheduling strategy.

[0132] Extract the direction of the processing time adjustment of each robot node in the optimized scheduling strategy, and identify the robot nodes whose processing time is advanced as high-risk nodes;

[0133] The high-risk node identifier is fed back to the scene generation module. The scene generation module calculates the ratio of the number of high-risk nodes to the total number of nodes as the scene enhancement coefficient. For the generational evolution edge corresponding to the high-risk node, the original scene generation intensity is multiplied by the scene enhancement coefficient to obtain the enhanced scene generation intensity. According to the enhanced scene generation intensity, the descendant technology activation event for the high-risk node is generated to form a new scene.

[0134] The strategy evaluation module recalculates the comprehensive loss value based on the new scenario and continues to optimize the scheduling strategy, repeating the feedback and scenario generation process.

[0135] When the change in the comprehensive loss value is lower than the preset convergence threshold in multiple consecutive iterations, the optimization process is determined to have converged, and the current scheduling strategy is output as the target scheduling strategy.

[0136] For example, the policy evaluation module receives candidate scenarios and state embedding vectors of each node in each iteration, and calculates the comprehensive loss value. Then, gradient backpropagation is performed on all learnable parameters of the current scheduling policy. Specifically, with Using an automatic differentiation mechanism to calculate the gradients of each parameter as the objective function, an adaptive learning rate optimization algorithm (such as Adam) is employed to update the parameters, thereby generating an optimized scheduling strategy. The optimized scheduling strategy includes disposal decisions for each robot node in the knowledge graph, including key dimensions such as the suggested purchase quantity, inventory holdings, and disposal time window. The direction of change in the disposal time window directly reflects the strategy's judgment on the risk level of each node: if the disposal time of a node is advanced compared to the previous iteration, it indicates that the strategy believes the robot model corresponding to that node faces a high risk of intergenerational depreciation, requiring early clearance or inventory reduction to avoid losses.

[0137] After completing one round of gradient optimization and generating the optimized scheduling strategy, the processing time of each robot node in the current round is compared with that in the previous round. The set of nodes whose processing time has been advanced is identified and marked as high-risk nodes. The identification logic for high-risk nodes is based on the following technical judgment: driven by the gradient signal of the comprehensive loss value, the scheduling strategy will advance the processing time of nodes with greater intergenerational depreciation pressure to reduce... Therefore, the direction of advancing the processing time itself constitutes an implicit score for the node risk level. The identifiers of all nodes identified as high-risk (i.e., their node numbers in the knowledge graph) are organized into a high-risk node set and passed to the scene generation module.

[0138] After receiving the set of high-risk nodes, the scene generation module calculates the number of high-risk nodes and the total number of robot nodes participating in scheduling in the knowledge graph. The proportion of this percentage will be used as the scene enhancement coefficient. ,Right now ,in This represents the number of high-risk nodes identified in the current round. Scene enhancement coefficient. The physical meaning is that when the proportion of high-risk nodes is large, the uncertainty of the overall scheduling environment is high, and more candidate samples covering extreme technical activation scenarios need to be generated to fully expose the vulnerability of the strategy in these scenarios; when the proportion of high-risk nodes is small, the scenario enhancement is relatively mild to avoid over-generating redundant scenarios and wasting computing resources.

[0139] For each high-risk node, scenario reinforcement generation is performed along its corresponding generational evolution edge. Let the original scenario generation intensity be... The enhanced scene generation intensity for high-risk nodes is then... The scene generation module is based on Generate higher density of descendant technology activation events: On the generational evolution edges corresponding to high-risk nodes, extract from the technology activation time distribution at a higher sampling frequency. In conjunction with the technical impact intensity New candidate scenarios are constructed. Compared to the original scenarios, these new scenarios have a wider coverage and higher density of extreme cases in terms of the activation time and impact intensity of descendant technologies at high-risk nodes. This forces the policy evaluation module to optimize scheduling parameters under more stringent adversarial conditions, thereby improving the robustness of the target scheduling policy to generational technology impacts. For non-high-risk nodes, the scenario generation intensity remains unchanged. No additional enhancements are performed to maintain a balance between generation efficiency and optimized quality.

[0140] After receiving the new set of enhanced scenarios, the policy evaluation module recalculates the state embedding vector of each node (if the node features are updated due to scene changes), and recalculates the comprehensive loss value based on the new scenario. Gradient optimization is performed again to update the scheduling strategy parameters. The updated scheduling strategy is then fed back into the high-risk node identification process to extract the direction of action time adjustment, generating a new set of high-risk nodes, which is then fed back to the scene generation module to generate the next round of new scenes. The above feedback and scene generation process form a closed-loop iteration. Each iteration is guided by the gradient signal of the previous round of optimization strategy, dynamically adjusting the focus of scene generation, so that the two modules of the generative adversarial network continuously improve their respective capabilities in the game between each other.

[0141] Convergence is determined by monitoring the magnitude of change in the overall loss value. At the end of each iteration, the overall loss value for that iteration is recorded. ,in This is the current iteration round number. Calculate the magnitude of change in the comprehensive loss value over multiple consecutive iterations, specifically, using the most recent iteration as the starting point. Within round iteration The difference between the maximum and minimum values ​​is used as a volatility metric. When this volatility metric is lower than a preset convergence threshold... At this point, the optimization process is considered to have converged. Convergence threshold. The setting is based on the actual business requirements for scheduling accuracy, and is usually taken as a value in the range of one-thousandth to one-hundredth of the comprehensive loss value. The typical value is 3 to 5 rounds to avoid misjudging convergence due to occasional small fluctuations in the loss value of a single round.

[0142] Once the convergence condition is met, the scheduling strategy parameters for the current round are solidified and output as the target scheduling strategy. The target scheduling strategy includes the optimal disposal time window for all robot nodes in the knowledge graph, suggested upper and lower inventory limits, and procurement rhythm suggestions, which can be directly used to guide actual robot inventory management operations. Through the above iterative feedback mechanism, the scenario generation module and the strategy evaluation module achieve collaborative optimization during the adversarial process: the scenario generation module continuously strengthens scenario coverage for weak nodes (i.e., high-risk nodes) of the strategy, while the strategy evaluation module continuously strengthens its ability to cope with the risk of generational devaluation under more stringent scenario pressures. The final target scheduling strategy output has strong robustness and adaptability when facing the impact of subsequent technology activation in the real market.

[0143] A second aspect of this invention provides an AI-based intelligent scheduling system for robot inventory management, comprising:

[0144] The information acquisition unit is used to acquire information on the technological evolution and market transactions of the inventory robot.

[0145] The knowledge construction unit is used to construct a robot inventory knowledge graph, mapping each robot to a knowledge graph node, establishing generational evolution edges based on the technology evolution information and assigning generational weights, wherein the generational weights represent the value decay rate of the previous generation of robots, and establishing demand substitution edges based on market transaction information.

[0146] The vector generation unit is used to generate state embedding vectors for each node in the knowledge graph by means of the message passing mechanism of the graph neural network.

[0147] The strategy optimization unit is used to construct an adversarial generative network including a scenario generation module and a strategy evaluation module. The scenario generation module generates candidate scenarios containing subsequent generation technology activation events along the generational evolution edge. The strategy evaluation module calculates the comprehensive loss value and optimizes the scheduling strategy based on the candidate scenarios and the state embedding vectors of each node. The optimized strategy is fed back to the scenario generation module to generate new scenarios that drive further optimization of the strategy. The process is iteratively executed until convergence, and then the target scheduling strategy is output. The comprehensive loss value consists of an inventory cost item and a generational depreciation cost item.

[0148] A third aspect of the present invention provides an electronic device, comprising:

[0149] processor;

[0150] Memory used to store processor-executable instructions;

[0151] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0152] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0153] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based intelligent scheduling method for robot inventory management, characterized in that, include: Obtain information on the technological evolution and market transactions of inventory robots; A robot inventory management knowledge graph is constructed, mapping each robot to a knowledge graph node. Based on the technological evolution information, generational evolution edges are established and assigned generational weights. The generational weights represent the value decay rate of previous generation robots. At the same time, demand substitution edges are established based on market transaction information. For knowledge graphs, a state embedding vector for each node is generated based on the message passing mechanism of a graph neural network to integrate the graph topology position. An adversarial generative network is constructed, comprising a scene generation module and a strategy evaluation module. The scene generation module generates candidate scenes containing subsequent generation technology activation events along the intergenerational evolution edge. The strategy evaluation module calculates a comprehensive loss value and optimizes the scheduling strategy based on the candidate scenes and the state embedding vectors of each node. The optimized strategy is fed back to the scene generation module to generate new scenes that drive further optimization of the strategy. The process is iteratively executed until convergence, and then the target scheduling strategy is output. The comprehensive loss value consists of an inventory cost term and an intergenerational depreciation cost term.

2. The method according to claim 1, characterized in that, Based on the aforementioned technological evolution information, intergenerational evolution edges are established and assigned intergenerational weights. These weights characterize the rate of value decay of previous generations of robots. Simultaneously, based on market transaction information, demand substitution edges are established, including: Map each robot's identifier to a knowledge graph node; Based on the aforementioned technological evolution information, identify the previous generation robot nodes and the next generation robot nodes that have a technological iteration relationship, and establish an intergenerational evolution edge between the previous generation robot nodes and the next generation robot nodes. The initial generation weight is calculated based on the differences in technical parameters between the previous generation and the next generation in the technology evolution information. The price change data of the previous generation robot after the release of the next generation technology is obtained from the historical market transaction information. The value decay coefficient is extracted based on the time decay characteristics of the price change data. The value decay coefficient is fused with the initial generation weight to generate the final generation weight and assigned the corresponding generation evolution edge. Based on the market transaction information, robot node pairs with market competition relationships are identified, and demand substitution edges are established between the robot node pairs.

3. The method according to claim 2, characterized in that, The value decay coefficient is extracted based on the time decay characteristics of the price change data, including: The number of transaction records of previous generation robots in historical market transaction information is statistically analyzed in each time period after the release of the next generation robot technology. The length of the time window is dynamically determined based on the distribution density of the number of transaction records. The length of the time window is inversely proportional to the distribution density of the number of transaction records. Identify abnormal data points where the price change exceeds a preset deviation threshold within each time window, determine a weighting coefficient based on the degree of deviation of the abnormal data points, and perform weighted processing or remove the abnormal data points. The local decay rate is calculated for the processed price change data within each time window, and a moving average mechanism is used to smooth the local decay rate of adjacent time windows over time. When the number of valid samples of price change data is lower than a preset sparsity threshold, the average decay rate of other robots of the same technology generation is obtained as the generational decay prior mode. The generational decay prior mode is then weighted and fused with the local decay rate of the previous generation robot in inverse proportion to the number of samples to generate a value decay coefficient.

4. The method according to claim 1, characterized in that, For knowledge graphs, a state embedding vector for each node is generated based on the message passing mechanism of a graph neural network, including: For each node in the knowledge graph, aggregate the features of its neighboring nodes on the generational evolution edge and the features of its neighboring nodes on the demand substitution edge. The characteristics of adjacent nodes on the generational evolution edge are assigned generational propagation weights, so that the generational propagation weights are positively correlated with the generational weights of the corresponding generational evolution edge. The characteristics of adjacent nodes on the demand substitution edge are assigned competitive propagation weights. The weighted features of adjacent nodes of the generational evolution edge and the features of adjacent nodes of the demand replacement edge are fused and aggregated to obtain the neighborhood aggregation features of the node. The initial features of the node are nonlinearly transformed with the neighborhood aggregation features to generate a state embedding vector.

5. The method according to claim 1, characterized in that, The scene generation module generates candidate scenes containing subsequent technology activation events along the generational evolution edge, including: Identify the previous generation robot nodes in the knowledge graph that have descendant robot nodes, and use the previous generation robot nodes as the starting nodes for generating candidate scenarios; for each starting node, trace along its generational evolution edge to the corresponding descendant robot node, and obtain the technical parameters and historical technology release cycle data of the descendant robot node. An initial probability distribution of subsequent technology activation events is constructed based on the aforementioned technical parameters and historical technology release cycle data. The subsequent technology activation events include the predicted release time and technical performance indicators of the subsequent technologies. Extract the generational weights of the generational evolution edges, and use the generational weights as a time adjustment factor to scale and transform the time axis of the initial probability distribution so that the predicted release time is negatively correlated with the generational weights. The initial probability distribution after scaling transformation is sampled to generate the technology activation time point of the descendant, and the technical impact intensity corresponding to the activation time point is determined according to the technical performance index of the descendant robot node. The time series containing the activation time points and technological impact intensity of the descendant technologies are constructed as candidate scenarios.

6. The method according to claim 1, characterized in that, The strategy evaluation module calculates a comprehensive loss value based on the candidate scenarios and the state embedding vectors of each node, including: Based on the current scheduling strategy, the inventory holding time of each robot node within the candidate scenario time range is obtained. The inventory cost of each node is calculated based on the inventory holding time and the unit time maintenance cost of each node. The inventory costs of all nodes are summed and normalized to obtain the inventory cost item. For the descendant technology activation events contained in the candidate scenarios, identify the generational evolution edge corresponding to the descendant technology activation event, trace back along the generational evolution edge to the previous generation robot node affected by the technological impact, obtain the state embedding vector of the previous generation robot node, and extract the feature components of the generational evolution dimension from the state embedding vector. The feature components are fused with the generational weights of the corresponding generational evolution edges to calculate the unit value decay of each previous generation robot node at the time of activation of the next generation technology. The predicted inventory quantity of the previous generation robot node at the time of activation of the next generation technology is obtained according to the current scheduling strategy. The unit value decay rate is multiplied by the predicted inventory quantity and the node unit price to obtain the intergenerational depreciation cost of each previous generation robot node. The intergenerational depreciation costs of all previous generation robot nodes are summed and normalized to obtain the intergenerational depreciation cost item. The weighted sum of the inventory cost item and the intergenerational depreciation cost item yields the comprehensive loss value.

7. The method according to claim 1, characterized in that, The optimization strategy is fed back to the scene generation module to generate new scenes that drive further optimization of the strategy. This process is iterated until convergence, at which point the target scheduling strategy is output, including: The strategy evaluation module performs gradient optimization on the parameters of the current scheduling strategy based on the comprehensive loss value, and generates an optimized scheduling strategy. Extract the direction of the processing time adjustment of each robot node in the optimized scheduling strategy, and identify the robot nodes whose processing time is advanced as high-risk nodes; The high-risk node identifier is fed back to the scene generation module. The scene generation module calculates the ratio of the number of high-risk nodes to the total number of nodes as the scene enhancement coefficient. For the generational evolution edge corresponding to the high-risk node, the original scene generation intensity is multiplied by the scene enhancement coefficient to obtain the enhanced scene generation intensity. According to the enhanced scene generation intensity, the descendant technology activation event for the high-risk node is generated to form a new scene. The strategy evaluation module recalculates the comprehensive loss value based on the new scenario and continues to optimize the scheduling strategy, repeating the feedback and scenario generation process. When the change in the comprehensive loss value is lower than the preset convergence threshold in multiple consecutive iterations, the optimization process is determined to have converged, and the current scheduling strategy is output as the target scheduling strategy.

8. An AI-based intelligent scheduling system for robot inventory management, used to implement the method as described in any one of claims 1-7, characterized in that, include: The information acquisition unit is used to acquire information on the technological evolution and market transactions of the inventory robot. The knowledge construction unit is used to construct a robot inventory knowledge graph, mapping each robot to a knowledge graph node, establishing generational evolution edges based on the technology evolution information and assigning generational weights, wherein the generational weights represent the value decay rate of the previous generation of robots, and establishing demand substitution edges based on market transaction information. The vector generation unit is used to generate state embedding vectors for each node in the knowledge graph by means of the message passing mechanism of the graph neural network. The strategy optimization unit is used to construct an adversarial generative network including a scenario generation module and a strategy evaluation module. The scenario generation module generates candidate scenarios containing subsequent generation technology activation events along the generational evolution edge. The strategy evaluation module calculates the comprehensive loss value and optimizes the scheduling strategy based on the candidate scenarios and the state embedding vectors of each node. The optimized strategy is fed back to the scenario generation module to generate new scenarios that drive further optimization of the strategy. The process is iteratively executed until convergence, and then the target scheduling strategy is output. The comprehensive loss value consists of an inventory cost item and a generational depreciation cost item.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images