An automobile part demand prediction method and system based on a knowledge graph

Abstract

The application discloses a kind of based on knowledge graph's automobile parts demand prediction method and system, comprising: collecting multi-source heterogeneous data, generate standardized multi-source heterogeneous data;Based on standardized multi-source heterogeneous data, generate automobile parts knowledge graph;Divide demand prediction scene, generate document word co-occurrence matrix;Improved PLSA topic model is constructed, and demand theme distribution is obtained;Optimal transmission mapping relationship is obtained using Sinkhorn algorithm, and aligned demand theme distribution representation is generated;Unified demand representation vector is constructed, and the demand prediction result of part is output.The application realizes the stable, comparable and transferable demand prediction of automobile parts in different regions, vehicle types and use scenarios by introducing improved PLSA topic model and Sinkhorn algorithm.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to a method and system for predicting the demand for automotive parts based on knowledge graphs. Background Technology

[0002] With the continuous expansion of the automotive industry and the increasing segmentation of vehicle configurations, the variety, compatibility, and usage scenarios of automotive parts are becoming increasingly complex, resulting in significant diversity and dynamic changes in demand. Traditional automotive parts demand forecasting methods primarily rely on historical sales data, maintenance statistics, or simple time series analysis models to extrapolate or statistically estimate demand trends. These methods typically assume that demand changes are stable or linear, making it difficult to fully characterize the structural differences in demand formation processes across different regions, vehicle models, and usage scenarios. Traditional methods often process multi-source data through simple aggregation, ignoring the complex relationships between parts, vehicle models, events, and scenarios, leading to insufficient stability and limited accuracy in cross-regional or cross-scenario applications.

[0003] Some existing technologies have begun to incorporate knowledge graphs or topic models for automotive parts demand analysis. These technologies enhance the expressive power of demand forecasting by constructing relationship networks between parts and vehicle models, faults, and maintenance events, or by mining the potential structure of demand through topic modeling. However, these technologies often focus on demand modeling within a single scenario or time scale. Significant differences often exist in the distribution of demand topics across different regions, vehicle models, or usage scenarios, and effective structural alignment and consistency constraints are lacking. This makes it difficult to directly compare or transfer demand topics across different scenarios, limiting the generalization ability and stable application performance of demand forecasting models under multi-scenario conditions.

[0004] Therefore, how to provide a knowledge graph-based method and system for predicting the demand for automotive parts is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a knowledge graph-based method and system for predicting automotive parts demand. This invention comprehensively utilizes knowledge graph modeling, topic modeling, and optimal transport computation methods to perform unified modeling and alignment analysis of the complex demand behaviors of automotive parts formed under different regions, vehicle models, and usage scenarios. By constructing an automotive parts knowledge graph, it systematically depicts the structural relationships in the demand formation process. An improved PLSA topic model is introduced to learn the semantics of demand behaviors under multiple scenarios and time scales, forming a stable distribution of potential demand topics. The Sinkhorn algorithm is used to perform entropy regularization optimal transport alignment on the demand topic distributions under different scenarios, achieving consistent expression and comparable analysis of demand structures across scenarios. This invention describes and implements structured modeling, cross-scenario alignment, and unified representation of heterogeneous demand data, overcoming the problems of unstable prediction and weak transferability in existing demand prediction methods under multiple scenario conditions. It possesses the advantages of high consistency of demand prediction results, strong adaptability across regions and vehicle models, and high prediction reliability.

[0006] A knowledge graph-based method for predicting automotive parts demand according to an embodiment of the present invention includes: Collect multi-source heterogeneous data related to automotive parts demand forecasting, preprocess the multi-source heterogeneous data, and generate standardized multi-source heterogeneous data. Based on standardized multi-source heterogeneous data, we extract entities such as parts, vehicle models, usage scenarios, regions, events, and time, and construct assembly relationships, adaptation relationships, triggering relationships, and attribution relationships between entities to generate an automotive parts knowledge graph. Demand prediction scenarios are divided according to the combination of region, vehicle type and usage scenario. In each demand prediction scenario, demand-related event subgraphs are extracted from the automotive parts knowledge graph, a set of demand behavior semantic units is constructed, and a document term co-occurrence matrix is ​​generated. An improved PLSA topic model is constructed, and topic learning is performed on the document term co-occurrence matrix under various demand prediction scenarios to obtain the corresponding potential demand topic distribution and the corresponding component probability distribution, which are then encapsulated as demand topic distribution. The Sinkhorn algorithm is used to perform entropy regularization optimal transmission calculation on the demand topic distribution and alignment target of each scene to be aligned, to obtain the optimal transmission mapping relationship and generate an alignment demand topic distribution representation; Based on the aligned demand topic distribution representation, and combined with component attribute features, event frequency features, and time features, a unified demand representation vector is constructed to output the demand prediction results for components.

[0007] Optionally, the multi-source heterogeneous data specifically includes basic attribute data of components, vehicle configuration information, vehicle operating condition data, maintenance and repair record data, fault event data, and regional and store data.

[0008] Optionally, the preprocessing of multi-source heterogeneous data specifically includes deduplication, missing data completion, anomaly removal, time alignment, and unified encoding processing.

[0009] Optionally, generating the automotive parts knowledge graph includes: Based on standardized multi-source heterogeneous data, we extract component entities, vehicle entities, usage scenario entities, region entities, event entities, and time entities to form a set of entity nodes. For the set of entity nodes, assembly and adaptation relationships are established between component entities and vehicle entities, triggering relationships are established between event entities and component entities, and attribution relationships are established between store entities and regional entities, forming a set of relationship edges; Within a unified time window, the number of occurrences of various relationships between corresponding entity pairs is counted, and a corresponding relationship weight is assigned to each relationship based on the relative frequency of occurrence of each relationship among similar relationships. A knowledge graph of automotive parts is generated based on the set of entity nodes, the set of relation edges, and the corresponding relation weights.

[0010] Optionally, the generation of the document term co-occurrence matrix includes: Based on regional and store data, vehicle configuration information and vehicle operating condition data, regional dimension identifiers, vehicle dimension identifiers and usage scenario dimension identifiers are determined respectively, and the three types of identifiers are combined to determine a demand prediction scenario, forming a set of demand prediction scenarios. For each demand forecasting scenario, based on the timestamp information corresponding to the time entity, the maintenance record data and fault event data within the demand forecasting scenario are sorted by time, and the time axis is divided according to the time window length to generate a continuous time window. Within each time window of each demand forecasting scenario, maintenance and repair event entities and fault event entities that occur within the time window are retrieved from the automotive parts knowledge graph. The event entities are used as starting nodes to expand the associated parts entities and vehicle models along the assembly relationship, adaptation relationship and triggering relationship to generate a demand-related event subgraph. Each demand-related event subgraph corresponding to each time window in each demand forecasting scenario is encapsulated into a demand behavior semantic unit, and each demand behavior semantic unit is defined as a document unit. Each document unit contains a set of component entities and the corresponding number of events. By using component entities as term dimensions and each requirement behavior semantic unit as document dimensions, a document term co-occurrence matrix is ​​generated based on the number of events occurring in each component entity within each requirement behavior semantic unit.

[0011] Optionally, the encapsulation as a demand topic distribution includes: An improved PLSA topic model is constructed. Under various demand prediction scenarios, the set of semantic units of demand behavior is sorted, and short-granular semantic windows, medium-granular semantic windows and long-granular semantic windows are constructed based on different time spans to form a multi-scale semantic window set. Based on a multi-scale semantic window set, a consistency guidance structure is introduced. The statistical results of demand semantics formed in the long-granularity semantic window are used as constraint information to guide the weight distribution of demand behavior semantic units in the short-granularity and medium-granularity semantic windows. At the same time, the demand semantic units that frequently appear in the short-granularity semantic window are fed back to the medium-granularity and long-granularity semantic windows, forming a cross-scale bidirectional constraint semantic consistency guidance relationship. Based on semantic consistency guidance, corresponding PLSA sub-models are constructed for short-granularity semantic windows, medium-granularity semantic windows, and long-granularity semantic windows, respectively. Based on the document term co-occurrence matrix, topic learning is performed in each PLSA sub-model. A cross-semantic scale consistency guidance structure is introduced during the topic learning process to constrain and smooth the update process of the demand topic distribution and component probability distribution at each scale, thereby obtaining the corresponding demand topic distribution sets under short-granularity, medium-granularity and long-granularity semantic windows. The distribution of demand topics under each semantic scale is compared in the continuous time dimension. The degree of change in the distribution of demand topics within adjacent time periods is calculated. An adaptive semantic time period positioning structure is introduced. When the degree of change in the distribution of demand topics exceeds the threshold, the corresponding time point is determined as a semantic change point, and the time axis is re-divided to generate an adaptive semantic time period. Based on adaptive semantic time periods, the demand topic distributions formed under short-granularity, medium-granularity, and long-granularity semantic windows are integrated to obtain the corresponding demand topic distributions.

[0012] Optionally, the generation of the alignment requirement topic distribution representation includes: Based on the demand topic distribution, the Sinkhorn algorithm is executed to determine the set of demand topic distributions for the scenarios to be aligned. Integrity checks are performed on the target demand topic distribution. When missing topics or abnormal topic weights are detected in the target demand topic distribution, an adaptive reconstruction structure for the alignment target is introduced. Based on the historical stable demand topic distribution set, the target demand topic distribution is reconstructed to generate a reconstructed target demand topic distribution. Based on information about region, vehicle type, and usage scenario, similarity is determined for the scenarios to be aligned. A positive and negative feedback channel is introduced. The distribution of demand topics that meet the consistency conditions with the scenarios to be aligned in terms of region, vehicle type, and usage scenario is taken as the set of positive demand topic distributions, and the distribution of demand topics that do not meet the consistency conditions is taken as the set of negative demand topic distributions. Based on the distribution of demand topics in the scenario to be aligned, the distribution of demand topics for reconstruction, and the set of positive and negative sample demand topics, the degree of topic difference between each potential demand topic is calculated, a topic transmission cost table is constructed, and a source-end topic weight sequence, a target-end topic weight sequence, and a negative constraint weight sequence are generated. Based on the topic transmission cost table, the source topic weight sequence, the target topic weight sequence, and the negative constraint weight sequence, the topic transmission matrix is ​​initialized, and iterative normalization update processing is performed until the convergence criterion is met, generating the optimal topic transmission matrix and the corresponding optimal transmission mapping relationship. Based on the optimal topic transfer matrix, the topic distribution of the scene to be aligned is mapped to a topic space consistent with the topic distribution of the reconstruction target, thereby generating a representation of the topic distribution of the alignment requirements.

[0013] Optionally, the demand forecast results for the output components include: Based on the alignment demand topic distribution representation, read the alignment demand topic weight sequence corresponding to the demand prediction scenario; Based on the basic attribute data of parts and vehicle configuration information, the attribute fields of each part entity and the set of compatible vehicle models are read from the automotive parts knowledge graph to generate a set of part attribute features. Then, the encoding is unified and normalized to obtain a standardized set of part attribute feature vectors. Based on maintenance and repair records and fault event data, the number of event triggers, maintenance occurrences and fault occurrences of each component entity are counted within the time range corresponding to the demand prediction scenario. A set of component event frequency features is generated, and time alignment and window aggregation processing are performed to obtain a set of component frequency feature vectors. Based on vehicle operation data and regional and store data, time and regional features corresponding to demand prediction scenarios are extracted and then spliced ​​and fused with the aligned demand theme weight sequence, standardized component attribute feature vector set and component frequency feature vector set to generate a unified demand representation vector. Based on a unified demand representation vector, the demand forecast values ​​for each component are output according to the component dimension, and the demand forecast results for each component are output.

[0014] According to an embodiment of the present invention, an automotive parts demand forecasting system based on a knowledge graph includes the following modules: The data acquisition module is used to collect multi-source heterogeneous data related to the forecasting of automotive parts demand. The data preprocessing module is used to preprocess multi-source heterogeneous data and generate standardized multi-source heterogeneous data. The knowledge graph construction module is used to generate automotive parts knowledge graphs based on standardized multi-source heterogeneous data. The demand scenario segmentation module is used to segment demand forecasting tasks according to region, vehicle type, and usage scenario, and generate demand forecasting scenarios. The semantic unit construction module is used to extract event subgraphs from the automotive parts knowledge graph and construct a set of demand behavior semantic units in various demand prediction scenarios. The document term co-occurrence modeling module is used to generate a document term co-occurrence matrix based on the set of semantic units of demand behavior. The demand topic modeling module is used to build an improved PLSA topic model, perform topic learning processing, and obtain the demand topic distribution. The demand topic alignment module is used to generate an aligned demand topic distribution representation by executing the Sinkhorn algorithm based on the demand topic distribution. The demand feature representation module is used to construct a demand representation vector by integrating component attribute features, event frequency features, and time features based on the aligned demand topic distribution representation. The demand forecasting output module is used to output the demand forecasting results for each component based on the demand representation vector.

[0015] The beneficial effects of this invention are: This invention proposes a knowledge graph-based method and system for predicting automotive parts demand. It comprehensively employs multi-source data structured modeling, topic semantic modeling, and optimal transport alignment calculation to uniformly model and predict the demand behavior of automotive parts across different regions, vehicle models, and usage scenarios. An automotive parts knowledge graph is constructed to structurally characterize the relationships between various elements in the demand formation process. Demand prediction scenarios are divided by region, vehicle model, and usage scenario, and demand-related event subgraphs are extracted within a unified time window to construct semantic units of demand behavior and a document term co-occurrence matrix.

[0016] This invention introduces an improved PLSA topic model, performing topic learning on demand behavior semantics from multiple time scales to uncover potential demand topic structures under different demand prediction scenarios, avoiding the incomplete demand expression caused by relying solely on a single time scale or scenario. By introducing entropy regularization optimal transport calculation based on the Sinkhorn algorithm, the demand topic distributions formed under different scenarios are structurally aligned, mapping the originally significantly different demand topic distributions to a unified topic space, achieving consistent demand structure expression across regions, vehicle models, and usage scenarios. The aligned demand topic distribution is then fused with component attribute features, event frequency features, and time features to construct a unified demand representation vector and output prediction results. This invention effectively solves the problems of lack of comparability of demand distributions across multiple scenarios, difficulty in transferring prediction results, and insufficient stability in existing automotive component demand prediction methods. It achieves structured expression of complex demand behavior and consistent prediction across scenarios, significantly improving the reliability, adaptability, and application value of demand prediction results. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a knowledge graph-based method for predicting the demand for automotive parts proposed in this invention. Figure 2 This is a functional flowchart of an improved PLSA topic model for a knowledge graph-based automotive parts demand forecasting method proposed in this invention. Figure 3 This is a schematic diagram of the structure of a knowledge graph-based automotive parts demand prediction system proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 and Figure 2 A knowledge graph-based method for predicting automotive parts demand includes: Collect multi-source heterogeneous data related to automotive parts demand forecasting, preprocess the multi-source heterogeneous data, and generate standardized multi-source heterogeneous data. Based on standardized multi-source heterogeneous data, we extract entities such as parts, vehicle models, usage scenarios, regions, events, and time, and construct assembly relationships, adaptation relationships, triggering relationships, and attribution relationships between entities to generate an automotive parts knowledge graph. Demand prediction scenarios are divided according to the combination of region, vehicle type and usage scenario. In each demand prediction scenario, demand-related event subgraphs are extracted from the automotive parts knowledge graph, a set of demand behavior semantic units is constructed, and a document term co-occurrence matrix is ​​generated. An improved PLSA topic model is constructed, and topic learning is performed on the document term co-occurrence matrix under various demand prediction scenarios to obtain the corresponding potential demand topic distribution and the corresponding component probability distribution, which are then encapsulated as demand topic distribution. The Sinkhorn algorithm is used to perform entropy regularization optimal transmission calculation on the demand topic distribution and alignment target of each scene to be aligned, to obtain the optimal transmission mapping relationship and generate an alignment demand topic distribution representation; Based on the aligned demand topic distribution representation, and combined with component attribute features, event frequency features, and time features, a unified demand representation vector is constructed to output the demand prediction results for components.

[0020] In this embodiment, the multi-source heterogeneous data specifically includes basic attribute data of components, vehicle configuration information, vehicle operating condition data, maintenance and repair record data, fault event data, and regional and store data.

[0021] In this embodiment, the preprocessing of multi-source heterogeneous data specifically includes deduplication, missing data completion, anomaly removal, time alignment, and encoding unification.

[0022] In this embodiment, generating the automotive parts knowledge graph includes: Based on standardized multi-source heterogeneous data, we extract component entities, vehicle entities, usage scenario entities, region entities, event entities, and time entities to form a set of entity nodes. For the set of entity nodes, assembly and adaptation relationships are established between component entities and vehicle entities, triggering relationships are established between event entities and component entities, and attribution relationships are established between store entities and regional entities, forming a set of relationship edges; Within a unified time window, the frequency of occurrence of various relationships between corresponding entity pairs is counted, and based on the relative frequency of each relationship within the same type of relationship, a corresponding relationship weight is assigned to each relationship, where: Assign a corresponding relation weight to each relation, specifically as follows: Within a unified time window, all relation records belonging to the same relation type are grouped by entity pairs. The number of relation records appearing for each entity pair within the time window is accumulated to obtain the number of relation occurrences for the entity pair within the current time window. The number of relation occurrences for all entity pairs under the same relation type is summed to obtain the total number of occurrences for the relation type within the current time window. The number of relation occurrences for each entity pair is divided by the total number of occurrences for the relation type to obtain the relative frequency of occurrence of the corresponding relation of the entity pair within the same type of relation. The relative frequency of occurrence is used as the relation weight between entity pairs of that relation type. Based on the set of entity nodes, the set of relation edges, and the corresponding relation weights, a knowledge graph of automotive parts is generated, where: Generate a knowledge graph of automotive parts, specifically: Each entity node is assigned a unique node identifier based on the entity node set, and the entity type and entity attributes are written into the node attribute field. The connection relationship between nodes is established one by one according to the entity pair identifiers recorded in the relationship edge set, and the corresponding relationship type identifier and relationship weight are written into the attribute field of the connection edge. After all nodes and relationship edges are written, an automotive parts knowledge graph containing entity nodes, relationship edges, relationship types and relationship weight information is formed.

[0023] In this embodiment, generating the document term co-occurrence matrix includes: Based on regional and store data, vehicle configuration information, and vehicle operating condition data, regional dimension identifiers, vehicle model dimension identifiers, and usage scenario dimension identifiers are determined respectively. These three types of identifiers are then combined to form a demand forecasting scenario set, where: A set of demand forecasting scenarios is formed, specifically: The system reads the regional identifier corresponding to each store from the regional and store data as the regional dimension identifier, reads the vehicle model code corresponding to the vehicle from the vehicle configuration information as the vehicle model dimension identifier, and extracts the scenario classification results corresponding to the vehicle's operating status, usage environment, and operating parameters from the vehicle operating condition data as the usage scenario dimension identifier. The system combines the regional dimension identifier, vehicle model dimension identifier, and usage scenario dimension identifier corresponding to the same data record in a fixed order to generate a unique scenario combination identifier. Different demand prediction scenarios are distinguished by different scenario combination identifiers, forming a demand prediction scenario set composed of multiple scenario combination identifiers. For each demand forecasting scenario, based on the timestamp information corresponding to the time entity, the maintenance record data and fault event data within the demand forecasting scenario are sorted by time, and the time axis is divided according to the time window length to generate a continuous time window. Within each time window of each demand forecasting scenario, maintenance and repair event entities and fault event entities occurring within the time window are retrieved from the automotive parts knowledge graph. Using these event entities as starting nodes, associated part entities and vehicle models are expanded along assembly, adaptation, and triggering relationships to generate a demand-related event subgraph, where: Generate a subgraph of demand-related events, specifically as follows: Within the current time window of the current demand forecasting scenario, maintenance and repair event entities and fault event entities whose occurrence time falls within the time window are selected from the automotive parts knowledge graph. The event entities are used as the starting nodes of the subgraph. Centered on the starting node, the parts entities directly associated with the event entities are retrieved along the established triggering relationships in the knowledge graph. Furthermore, the vehicle models associated with the parts entities are retrieved along the assembly and adaptation relationships. Only the nodes and relationships between entities that have associated records under the current demand forecasting scenario are retained. Finally, the starting event entity, the expanded parts entities and vehicle models, and the relationships between them are combined to form the demand-related event subgraph. Each demand-related event subgraph corresponding to each time window in each demand forecasting scenario is encapsulated into a demand behavior semantic unit, and each demand behavior semantic unit is defined as a document unit. Each document unit contains a set of component entities and the corresponding number of events. Using component entities as the term dimension and each requirement behavior semantic unit as the document dimension, a document term co-occurrence matrix is ​​generated based on the number of events occurring for each component entity within each requirement behavior semantic unit, where: Generate a document term co-occurrence matrix, specifically as follows: Using component entities as term sets and each requirement behavior semantic unit as document set, the corresponding component entity set and the number of events of each component entity within the time window are read one by one for each requirement behavior semantic unit. With document units as rows and component entities as columns, the number of events of the corresponding component entities in each document unit is written into the corresponding matrix unit position. When no event occurs for a component entity in the corresponding document unit, the value of the corresponding matrix unit is set to zero, forming a complete document term co-occurrence matrix.

[0024] In this embodiment, the encapsulation as a demand topic distribution includes: An improved PLSA topic model is constructed. Under various demand forecasting scenarios, the set of semantic units representing demand behaviors is ranked, and short-granularity, medium-granularity, and long-granularity semantic windows are constructed based on different time spans, forming a multi-scale semantic window set. A multi-scale semantic window set is formed, specifically as follows: In the current demand forecasting scenario, the set of demand behavior semantic units is sorted according to time sequence. Based on the sorted demand behavior semantic units, short-granularity semantic windows, medium-granularity semantic windows, and long-granularity semantic windows are determined according to the time span parameter. The short-granularity semantic window covers 3 consecutive demand behavior semantic units, the medium-granularity semantic window covers 4 consecutive short-granularity semantic windows, and the long-granularity semantic window covers 5 consecutive medium-granularity semantic windows, forming a multi-scale semantic window structure from short-term demand changes to long-term demand trends. Based on a multi-scale semantic window set, a consistency guidance structure is introduced. The statistical results of demand semantics formed in the long-granularity semantic window are used as constraint information to guide the weight distribution of demand behavior semantic units in the short-granularity and medium-granularity semantic windows. Simultaneously, frequently occurring demand semantic units in the short-granularity semantic window are fed back to the medium-granularity and long-granularity semantic windows, forming a cross-scale, bidirectional constraint semantic consistency guidance relationship. A consistency-guided structure refers to establishing cross-scale constraints between multi-scale semantic windows, so that the distribution of demand semantics formed at different time scales maintains comparability and continuity in terms of overall trends and local changes. The weight distribution of demand behavior semantic units within short-granularity and medium-granularity semantic windows is guided as follows: Within the long-granularity semantic window, the cumulative frequency of each demand behavior semantic unit within the corresponding time range is statistically analyzed to form a long-granularity semantic statistical result. In the short-granularity and medium-granularity semantic windows, the initial weights of the corresponding demand behavior semantic units are adjusted and brought closer to the relative proportions in the long-granularity semantic statistical result during the update process. For demand behavior semantic units whose frequency of occurrence in the short-granularity semantic window is higher than the threshold, the weight information is fed back to the medium-granularity and long-granularity semantic windows to correct the weight distribution of the corresponding semantic units at a larger time scale, forming a cross-scale bidirectional weight constraint relationship. Based on semantic consistency guidance, corresponding PLSA sub-models are constructed for short-granularity semantic windows, medium-granularity semantic windows, and long-granularity semantic windows, respectively. Based on the document term co-occurrence matrix, topic learning is performed in each PLSA sub-model. A cross-semantic scale consistency guidance structure is introduced during topic learning to constrain and smooth the update process of demand topic distribution and component probability distribution at each scale, resulting in demand topic distribution sets corresponding to short-granularity, medium-granularity, and long-granularity semantic windows. Topic learning processing is performed in each PLSA sub-model, specifically as follows: In the PLSA sub-model corresponding to each semantic scale, the document term co-occurrence matrix is ​​used as input to initialize the demand topic distribution and component probability distribution. In each iteration, the weights of each demand behavior semantic unit to different topics are recalculated based on the current topic distribution and component probability distribution, and the topic distribution and component probability distribution are updated until the iteration termination condition is met, where: The iteration termination condition is that in two consecutive iterations, the change in the weight of each topic in the demand topic distribution is less than the first threshold, and the change in the probability of each component under the corresponding topic in the component probability distribution is less than the second threshold. Cross-semantic scale consistency guidance structure refers to a structure that introduces topic distribution information from semantic windows of other scales during the iterative update process of PLSA sub-models corresponding to different semantic scales, and imposes constraints on the update magnitude and direction of topic distribution at the current scale. The distribution sets of demand topics corresponding to short-granularity, medium-granularity, and long-granularity semantic windows are obtained, specifically as follows: After each PLSA sub-model completes its iterative update, the corresponding demand topic distribution results under the short-granularity semantic window, medium-granularity semantic window and long-granularity semantic window are output respectively, and encapsulated according to semantic scale and time order to form a demand topic distribution set composed of demand topic distributions at multiple scales. The distribution of demand topics at various semantic scales is compared along a continuous time dimension. The degree of change in the distribution of demand topics within adjacent time periods is calculated, and an adaptive semantic time period positioning structure is introduced. When the degree of change in the distribution of demand topics exceeds a threshold, the corresponding time point is determined as a semantic change point, and the time axis is re-divided to generate an adaptive semantic time period, wherein: Calculate the degree of change in the distribution of demand themes within adjacent time periods, specifically as follows: In the continuous time dimension, for the distribution of demand themes in two adjacent time periods, the differences in the distribution ratio of each theme in the two time periods are compared one by one, and the differences in the distribution of each theme are summarized to obtain a change index reflecting the overall degree of change of demand themes in adjacent time periods. Semantic time period adaptive positioning structure refers to a structure that dynamically adjusts the original time division based on the degree of change in the distribution of demand topics in the time dimension, identifies the time position where the semantics of demand change significantly and uses it as the new time segment boundary. Generate adaptive semantic time periods, specifically: When the change in the distribution of demand topics within adjacent time periods exceeds the threshold, the corresponding time point is marked as a semantic change point, and the original time axis is re-divided with the semantic change point as the boundary, ultimately forming an adaptive semantic time period composed of multiple semantically stable intervals. Based on adaptive semantic time periods, the demand topic distributions formed under short-granularity, medium-granularity, and long-granularity semantic windows are integrated to obtain the corresponding demand topic distributions, where: The corresponding demand topic distribution is obtained as follows: Within each adaptive semantic time period, the distribution results of demand topics formed under short-granularity, medium-granularity, and long-granularity semantic windows are summarized respectively, and the topic distributions under different semantic scales are integrated to generate a demand topic distribution with a unified semantic representation within the current adaptive semantic time period.

[0025] In this embodiment, generating the alignment requirement topic distribution representation includes: Based on the demand topic distribution, the Sinkhorn algorithm is executed to determine the set of demand topic distributions for the scenarios to be aligned. Integrity checks are performed on the target demand topic distribution. When missing topics or abnormal topic weights are detected, an adaptive reconstruction structure for the alignment target is introduced. Based on a historically stable set of demand topic distributions, the target demand topic distribution is reconstructed to generate a reconstructed target demand topic distribution, where: The scenario to be aligned refers to the demand forecasting scenario corresponding to the region, vehicle type and usage scenario combination that needs to be mapped with the selected alignment target in the current demand forecasting process, where the demand theme distribution needs to be consistent with the theme space. The target demand topic distribution refers to the demand topic distribution that serves as an alignment reference benchmark. It is derived from the demand topic distribution results in historical demand forecasting scenarios where the topic structure is stable, the data is complete, and the time continuity requirement is met. Alignment target adaptive reconstruction structure refers to a structure that completes and corrects the theme composition and weight structure of the target demand theme distribution based on a set of historical stable demand theme distributions when there are missing themes or abnormal theme weights in the target demand theme distribution. Generate a distribution of refactoring target requirements topics, specifically as follows: When it is detected that there are potential demand topics that have not appeared in the target demand topic distribution, or that the weights of some topics deviate from the historical stable range, the average weight distribution of the corresponding topics is extracted from the historical stable demand topic distribution set. The missing topics are filled in according to the topic index, and the weights of abnormal topics are corrected to form a reconstructed target demand topic distribution with a complete topic set and a continuous weight distribution. Based on region, vehicle model, and usage scenario information, a similarity determination is made for the scenarios to be aligned. A dual-sided feedback channel of positive and negative samples is introduced. The distribution of demand topics that meet the consistency conditions with the scenarios to be aligned in terms of region, vehicle model, and usage scenario is designated as the positive sample demand topic distribution set, while the distribution of demand topics that do not meet the consistency conditions is designated as the negative sample demand topic distribution set. Where: The positive and negative sample dual-sided feedback channel refers to the structure in which, during the topic alignment process, the distribution of demand topics that are similar in attributes to the scenario to be aligned is introduced as a positive constraint sample, and the distribution of demand topics that are dissimilar in attributes is introduced as a negative inhibition sample, so as to jointly constrain the topic alignment process. The consistency condition refers to the judgment condition that the demand forecast scenario is completely consistent with the scenario to be aligned in terms of regional dimension identifier, vehicle type dimension identifier, and usage scenario dimension identifier, or meets the similarity threshold. Based on the distribution of demand topics in the scenario to be aligned, the distribution of target demand topics for reconstruction, and the sets of positive and negative sample demand topic distributions, the degree of topic difference between each potential demand topic is calculated, a topic transmission cost table is constructed, and source-end topic weight sequences, target-end topic weight sequences, and negative constraint weight sequences are generated, where: Calculate the degree of thematic differences among the potential demand themes, specifically as follows: For each potential demand theme in the distribution of demand themes in the scenario to be aligned and the distribution of demand themes in the target demand theme for reconstruction, the differences in the probability distribution of components are compared, and the corresponding component probability differences are summarized to form a difference value that reflects the degree of structural difference between the themes. Construct the topic transport cost table as follows: Arrange the calculated difference values ​​between each potential demand topic in the order of topic index, and construct a topic transmission cost table that represents the transmission cost between topics, with the scenario topic to be aligned as the row and the target demand topic to be reconstructed as the column. The source-side topic weight sequence, target-side topic weight sequence, and negative constraint weight sequence are generated as follows: The weights corresponding to each potential demand topic are extracted from the demand topic distribution of the scene to be aligned, forming a source-end topic weight sequence. The weights corresponding to each potential demand topic are extracted from the target demand topic distribution, forming a target-end topic weight sequence. At the same time, based on the frequency of occurrence of each topic in the negative sample demand topic distribution set, a negative constraint weight sequence to suppress inconsistent topic mapping is generated. Based on the topic transport cost table, the source-end topic weight sequence, the target-end topic weight sequence, and the negative constraint weight sequence, the topic transport matrix is ​​initialized, and iterative normalization update processing is performed until the convergence criterion is met, generating the optimal topic transport matrix and the corresponding optimal transport mapping relationship, where: Initialize the topic transport matrix as follows: Based on the row and column structure of the topic transmission cost table, a corresponding initial topic transmission matrix is ​​constructed, and the source-end topic weight sequence and the target-end topic weight sequence are used as the initial conditions for the row and column constraints of the transmission matrix, respectively. Perform iterative normalization update processing, specifically as follows: While keeping the topic transmission cost table unchanged, the row and column directions of the topic transmission matrix are alternately normalized and adjusted. The weighted sum of each row gradually approximates the topic weight sequence of the source end, and the weighted sum of each column gradually approximates the topic weight sequence of the target end. A negative constraint weight sequence is introduced to suppress unreasonable topic mapping paths. The convergence criterion refers to the criterion that, during multiple consecutive iterations of updates, the change magnitude of each element in the topic transmission matrix is ​​lower than the threshold, and both row and column constraints simultaneously meet the error requirements. The optimal topic transport matrix and its corresponding optimal transport mapping relationship are generated as follows: When the convergence criterion is met, the topic transfer matrix obtained in the current iteration is taken as the optimal topic transfer matrix, and the transfer mapping relationship between the topic requirement to be aligned and the target requirement topic to be reconstructed is determined based on the transfer weight relationship between each topic in the optimal topic transfer matrix. Based on the optimal topic transfer matrix, the topic distribution of the scene to be aligned is mapped to a topic space consistent with the topic distribution of the reconstruction target, generating a representation of the alignment requirement topic distribution, where: Generate an alignment requirement topic distribution representation, specifically as follows: Based on the optimal topic transfer matrix, the weights of each potential demand topic in the distribution of demand topics to be aligned are redistributed to the topic space of the target demand topic to be reconstructed according to the corresponding optimal transfer mapping relationship. The redistributed topic weights are then normalized to generate an aligned demand topic distribution representation that is consistent with the target demand topic distribution in terms of topic structure.

[0026] In this embodiment, the demand forecast results for the output components include: Based on the alignment demand topic distribution representation, read the alignment demand topic weight sequence corresponding to the demand prediction scenario; Based on the basic attribute data of components and vehicle configuration information, the attribute fields of each component entity and the set of compatible vehicle models are read from the automotive component knowledge graph to generate a component attribute feature set. Then, encoding unification and normalization processing are performed to obtain a standardized component attribute feature vector set, where: The standardized component attribute feature vector set is obtained as follows: From the automotive parts knowledge graph, the basic attribute fields and the set of compatible vehicle models corresponding to each part entity involved in the demand prediction scenario are read one by one. The basic attribute fields include part category, specifications, material type and applicable years. The set of compatible vehicle models is represented in the form of a list of vehicle model identifiers. The discrete attributes in the basic attribute fields are mapped to numerical representations. The set of compatible vehicle models is converted into a multi-dimensional existence identifier vector. A uniform scaling transformation is performed on each numerical dimension to generate a standardized part attribute feature vector with consistent length and fixed dimensions for each part entity. These vectors are then aggregated to form a standardized set of part attribute feature vectors. Based on maintenance and repair records and fault event data, the number of event triggers, maintenance occurrences, and fault occurrences for each component entity are statistically analyzed within the time frame corresponding to the demand forecasting scenario. This generates a component event frequency feature set, which is then processed with time alignment and window aggregation to obtain a component frequency feature vector set, where: The set of frequency feature vectors of the components is obtained as follows: Within the time frame corresponding to the demand forecast scenario, maintenance and repair records and fault event data are retrieved one by one for each component entity. The number of event triggers, maintenance occurrences, and fault occurrences for each component entity within each time window are counted. The statistical results from different sources are aligned according to a unified time axis. Multiple counts for the same component within each time window are summarized. The statistical results under each time window are arranged in chronological order to form a frequency vector reflecting the change in the intensity of component events over time. This is then used to generate a set of component frequency feature vectors. Based on vehicle operating condition data and regional and store data, time and regional features corresponding to demand forecasting scenarios are extracted and then concatenated with aligned demand theme weight sequences, standardized component attribute feature vector sets, and component frequency feature vector sets to generate a unified demand representation vector, wherein: Generate a unified demand representation vector, specifically as follows: For demand forecasting scenarios, operational status indicators within the corresponding time range are extracted from vehicle operation data, and regional identifiers and store distribution features matching the scenario are extracted from regional and store data. After converting the extracted time and regional features into numerical vector representations, they are concatenated with the aligned demand theme weight sequence, standardized component attribute feature vectors, and component frequency feature vectors. Consistency checks and scale unification processing are performed on each dimension to form a unified demand representation vector with a fixed structure and complete semantics. Based on a unified demand representation vector, the demand forecast values ​​for each component are output according to the component dimension, and the demand forecast results for each component are output, where: The demand forecast values ​​for each component are output according to the component dimension, specifically: Using the vector representation of the corresponding component in the unified demand representation vector as input, the demand representation vector is read one by one according to the component identifier, and the values ​​of each dimension are used as input elements for demand intensity calculation to generate demand value results that correspond one-to-one with the component. The output demand forecast results for the components are as follows: The required numerical results for each component are associated and encapsulated with the component identifier to form a data set containing the component identifier and the corresponding demand forecast value. The data is then output according to the demand forecast scenario as the demand forecast result for automotive components.

[0027] refer to Figure 3 A knowledge graph-based automotive parts demand forecasting system includes the following modules: The data acquisition module is used to collect multi-source heterogeneous data related to the forecasting of automotive parts demand. The data preprocessing module is used to preprocess multi-source heterogeneous data and generate standardized multi-source heterogeneous data. The knowledge graph construction module is used to generate automotive parts knowledge graphs based on standardized multi-source heterogeneous data. The demand scenario segmentation module is used to segment demand forecasting tasks into scenarios according to region, vehicle type, and usage scenario, and generate demand forecasting scenarios. The semantic unit construction module is used to extract event subgraphs from the automotive parts knowledge graph and construct a set of demand behavior semantic units in various demand prediction scenarios. The document term co-occurrence modeling module is used to generate a document term co-occurrence matrix based on the set of semantic units of demand behavior. The demand topic modeling module is used to build an improved PLSA topic model, perform topic learning processing, and obtain the demand topic distribution. The demand topic alignment module is used to generate an aligned demand topic distribution representation by executing the Sinkhorn algorithm based on the demand topic distribution. The demand feature representation module is used to construct a demand representation vector by aligning the demand topic distribution representation and integrating component attribute features, event frequency features, and time features. The demand forecasting output module is used to output the demand forecasting results for each component based on the demand representation vector.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a parts demand forecasting scenario in a provincial-level automotive aftermarket service network. This network covers three cities along the eastern coast and surrounding areas, involving multiple car brands and different usage scenarios. The region has a large vehicle fleet, with frequent model updates. Maintenance and repair needs are significantly affected by regional climate differences, vehicle usage intensity, and model configuration variations. Traditional parts demand forecasting methods based on historical statistics have large prediction errors when applied across cities and car models, making it difficult to provide a stable basis for store inventory allocation.

[0029] In this scenario, basic attribute data of auto parts, vehicle configuration information, vehicle operating condition data, maintenance records, fault event data, and regional and store data from the past three years are collected to construct an automotive parts knowledge graph. Demand prediction scenarios are divided according to city region, vehicle type, and vehicle usage scenario. Within a unified time window, event subgraphs related to maintenance and fault events are extracted from the knowledge graph, and demand behavior semantic units are constructed. An improved PLSA topic model is introduced into each demand prediction scenario for topic learning to obtain the potential demand topic distribution under each scenario. Entropy regularization optimal transport calculation based on the Sinkhorn algorithm is used to perform structural alignment on the demand topic distribution of each scenario, mapping it to a unified topic space. The aligned demand topic distribution is then fused with part attribute features, event frequency features, and time features to output the demand prediction results for each part in different stores.

[0030] Experimental results show that, during the implementation of the system of this invention, compared with the traditional prediction method based on historical averages, the average absolute error of the predicted demand for braking system components and wear parts in this embodiment was reduced from 18.6% to 9.3%, and the prediction fluctuation amplitude in cross-city scenarios was reduced by about 42%. In three sample stores, the average amount of inventory backlog caused by prediction deviation was reduced by about 27%, and the number of emergency transfers was reduced by 19, verifying the significant improvement of the stability of demand prediction and practical application value of this invention under multi-scenario conditions.

[0031] Table 1. Comparison of Parts Demand Prediction Performance of Different Methods in Multi-Region, Multi-Vehicle Scenarios

[0032] As shown in Table 1, there are significant differences in prediction accuracy and practical application effect between traditional statistical prediction methods and the method of this invention in demand forecasting scenarios involving multiple regions, vehicle models, and component categories. Taking high-frequency demand components such as brake pads, suspension components, and electronic control modules as examples, the prediction error rate of traditional statistical methods in various cities is generally between 18% and 21%, with the predicted results significantly higher than the actual demand, easily leading to inventory redundancy. In contrast, the prediction error rate of the method of this invention is stably controlled within the range of 2% to 5% for various components. For example, the prediction error rate of brake pads in city A decreased from 18.75% to 2.34%, and the prediction error rate of suspension and steering components in city B decreased from over 17% to approximately 3%, indicating that the method of this invention can maintain high prediction consistency and stability under different regional and vehicle model conditions.

[0033] From the perspective of inventory management effectiveness, due to the significant improvement in forecasting accuracy, this invention has effectively reduced the amount of inventory backlog in stores across various cities. Taking city C as an example, under traditional forecasting conditions, 260 units of easily damaged rubber parts for compact cars were backlogged. However, after adopting the method of this invention, the backlog was significantly reduced, and inventory allocation became closer to actual demand. The number of emergency transfers also decreased simultaneously. In cities A and B, the number of transfers across multiple parts categories was reduced to about half of the original level, indicating that the demand forecasting results have higher guiding value for actual replenishment and transfer decisions.

[0034] Overall, the data shows that this invention effectively mitigates the impact of differences in demand distribution across different regions, vehicle models, and usage scenarios on the prediction results by uniformly modeling and aligning the demand theme structure across multiple scenarios. This makes the demand prediction results more stable and reliable when applied across different scenarios. As the number of samples increases and the scope of scenario coverage expands, the prediction error of this invention remains at a low level for various components, without significant fluctuations, demonstrating the continued effectiveness and engineering feasibility of the method in complex real-world application environments.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based method for predicting the demand for automotive parts, characterized in that, include: Collect multi-source heterogeneous data related to automotive parts demand forecasting, preprocess the multi-source heterogeneous data, and generate standardized multi-source heterogeneous data. Based on standardized multi-source heterogeneous data, we extract entities such as parts, vehicle models, usage scenarios, regions, events, and time, and construct assembly relationships, adaptation relationships, triggering relationships, and attribution relationships between entities to generate an automotive parts knowledge graph. Demand prediction scenarios are divided according to the combination of region, vehicle type and usage scenario. In each demand prediction scenario, demand-related event subgraphs are extracted from the automotive parts knowledge graph, a set of demand behavior semantic units is constructed, and a document term co-occurrence matrix is ​​generated. An improved PLSA topic model is constructed, and topic learning is performed on the document term co-occurrence matrix under various demand prediction scenarios to obtain the corresponding potential demand topic distribution and the corresponding component probability distribution, which are then encapsulated as demand topic distribution. The Sinkhorn algorithm is used to perform entropy regularization optimal transmission calculation on the demand topic distribution and alignment target of each scene to be aligned, to obtain the optimal transmission mapping relationship and generate an alignment demand topic distribution representation; Based on the aligned demand topic distribution representation, and combined with component attribute features, event frequency features, and time features, a unified demand representation vector is constructed to output the demand prediction results for components.

2. The knowledge graph-based method for predicting automotive parts demand according to claim 1, characterized in that, The multi-source heterogeneous data specifically includes basic attribute data of components, vehicle configuration information, vehicle operating condition data, maintenance and repair record data, fault event data, and regional and store data.

3. The knowledge graph-based method for predicting automotive parts demand according to claim 1, characterized in that, The preprocessing of multi-source heterogeneous data specifically includes deduplication, missing data completion, anomaly removal, time alignment, and unified encoding.

4. The knowledge graph-based method for predicting automotive parts demand according to claim 1, characterized in that, The generation of the automotive parts knowledge graph includes: Based on standardized multi-source heterogeneous data, we extract component entities, vehicle entities, usage scenario entities, region entities, event entities, and time entities to form a set of entity nodes. For the set of entity nodes, assembly and adaptation relationships are established between component entities and vehicle entities, triggering relationships are established between event entities and component entities, and attribution relationships are established between store entities and regional entities, forming a set of relationship edges; Within a unified time window, the number of occurrences of various relationships between corresponding entity pairs is counted, and a corresponding relationship weight is assigned to each relationship based on the relative frequency of occurrence of each relationship among similar relationships. A knowledge graph of automotive parts is generated based on the set of entity nodes, the set of relation edges, and the corresponding relation weights.

5. The knowledge graph-based method for predicting automotive parts demand according to claim 1, characterized in that, The generated document term co-occurrence matrix includes: Based on regional and store data, vehicle configuration information and vehicle operating condition data, regional dimension identifiers, vehicle dimension identifiers and usage scenario dimension identifiers are determined respectively, and the three types of identifiers are combined to determine a demand prediction scenario, forming a set of demand prediction scenarios. For each demand forecasting scenario, based on the timestamp information corresponding to the time entity, the maintenance record data and fault event data within the demand forecasting scenario are sorted by time, and the time axis is divided according to the time window length to generate a continuous time window. Within each time window of each demand forecasting scenario, maintenance and repair event entities and fault event entities that occur within the time window are retrieved from the automotive parts knowledge graph. The event entities are used as starting nodes to expand the associated parts entities and vehicle models along the assembly relationship, adaptation relationship and triggering relationship to generate a demand-related event subgraph. Each demand-related event subgraph corresponding to each time window in each demand forecasting scenario is encapsulated into a demand behavior semantic unit, and each demand behavior semantic unit is defined as a document unit. Each document unit contains a set of component entities and the corresponding number of events. By using component entities as term dimensions and each requirement behavior semantic unit as document dimensions, a document term co-occurrence matrix is ​​generated based on the number of events occurring in each component entity within each requirement behavior semantic unit.

6. The knowledge graph-based automotive parts demand forecasting method according to claim 1, characterized in that, The encapsulation is a demand topic distribution, including: An improved PLSA topic model is constructed. Under various demand prediction scenarios, the set of semantic units of demand behavior is sorted, and short-granular semantic windows, medium-granular semantic windows and long-granular semantic windows are constructed based on different time spans to form a multi-scale semantic window set. Based on a multi-scale semantic window set, a consistency guidance structure is introduced. The statistical results of demand semantics formed in the long-granularity semantic window are used as constraint information to guide the weight distribution of demand behavior semantic units in the short-granularity and medium-granularity semantic windows. At the same time, the demand semantic units that frequently appear in the short-granularity semantic window are fed back to the medium-granularity and long-granularity semantic windows, forming a cross-scale bidirectional constraint semantic consistency guidance relationship. Based on semantic consistency guidance, corresponding PLSA sub-models are constructed for short-granularity semantic windows, medium-granularity semantic windows, and long-granularity semantic windows, respectively. Based on the document term co-occurrence matrix, topic learning is performed in each PLSA sub-model. A cross-semantic scale consistency guidance structure is introduced during the topic learning process to constrain and smooth the update process of the demand topic distribution and component probability distribution at each scale, thereby obtaining the corresponding demand topic distribution sets under short-granularity, medium-granularity and long-granularity semantic windows. The distribution of demand topics under each semantic scale is compared in the continuous time dimension. The degree of change in the distribution of demand topics within adjacent time periods is calculated. An adaptive semantic time period positioning structure is introduced. When the degree of change in the distribution of demand topics exceeds the threshold, the corresponding time point is determined as a semantic change point, and the time axis is re-divided to generate an adaptive semantic time period. Based on adaptive semantic time periods, the demand topic distributions formed under short-granularity, medium-granularity, and long-granularity semantic windows are integrated to obtain the corresponding demand topic distributions.

7. The knowledge graph-based method for predicting automotive parts demand according to claim 1, characterized in that, The generated alignment requirement topic distribution representation includes: Based on the demand topic distribution, the Sinkhorn algorithm is executed to determine the set of demand topic distributions for the scenarios to be aligned. Integrity checks are performed on the target demand topic distribution. When missing topics or abnormal topic weights are detected in the target demand topic distribution, an adaptive reconstruction structure for the alignment target is introduced. Based on the historical stable demand topic distribution set, the target demand topic distribution is reconstructed to generate a reconstructed target demand topic distribution. Based on information about region, vehicle type, and usage scenario, similarity is determined for the scenarios to be aligned. A positive and negative feedback channel is introduced. The distribution of demand topics that meet the consistency conditions with the scenarios to be aligned in terms of region, vehicle type, and usage scenario is taken as the set of positive demand topic distributions, and the distribution of demand topics that do not meet the consistency conditions is taken as the set of negative demand topic distributions. Based on the distribution of demand topics in the scenario to be aligned, the distribution of demand topics for reconstruction, and the set of positive and negative sample demand topics, the degree of topic difference between each potential demand topic is calculated, a topic transmission cost table is constructed, and a source-end topic weight sequence, a target-end topic weight sequence, and a negative constraint weight sequence are generated. Based on the topic transmission cost table, the source topic weight sequence, the target topic weight sequence, and the negative constraint weight sequence, the topic transmission matrix is ​​initialized, and iterative normalization update processing is performed until the convergence criterion is met, generating the optimal topic transmission matrix and the corresponding optimal transmission mapping relationship. Based on the optimal topic transfer matrix, the topic distribution of the scene to be aligned is mapped to a topic space consistent with the topic distribution of the reconstruction target, thereby generating a representation of the topic distribution of the alignment requirements.

8. The knowledge graph-based method for predicting automotive parts demand according to claim 1, characterized in that, The demand forecast results for the output components include: Based on the alignment demand topic distribution representation, read the alignment demand topic weight sequence corresponding to the demand prediction scenario; Based on the basic attribute data of parts and vehicle configuration information, the attribute fields of each part entity and the set of compatible vehicle models are read from the automotive parts knowledge graph to generate a set of part attribute features. Then, the encoding is unified and normalized to obtain a standardized set of part attribute feature vectors. Based on maintenance and repair records and fault event data, the number of event triggers, maintenance occurrences and fault occurrences of each component entity are counted within the time range corresponding to the demand prediction scenario. A set of component event frequency features is generated, and time alignment and window aggregation processing are performed to obtain a set of component frequency feature vectors. Based on vehicle operation data and regional and store data, time and regional features corresponding to demand prediction scenarios are extracted and then spliced ​​and fused with the aligned demand theme weight sequence, standardized component attribute feature vector set and component frequency feature vector set to generate a unified demand representation vector. Based on a unified demand representation vector, the demand forecast values ​​for each component are output according to the component dimension, and the demand forecast results for each component are output.

9. A knowledge graph-based automotive parts demand forecasting system, comprising executing the knowledge graph-based automotive parts demand forecasting method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition module is used to collect multi-source heterogeneous data related to the forecasting of automotive parts demand. The data preprocessing module is used to preprocess multi-source heterogeneous data and generate standardized multi-source heterogeneous data. The knowledge graph construction module is used to generate automotive parts knowledge graphs based on standardized multi-source heterogeneous data. The demand scenario segmentation module is used to segment demand forecasting tasks into scenarios according to region, vehicle type, and usage scenario, and generate demand forecasting scenarios. The semantic unit construction module is used to extract event subgraphs from the automotive parts knowledge graph and construct a set of demand behavior semantic units in various demand prediction scenarios. The document term co-occurrence modeling module is used to generate a document term co-occurrence matrix based on the set of semantic units of demand behavior. The demand topic modeling module is used to build an improved PLSA topic model, perform topic learning processing, and obtain the demand topic distribution. The demand topic alignment module is used to generate an aligned demand topic distribution representation by executing the Sinkhorn algorithm based on the demand topic distribution. The demand feature representation module is used to construct a demand representation vector by aligning the demand topic distribution representation and integrating component attribute features, event frequency features, and time features. The demand forecasting output module is used to output the demand forecasting results for each component based on the demand representation vector.

Images