Artificial intelligence-driven enterprise asset data medium construction method and system

Abstract

The invention discloses an artificial intelligence-driven enterprise asset data table construction method and system, and particularly relates to the technical field of asset data management.The method comprises the steps that a structured asset information flow is accessed through event driving, a unified time grid and a rolling window are adopted to align a time sequence, and unified asset entries are generated; an asset knowledge graph is constructed on the basis, and potential links are complemented and credibility is updated through graph reasoning; calculating a load fluctuation index and a use reliability index in a graph context, and obtaining an operation adaptation coefficient through a neural fuzzy reasoning model; the operation adaptation coefficient and the dynamic change trend index are injected into reinforcement learning, a fact chain and offline playback buffer are combined to output a maintenance and allocation decision suggestion set, multi-dimensional semantic unification and associated dynamic update are realized, the consistency of access calibers and the traceability of results are ensured, the decision suggestion set and unified asset items are bound and put into a disk, and the decision suggestion set is stored in a database. And the problem of decision inaccuracy caused by cross-system asset state inconsistency and association missing is solved.

Description

Technical Field

[0001] This invention relates to the field of asset data management technology, and more specifically, to an artificial intelligence-driven method and system for building an enterprise asset data platform. Background Technology

[0002] In real-world business scenarios across the entire lifecycle of enterprise assets, individual assets are not isolated records but exist within a multi-dimensional network of relationships comprised of locations, projects, personnel, and maintenance plans. Every change in an asset's status, from procurement and shelving to allocation, maintenance, and eventual decommissioning, is accompanied by the generation, updating, and invalidation of related relationships. Current practices primarily focus on field-level standardization and ledger aggregation, neglecting the semantic management of cross-system, cross-departmental, and cross-time-window relationships. On the one hand, business systems from different sources differ in their identification methods, timestamp granularity, and update rhythms, making it difficult to create a unified asset entry for the same asset globally. Status and relationships may lag, be missing, or conflict at different stages. On the other hand, the relationships between assets and locations, projects, personnel, and maintenance plans have significant temporal and conditional characteristics. These relationships include explicit one-to-one or one-to-many bindings as well as diverse relationships that are temporarily generated and dynamically evolved during allocation and collaborative operations. The lack of continuous validity verification and credibility management makes it difficult to draw a complete context whenever a global perspective is needed.

[0003] The aforementioned shortcomings directly affect intelligent processing for decision-making: due to the lack of structured relational semantics and evolutionary trajectories, asset knowledge cannot be organized and reasoned in a graph-like manner. Artificial intelligence models lack the necessary context when predicting, optimizing, or recommending, which can easily lead to contradictory results for the same asset in different scenarios. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in the prior art, this invention provides an AI-driven method and system for constructing an enterprise asset data platform. This system accesses asset information flow through event-driven mechanisms, generates unified asset entries, and constructs an asset knowledge graph. It calculates load fluctuation and reliability indices, obtains operational adaptability coefficients via neural fuzzy logic, and uses reinforcement learning to output maintenance and allocation decision suggestions, thereby addressing the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing an enterprise asset data platform driven by artificial intelligence, comprising: Step S1: Integrate asset status update events from procurement, acceptance, use and maintenance stages through an event-driven architecture, and perform semantic annotation, entity extraction and structure transformation on the raw event data to form an asset information flow; Step S2: Perform multimodal feature embedding and similarity matching on the asset information stream, map and merge it into a unified asset entry, and update the cross-source asset mapping rules based on the incremental learning mechanism to obtain a unified asset record set; Step S3: Construct a dynamic asset knowledge graph based on a unified asset record set, establish multi-dimensional associations between assets and locations, projects, personnel and maintenance plans, and use graph reasoning algorithms to complete and update potential links; Step S4: Graph Representation and Anomaly Identification Train and apply a graph neural network model on the asset knowledge graph to generate representations of nodes and edges, and combine structural and temporal features to identify abnormal patterns of resource conflicts and uneven utilization, outputting abnormal early warning results and a candidate set of abnormal root causes. Step S5: Using the anomaly warning results and unified asset entries as input, perform strategy evaluation and optimization. Through strategy evaluation and reinforcement learning, obtain adaptive adjustment suggestions for maintenance plans and allocation strategies, and output a set of decision suggestions for management deployment.

[0006] Preferably, an event-driven architecture is used to access asset status update events in the procurement, acceptance, use and maintenance stages in real time, and semantic annotation, unstructured information extraction and structure transformation are performed on the captured event data to form a real-time asset information stream containing multi-dimensional attributes such as time series, personnel, location, project and maintenance plan. The asset information stream can be directly input into the unified asset item matching and knowledge graph construction steps.

[0007] Preferably, when performing unified matching processing on structured asset information flows, multimodal feature embedding, similarity matching, and normalization are introduced, and the cross-source mapping rules are dynamically optimized by combining an incremental learning mechanism, wherein: Multimodal feature embedding refers to aligning text logs, structured fields, and time series associated with a unified asset entry under a unified time raster, mapping them through an encoder to an asset event feature representation in the same vector space, and attaching source labels and time window labels to ensure data traceability; Similarity matching refers to determining whether events from different sources belong to the same unified asset item in a candidate unified asset item set based on a dual-channel order decision of semantic vector similarity and temporal consistency, and outputting a unique merging result within the matching threshold. Normalization refers to performing quantile normalization on the key numerical features and related indicators of the embedded vector according to the location and project dimensions to a closed interval of zero to one, and archiving the mapping parameters, time window labels and asset knowledge graph version numbers for replay and auditing. Dynamically optimizing cross-source mapping rules by combining incremental learning mechanisms refers to using the fact chain composed of confirmed matching and rejection samples as supervision, updating the embedding parameters, matching threshold, and rule strength online under the constraints of minimum evidence number and small step size, and binding them to the asset knowledge graph version number to reduce false inclusions and missed matches.

[0008] Preferably, the graph reasoning algorithm includes: A path pattern library with time-series constraints is established in the asset knowledge graph. Path patterns are defined by relation sequences of no more than two lengths, along with time window labels, source labels, relation types, and role labels. Attribute nearest neighbor channels are constructed, embedding the text, structured fields, and time series corresponding to unified asset entries into a shared vector space. A fixed number of nearest neighbors are selected to generate candidate links under the criteria of category consistency and quantile normalization. A context consistency gate is introduced to screen the feasibility of candidate links using location capacity labels and project time limit labels. Candidates meeting the capacity and time limit requirements are retained, while those not meeting these requirements are excluded. Select and transfer to cold storage and record the reason; perform intersection and union judgment on the output of path pattern channel and attribute nearest neighbor channel, the intersection is directly stored in the database, and the union is marked as a link to be verified; maintain the credibility of each candidate link, the credibility is updated according to the occurrence count and recent confirmed events according to the time decay function, and the link below the threshold is transferred to cold storage and the evidence reference is retained; use the change event of unified asset entry as the trigger to incrementally recalculate the candidate link and credibility in its K-hop local neighborhood, and bind all new, update and cold storage operations to the asset knowledge graph version number and write the explanation information, which is used as a traceable context for subsequent graph representation and anomaly identification.

[0009] Preferably, step S4 includes: constructing a dual-channel graph neural network with relation type awareness and event string temporal encoding within the version freeze view of the asset knowledge graph; generating node representations and edge representations by weighted message passing according to relation type; and merging the structural channels and temporal channels at the node and edge levels using gating; simultaneously applying contrast constraints to the same node and its neighborhood distribution in adjacent versions to stabilize cross-version representations; calculating local consistency scores based on the resource conflict pattern library; introducing location capacity labels and project time limit labels as hard constraint filters to form an anomaly candidate list; comparing the differences of the usage allocation vectors of location and project dimensions within the same time window label; and completing the anomaly sorting by combining the migration trajectory and maintenance plan overlap relationship of adjacent versions; generating a minimum evidence subgraph and explanation field by backtracking the path within the limited number of hops; writing the anomaly warning results and anomaly root cause candidate set together with the time window label and version number back to the asset knowledge graph and binding them with the unified asset entry for use as traceable input for subsequent strategy evaluation and optimization.

[0010] Preferably, the local consistency score is obtained as follows: within the version-frozen view of the asset knowledge graph, node representations and edge representations, as well as relationship types, role labels, and time window labels are read for the target asset and relationship instances. Within a limited number of hops neighborhood, neighborhood representations are aggregated according to relationship type and role label to form a structural prototype. A temporal prototype is generated based on the time sequence and change interval of the edge event string. The similarity and continuity between the target instance representation and the structural prototype and temporal prototype are jointly measured as a consistency measure. When the constraints of location capacity label and project time limit label are not met, a penalty is applied and a local consistency score is output. The significance of the local consistency score is to characterize the degree of fit between the target instance and its neighborhood structural pattern and temporal evolution trajectory. A lower local consistency score indicates deviation from the normal pattern and potential resource conflicts or unbalanced utilization. The purpose of the local consistency score is to generate and sort abnormal candidates, trigger thresholded alarms, and drive the minimum evidence subgraph backtracking. At the same time, it is bound to the unified asset entry for direct data retrieval and interpretation for strategy evaluation and optimization in step S5.

[0011] Preferably, it also includes the steps for determining the enterprise asset maintenance window, including: A load fluctuation index and a usage reliability index are introduced into the asset knowledge graph. The load fluctuation index is calculated by weighting the ratio of peak load to average load of the asset's historical usage curve with the frequency of operating condition disturbances. The usage reliability index is calculated by combining the average failure interval of the asset with the logarithmic ratio of usage time to maintenance interval. The load fluctuation index and the usage reliability index are input into a neural fuzzy inference model to obtain the operating fitness coefficient. When the operating fitness coefficient is within the hysteresis threshold band, only the membership snapshot and trigger rule number are recorded without performing parameter updates. Parameter updates are performed only when the coefficient is higher than the upper threshold or lower than the lower threshold.

[0012] Preferably, the method for obtaining the operational fitness coefficient includes: calculating the load fluctuation index and the reliability index within the same time window label, and performing quantile normalization to a closed interval of zero to one according to the location dimension and the project dimension; when the location or project sample is insufficient, calling the category-level quantile snapshot or the previous version snapshot of the same type of asset and marking the source; entering the normalized data into the neural fuzzy inference model, defining three groups of membership functions of "low", "medium" and "high" respectively, and performing context shaping according to asset category, location characteristics and maintenance plan rigor; setting interpretable rules and setting rule strength, calculating membership degree and obtaining the activation degree of each rule by product combination, and performing centroid defuzzification after aggregating the rule conclusions to obtain the operational fitness coefficient in the range of zero to one.

[0013] Preferably, the initialization of the hysteresis threshold is set based on category engineering experience, and then small-step adjustments are made using the false trigger rate and missed trigger rate obtained from the fact chain statistics as correction signals. The threshold parameters are archived together with the asset knowledge graph version number and time window label to achieve traceable reproducibility during the generation of decision suggestion sets and replay auditing.

[0014] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence-driven enterprise asset data platform construction system, comprising: The event access module uses an event-driven architecture to access asset status update events from the procurement, acceptance, use and maintenance stages, and performs semantic annotation, entity extraction and structure transformation on the accessed raw event data to form an asset information flow. The multimodal matching and merging module performs multimodal feature embedding and similarity matching on the asset information flow, maps and merges them into unified asset entries, and updates the cross-source asset mapping rules based on an incremental learning mechanism to obtain a unified asset record set. The knowledge graph construction module builds a dynamic asset knowledge graph based on a unified asset record set, establishing multi-dimensional associations between assets and locations, projects, personnel and maintenance plans, and using graph reasoning algorithms to complete and update potential links; The graph representation anomaly identification module trains and applies a graph neural network model on the asset knowledge graph to generate representations of nodes and edges, and combines structural and temporal features to identify abnormal patterns of resource conflicts and uneven utilization, outputting anomaly warning results and anomaly root cause candidate set. The strategy evaluation module takes the anomaly warning results and unified asset entries as input to perform strategy evaluation and optimization. Through strategy evaluation and reinforcement learning, it obtains adaptive adjustment suggestions for maintenance plans and allocation strategies.

[0015] The technical effects and advantages of this invention are as follows: (1) This invention generates a structured asset information flow under event-driven access, constructs an asset knowledge graph based on unified asset entries, and uses graph reasoning to complete the incremental updates of potential links and credibility management, thereby achieving consistent expression and traceable association across sources, time windows, and dimensions, and comprehensively presenting the multidimensional semantics of assets, locations, projects, personnel, and maintenance plans. This significantly improves data alignment and contextual completeness, and effectively solves the problems of state fragmentation and limited perspective caused by the traditional approach of only recording fields and lacking associated semantics.

[0016] (2) This invention calculates the load fluctuation index and the reliability index in the context of asset knowledge graph, inputs them into the neural fuzzy reasoning model to generate the operation fitness coefficient, and injects the operation fitness coefficient and the dynamic change trend index M into reinforcement learning. Combined with the fact chain and offline playback buffer, an interpretable closed loop is formed, realizing the adaptive optimization of maintenance and allocation strategies and the output of decision suggestion sets, ensuring the implementation and continuous evolution of strategies, and effectively solving the problems of disconnect between anomaly judgment and execution, uneven resource utilization and unbalanced maintenance timing. Attached Figure Description

[0017] Figure 1 This is a flowchart of the enterprise asset data platform construction method of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0020] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0021] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0022] Example 1, see Figure 1 The present invention provides a flowchart of a method for constructing an enterprise asset data middle platform. Figure 1 The AI-driven enterprise asset data platform construction method shown includes: Step S1: Integrate asset status update events from procurement, acceptance, use and maintenance stages through an event-driven architecture, and perform semantic annotation, entity extraction and structure transformation on the raw event data to form an asset information flow; Summary of Step S1: In the process of enterprise asset lifecycle management, from procurement and acceptance to use and maintenance, assets generate multi-source and heterogeneous data records. These data often have problems such as inconsistent formats, inconsistent content, and asynchronous time sequences. In order to achieve the construction of a unified data platform, it is necessary to use artificial intelligence methods to automatically process the raw event data in the real-time scenario of asset status updates, thereby transforming various scattered, heterogeneous, and unstructured asset status information into a structured information flow that can be used for unified analysis. The subject of this step is the data collection and processing process of the enterprise asset management department, and the time of occurrence covers the real-time process of asset status updates. The scenarios include data generated in procurement and acceptance, use scheduling, and maintenance operations.

[0023] Furthermore, step S1 includes the following sub-steps: Step S101: Real-time access to asset status update events, using an event-driven architecture to capture event records from procurement, acceptance, use and maintenance stages, including operation timestamps, event types, relevant personnel or department information and asset identifiers, ensuring that each event has complete contextual information when it is generated; Step S102: Perform semantic annotation on the captured raw event data, including automatic identification and classification of asset categories, operation types, and event impact scope. At the same time, perform natural language processing on unstructured text content to extract key information, such as maintenance instructions and fault descriptions, so that asset status information can be converted into standardized labels. Step S103: Perform structured transformation on the labeled asset event data, mapping the text and tag information to a unified data model to generate a structured information flow that can be directly input into unified asset item matching and knowledge graph construction; during the transformation process, the time sequence of events, related personnel, locations, projects and maintenance plan information are preserved to ensure that the multi-dimensional relationships between various attributes are not lost; Step S104: Perform a preliminary comparison between the structured information flow and historical asset event data to detect potential inconsistencies or missing fields, and mark the data quality status (such as missing, abnormal, or pending confirmation). Step S105: After structuring and labeling are completed, the event information flow is transformed into a real-time asset information flow and continuously output to the next step of the unified asset entry generation process to achieve dynamic fusion of multi-source events.

[0024] Step S2: Perform multimodal feature embedding and similarity matching on the asset information stream, map and merge it into a unified asset entry, and update the cross-source asset mapping rules based on the incremental learning mechanism to obtain a unified asset record set; The goal of step S2 is to eliminate duplicate, ambiguous, or inconsistent information across systems while maintaining the integrity of asset information, so as to provide high-quality unified asset entries for subsequent construction of dynamic knowledge graphs and anomaly identification, thereby achieving intelligent fusion and global management of cross-system data.

[0025] Furthermore, step S2 includes the following sub-steps: Step S201: Input the generated real-time asset information stream into the matching process. Perform preliminary aggregation on each asset event based on the unique asset identifier, event type and timestamp, and identify duplicate or conflicting asset entries through fuzzy matching method to ensure that events of the same asset from different sources can be accurately associated. Step S202: Perform multimodal feature embedding processing on the initially aggregated asset entries, including text description vectorization, numerical attribute normalization, and time series encoding, to form a unified asset representation vector. Step S203: Perform similarity matching and normalization operations on the asset representation vector to map and merge events from different sources into unified asset entries, mark any possible data inconsistencies or missing information, and provide complete and reliable asset entries for subsequent knowledge graph construction. Step S204: Introduce an incremental learning mechanism to dynamically update cross-source asset mapping rules based on feedback from manual confirmation or automatic verification, including similarity thresholds, matching weights, and conflict handling strategies. This will continuously optimize the unified asset entry generation process as new data streams are added, thereby improving mapping accuracy and stability. Step S205: Output a unified set of asset entries, including complete time series, event types, multidimensional attributes and data quality status, to provide directly usable input data for the knowledge graph construction in step S3.

[0026] Step S3: Construct a dynamic asset knowledge graph based on a unified asset record set, establish multi-dimensional associations between assets and locations, projects, personnel and maintenance plans, and use graph reasoning algorithms to complete and update potential links.

[0027] Furthermore, step S3 includes the following sub-steps: Step S301: Extract key attributes such as asset identifier, location, project, personnel and maintenance plan from the unified asset entry set, and perform initial mapping according to the predefined relationship pattern to generate multi-dimensional entities and candidate associations between them, providing basic input for knowledge graph construction; Step S302: Perform AI-driven semantic disambiguation on candidate relationships, including automatic identification and unification of synonyms, abbreviations, and contextual references to avoid relationship redundancy or ambiguity caused by multiple system sources; Step S303: Based on semantic consistency, use graph structure learning methods to calculate the potential connection strength between entities, add entities with high potential connections as new edges, and label weak connections with confidence to form a dynamically updated multidimensional relationship link. Step S304: Deploy a dynamic update mechanism for the knowledge graph. When new event data is added, new entity nodes are automatically added or existing relationships are corrected by comparing with the existing graph, ensuring the continuity and integrity of the knowledge graph in the time dimension. Step S305: Use the updated knowledge graph as an extended representation of the unified asset entries and output it to the subsequent anomaly identification and optimization steps to ensure that subsequent processing can be carried out on the basis of more complete contextual semantics and relationships.

[0028] Step S4: Graph Representation and Anomaly Identification Train and apply a graph neural network model on the asset knowledge graph to generate representations of nodes and edges, and combine structural and temporal features to identify abnormal patterns of resource conflicts and uneven utilization, outputting abnormal early warning results and a candidate set of abnormal root causes. Summary of Step S4: After completing the incremental update of the asset knowledge graph in Step S3, the graph representation and anomaly identification process is immediately initiated under the same time window label where the graph version is frozen. The business scenario is asset scheduling and maintenance collaboration across locations, projects, and personnel. The purpose is to generate representations of nodes and edges in a context consistent with structure and time sequence and to identify abnormal patterns of resource conflicts and uneven utilization. Taking the intelligent processing process of the asset data platform as the main body, the incremental changes bound to the unified asset entries are obtained through the event bus and aligned with the graph snapshot before entering the modeling and identification stage.

[0029] Furthermore, step S4 includes the following sub-steps: Step S401: Graph Representation Construction and Temporal Alignment (a) Freezing of map view and relationship-aware sampling On the asset knowledge graph output in step S3, the graph view corresponding to the same time window label is frozen with the version number as the anchor point. The graph contains five types of nodes: assets, locations, projects, personnel, and maintenance plans, as well as multi-dimensional association edges carrying relationship type, source label, time window label, validity status, and credibility. To ensure the reproducibility and interpretability of representation learning, the node feature pattern is first unified, mapping asset category, lifecycle status, location capacity label, project time limit label, personnel skill label, and maintenance plan severity to fixed fields, and the edge features are unified to relationship type, creation time, last change time, and source. Then, relationship-aware neighborhood sampling is performed, extracting the adjacency set according to the relationship type ratio within a local subgraph that does not exceed a predetermined number of hops. At the same time, a negative sample set that follows the constraints of location capacity label and project time limit label is constructed. Negative samples are only generated in edge pairs that are logically allowed but do not currently exist, in order to reduce training noise caused by invalid comparisons. The sampled subgraphs and the metadata of the frozen views are stored in the database to ensure that subsequent training, inference, and reproduction experiments are carried out in a consistent input context. (ii) Fusion of temporal alignment and representation stabilization To address the issues of non-uniform event timing and asynchronous arrival from different sources, incremental events in the structured asset information flow are aggregated into time-slice sequences using a unified time grid. The change trajectory of each edge is encoded into an ordered event string, forming two input paths: a "structural channel" and a "temporal channel." In representation modeling, a channel-level fusion strategy is employed. First, basic representations of nodes and edges are generated within the structural channel using relationship-type-aware message passing. Then, the event strings are sequence-encoded within the temporal channel to extract elements of change direction and rate. Finally, gating fusion merges the two representations at the node and edge levels, thus preserving both stable topological semantics and sensitive temporal dynamics. To improve the consistency of cross-version representations, comparative constraints are applied to the same node and its neighborhood distribution in adjacent versions, promoting smooth evolution of representations with business changes rather than drastic fluctuations. The fused node and edge representations, along with version numbers, time window labels, and feature lists, are written back for direct data retrieval in the next sub-step for anomaly identification. Step S402: Joint structural and temporal anomaly identification and root cause analysis (I) Resource Conflict Pattern Library and Constraint-Driven Decoding A resource conflict pattern library is established on the frozen view, covering business scenarios such as the same person binding multiple assets within overlapping time windows, the same location having its capacity label exceeding the limit within a time window, the same maintenance plan being occupied by multiple assets within overlapping time windows, and cross-project transfers causing the remaining capacity of the source project to fall below the required lower limit. Local consistency scores are calculated using the node and edge representations obtained in step 401, and location capacity labels and project time limit labels are used as hard constraint filters to first eliminate candidates that violate basic constraints. Then, a decoder guided by relationship type and role labels generates an abnormal candidate list on the remaining candidates. To ensure interpretability, the decoder returns the neighborhood evidence that reached the candidate when outputting each candidate, including the triggering relationship type, the involved time window, and the source label, forming an auditable structured evidence fragment. The candidate list is sorted by consistency score and constraint triggering degree, marked with warning levels, and a one-to-many reference relationship is established with the unified asset entry, facilitating subsequent strategy evaluation at the entry level. (ii) Using Imbalanced Identification and Local Evidence Subgraph Collection To identify uneven asset utilization at the location or project level, the neighborhood distribution is first statistically analyzed on the frozen view based on edge-level usage relationships. Within the same time window label, allocation vectors for locations and projects are constructed. The differences between the allocation vectors surrounding a target node and the reference allocation vectors of its corresponding location and project are compared. Significant differences that consistently cross adjacent versions are marked as imbalance risks. Furthermore, the migration trajectory of assets between adjacent versions and their overlap with maintenance plan edges are monitored. If migration accumulates in a few locations or projects and is accompanied by concentrated maintenance windows, the area is marked as a hotspot. For each imbalance risk, a path tracing with limited hops is performed to collect a minimum evidence subgraph. This subgraph includes the assets, locations, projects, personnel, and maintenance plan nodes in the trigger area, along with their key edges, and includes time window and source labels, forming a candidate set of anomaly root causes. Finally, a structured object containing anomaly warning results and the candidate set of root causes is output, with version numbers and explanation fields, for direct reference in the strategy evaluation and optimization of step S5, achieving seamless integration from identification to handling.

[0030] Step S5: Using the abnormal warning results and unified asset entries as input, perform strategy evaluation and optimization. Through strategy evaluation and reinforcement learning, obtain adaptive adjustment suggestions for maintenance plans and allocation strategies, and output a set of decision suggestions for management deployment. The purpose of step S5 is to take the abnormal warning results and unified asset entries as input, and combine the location capacity tags, project time limit tags and personnel availability tags of the asset knowledge graph to generate adaptive adjustment suggestions for maintenance plans and allocation strategies and form a set of decision suggestions; the frozen view provides constraints and evidence, and the event bus continuously backfills the execution results to improve the quality of the strategy in a closed loop.

[0031] Furthermore, step S5 includes the following sub-steps: Step S501: Construction of the strategy evaluation environment and constraints on the action space (a) Alignment of State Expressions and Evidence Binding The processing flow uses unified asset entries as indexes, reading the alert type, alert level, and candidate root cause set related to the target entry from the anomaly alert results. It then obtains node and edge representations, location capacity labels, project time limit labels, and personnel availability labels under the version-frozen view of the asset knowledge graph, forming a two-layer state representation. The observation layer carries a structural summary of local evidence and minimal evidence subgraphs from the anomaly alert results, including the triggering relationship type, the involved time window, source label, and affected neighborhood. The graph semantic layer carries executable constraints and resource profiles, including capacity characterization of the target location, time limit characterization of the target project, availability characterization of relevant personnel, and binding relationships with the maintenance plan. To eliminate time jitter in cross-source event arrivals, state representations undergo uniform temporal grid alignment and explicit missing labeling before entering training and evaluation. All data retrievals are bound to time window labels and asset knowledge graph version numbers. To avoid amplifying isolated anomalies into global optimization targets, state construction introduces a neighborhood truncation strategy: evidence is aggregated only within a subgraph with a limited number of hops, and the evidence is categorized by pattern using a resource conflict pattern library, ensuring that each state can be restored to a directed path of "evidence, context, and constraints". After alignment, state representations are stored on disk as replayable slices and written to an offline replay buffer to support subsequent reinforcement learning using a consistent input view. (ii) Action set trimming and safety boundary injection The action set originates from two types of interventions: maintenance and allocation. These interventions include early maintenance, delayed maintenance, cross-site allocation, intra-project reallocation, and observation delay. To ensure executability, the process first eliminates candidate actions that would cause the target location's capacity to exceed its limit based on the location capacity label. Then, it eliminates candidate actions that would exceed the target project's time limit based on the project time limit label. Finally, it verifies the scheduling feasibility of actions requiring manpower support based on the personnel availability label, thus obtaining a subset of actions that meet the hard constraints. To avoid high-frequency switching during uncertain phases, the process introduces safety boundaries: when the anomaly warning result points to a resource conflict mode, actions that can directly alleviate the conflict (such as intra-project reallocation and cross-site allocation) are prioritized; when the anomaly warning result points to an unbalanced utilization mode, actions that can revert to the reference allocation vector are prioritized (such as delayed maintenance or small-scale allocation); when the anomaly warning result is at the observation level, only the observation delay is retained, along with a small-scale reallocation. To shorten the cold start convergence time, the processing flow extracts "state, action, and result" slices similar to the current evidence from the historical fact chain and puts them into an offline playback buffer for reuse in the early assessment stage; all action candidates, reasons for elimination, and feasibility verification results are bound to unified asset entries and written back to ensure traceability in subsequent audits. Step S502: Reinforcement Learning Optimization and Policy Clearing (I) Exploration of Multi-Objective Benefit Order Synthesis and Constraint Perception The strategy evaluation is driven by a multi-objective benefit ordering synthesis: the first order is risk exposure reduction, reflected in a decrease in the level of abnormal warning results and an increase in local consistency scores; the second order is the timeliness of maintenance plan execution, reflected in the improved alignment between maintenance windows and project deadlines; the third order is resource utilization balance, reflected in the convergence of location and project dimension allocation vectors towards the reference allocation. Ordering synthesis avoids the conflict between objectives caused by simple weighting, ensuring that higher-level objectives are satisfied first before pursuing lower-level optimization. The exploration process introduces a constraint-aware mechanism, limiting the sampling frequency of high-risk actions based on location capacity labels and project deadline labels, and automatically reverting to suboptimal actions when scheduling is not feasible based on personnel availability labels; to reduce invalid searches, the evaluation phase sets negative memories for poorly performing action fragments in the fact chain, reducing repeated attempts in the same context. The state transition is based on the two-layer state representation in step S501. The anomaly warning result and the update of the minimum evidence subgraph affect the expected return of the next step, enabling the strategy to respond proactively to changes in evidence. Each round of training and evaluation is bound to the time window label and the asset knowledge graph version number to ensure that it can be replayed at any point in time and obtain consistent results. (ii) Strategy clearing, explanation package and closed-loop write-back Once the strategy reaches a stable range on the validation set, the processing flow enters the strategy clearing phase: abstract actions are mapped to specific maintenance windows, allocation paths, and personnel schedules, and affected assets, locations, projects, and maintenance plan nodes are located in the asset knowledge graph to form an executable draft work order. Clearing simultaneously generates an interpretation package, which contains three parts: an input snapshot (anomaly warning results, a minimal evidence subgraph summary, and state data retrieval criteria), a strategy reasoning snapshot (the adopted ordinal synthesis path, the set of restricted actions and the list of rejected actions, the generation order, and rollback records), and context fragments (referenced location capacity tags, project time limit tags, and personnel availability tags). The interpretation package is bidirectionally bound to the unified asset entry and written back to the asset knowledge graph as a reference for subsequent execution monitoring and re-evaluation; the event bus continues to collect execution results, including maintenance completion records, resource fallback trajectories, and changes in anomaly warning results, updating the structured asset information flow and automatically completing the "action, result" fragments via the fact chain; the written-back data simultaneously enters the offline playback buffer for incremental updates in the next round of reinforcement learning. Through a closed loop of "state alignment, action pruning, constraint awareness, strategy clearing, explanation writing back, and fact chain backfilling", step S5 and step S4 form a mutually supportive collaborative mechanism: the former seeks optimization within the executable space, while the latter improves the quality of judgment with structural and temporal evidence. The combination of the two produces a robust set of decision recommendations and maintains the ability to adapt to changes in business.

[0032] Furthermore, step S5 includes: constructing a two-layer state representation using the abnormal warning results and the minimum evidence subgraph as the observation layer, and using the location capacity label, project time limit label, and personnel availability label as the graph semantic layer; and using the location capacity label and project time limit label to prune the action set to form a hard constraint action subset; in the training and evaluation process, multi-objective benefit order synthesis is adopted, first satisfying the risk exposure reduction, then satisfying the timeliness of maintenance plan execution, and finally improving the resource utilization balance; and setting negative memory restrictions on invalid exploration for action fragments that perform poorly in the fact chain; and binding all state slices and action pruning results with time window labels and asset knowledge graph version numbers and writing them back. After the strategy stabilizes, strategy clearing is performed, mapping abstract actions to maintenance windows, allocation paths, and personnel scheduling, and generating an explanation package containing input snapshots, strategy reasoning snapshots, and context fragments. The explanation package is bidirectionally bound to the unified asset entries and written back to the asset knowledge graph. The event bus collects the execution results to update the structured asset information flow, and the fact chain completes the "action and result" fragments. The written-back data enters the offline playback buffer for the next round of incremental updates, thus forming a traceable closed loop for decision suggestion sets.

[0033] Furthermore, step S1 uses an event-driven architecture to access asset status update events in real time during the procurement, acceptance, use, and maintenance processes. It then performs semantic annotation, unstructured information extraction, and structured transformation on the captured event data to form a real-time asset information stream containing multi-dimensional attributes such as time series, personnel, location, project, and maintenance plan. This asset information stream can be directly input into the unified asset item matching and knowledge graph construction steps.

[0034] Furthermore, when performing unified matching processing on structured asset information flows, multimodal feature embedding, similarity matching, and normalization are introduced, and the cross-source mapping rules are dynamically optimized by combining an incremental learning mechanism, wherein: Multimodal feature embedding refers to aligning text logs, structured fields, and time series associated with a unified asset entry under a unified time raster, mapping them through an encoder to an asset event feature representation in the same vector space, and attaching source labels and time window labels to ensure data traceability; Similarity matching refers to determining whether events from different sources belong to the same unified asset item in a candidate unified asset item set based on a dual-channel order decision of semantic vector similarity and temporal consistency, and outputting a unique merging result within the matching threshold. Normalization refers to performing quantile normalization on the key numerical features and related indicators of the embedded vector according to the location and project dimensions to a closed interval of zero to one, and archiving the mapping parameters, time window labels and asset knowledge graph version numbers for replay and auditing. Dynamically optimizing cross-source mapping rules by combining incremental learning mechanisms refers to using the fact chain composed of confirmed matching and rejection samples as supervision, updating the embedding parameters, matching threshold, and rule strength online under the constraints of minimum evidence number and small step size, and binding them to the asset knowledge graph version number to reduce false inclusions and missed matches.

[0035] Furthermore, the graph reasoning algorithm includes: A path pattern library with time-series constraints is established in the asset knowledge graph. Path patterns are defined by relation sequences of no more than two lengths, along with time window labels, source labels, relation types, and role labels. Attribute nearest neighbor channels are constructed, embedding the text, structured fields, and time series corresponding to unified asset entries into a shared vector space. A fixed number of nearest neighbors are selected to generate candidate links under the criteria of category consistency and quantile normalization. A context consistency gate is introduced to screen the feasibility of candidate links using location capacity labels and project time limit labels. Candidates meeting the capacity and time limit requirements are retained, while those not meeting these requirements are excluded. Select and transfer to cold storage and record the reason; perform intersection and union judgment on the output of path pattern channel and attribute nearest neighbor channel, the intersection is directly stored in the database, and the union is marked as a link to be verified; maintain the credibility of each candidate link, the credibility is updated according to the occurrence count and recent confirmed events according to the time decay function, and the link below the threshold is transferred to cold storage and the evidence reference is retained; use the change event of unified asset entry as the trigger to incrementally recalculate the candidate link and credibility in its K-hop local neighborhood, and bind all new, update and cold storage operations to the asset knowledge graph version number and write the explanation information, which is used as a traceable context for subsequent graph representation and anomaly identification.

[0036] Furthermore, step S4 includes: constructing a dual-channel graph neural network with relation type awareness and event string temporal encoding within the version freeze view of the asset knowledge graph; generating node representations and edge representations by weighted message passing according to relation type; and merging the structural channels and temporal channels at the node and edge levels using gating; simultaneously applying contrast constraints to the same node and its neighborhood distribution in adjacent versions to stabilize cross-version representations; calculating local consistency scores based on the resource conflict pattern library; introducing location capacity labels and project time limit labels as hard constraint filters to form an anomaly candidate list; comparing the differences of usage allocation vectors for location and project dimensions within the same time window label; and combining the overlapping relationship between migration trajectories and maintenance plans of adjacent versions to complete anomaly sorting; generating a minimum evidence subgraph and explanation field by backtracking the path within the limited number of hops; writing the anomaly warning results and anomaly root cause candidate set, along with the time window label and version number, back to the asset knowledge graph and binding them with unified asset entries for use as traceable inputs for subsequent strategy evaluation and optimization.

[0037] Furthermore, the local consistency score is obtained as follows: within the version-frozen view of the asset knowledge graph, node representations and edge representations, as well as relationship types, role labels, and time window labels are read for the target asset and relationship instances. Within a limited number of hops neighborhood, neighborhood representations are aggregated according to relationship type and role label to form a structural prototype. A temporal prototype is generated based on the time sequence and change interval of the edge event string. The similarity and continuity between the target instance representation and the structural prototype and temporal prototype are jointly measured as a consistency measure. When the constraints of location capacity label and project time limit label are not met, a penalty is applied and a local consistency score is output. The significance of the local consistency score is to characterize the degree of fit between the target instance and its neighborhood structural pattern and temporal evolution trajectory. A lower local consistency score indicates deviation from the normal pattern and potential resource conflicts or unbalanced utilization. The purpose of the local consistency score is to generate and sort abnormal candidates, trigger thresholded alarms, and drive the minimum evidence subgraph backtracking. At the same time, it is bound to the unified asset entry for direct data retrieval and interpretation for strategy evaluation and optimization in step S5.

[0038] Example 2: To address the issue of fragmented enterprise asset usage intensity and reliability signals, preventing the stable triggering of executable maintenance windows within a unified semantic framework, this embodiment of the invention further includes an enterprise asset maintenance window determination step, specifically comprising: The load fluctuation index and the reliability index are introduced into the asset knowledge graph. The load fluctuation index is calculated by weighting the peak load to average load ratio of the asset's historical usage curve with the operating condition disturbance frequency. The reliability index is calculated by combining the asset's average failure interval with the logarithmic ratio of usage time to maintenance interval. The load fluctuation index and the reliability index are input into the neural fuzzy inference model to obtain the operation fitness coefficient. The operation fitness coefficient is used to determine whether there is a need for early maintenance adjustment. The operational adaptability coefficient reflects the level of asset usage pressure and reliability under different scenarios, and is used to determine whether the maintenance plan needs to be adjusted in advance. It is dimensionless. In summary, the adaptive coefficient forms a stable triggering logic based on the hysteresis threshold, avoiding frequent switching caused by short-term noise. The triggering result is bound to a unified asset entry and stored on disk, along with the dynamic trend indicator M, which serves as a state element and fact chain evidence in the offline replay buffer. This drives reinforcement learning to converge to an executable set of maintenance and allocation strategies and decision suggestions. This solves the problems of inconsistent data across different sources, unstable alarm triggering, unbalanced maintenance timing, and uneven resource utilization.

[0039] Furthermore, the method for obtaining the operational fitness coefficient is as follows: Calculate the load fluctuation index and the reliability index within the same time window label; perform quantile normalization on the two indices to a closed interval of zero to one according to the location and project dimensions, and archive the mapping parameters, time window labels, and asset knowledge graph version number. The two normalized inputs are then fed into the membership function for context shaping and rule calculation to generate the operational fitness coefficient; when the location or project sample is insufficient, call the category-level quantile snapshot or the previous version snapshot of similar assets and indicate the source; use the two normalized inputs as the unique values ​​for membership function calculation, and the original index is only used for retention and traceability; define three sets of membership functions for the two inputs respectively: "low," "medium," and "high," and then... The membership function position and span are context-shaped based on category information, location characteristics, and maintenance plan rigor in the asset knowledge graph. Nine interpretable rules are established according to the three-by-three combination of two input semantics, and the rule strength is set. The membership degree of the two inputs on the corresponding membership function is calculated, and the activation degree of each rule is obtained by product conjunction. All rule conclusions are aggregated, and centroid defuzzification is performed using the representative values ​​of the output semantics "low", "medium", and "high" to obtain the running fitness coefficient in the range of zero to one. The running fitness coefficient, trigger rule list, time window label, and asset knowledge graph version number are written back and used as state variables for maintaining advance adjustment judgments and strategy optimization.

[0040] The definitions of the three sets of membership functions are as follows: Three sets of semantic membership functions are defined for each input, named "low," "medium," and "high," with curve types of trigonometric or trapezoidal functions. The initial position is determined based on the low, medium, and high quantiles of the historical distribution of similar assets at the current location and in the project dimension. Context shaping is controlled by two types of parameters: offset, used for finite displacement along the input axis to adjust the sensitive interval, with a value being a bounded real number within the input interval; and scaling factor, used to expand or shrink the coverage span, with a value being a bounded real number greater than zero. Shaping parameters, along with time window labels and asset knowledge graph version numbers, are archived to ensure reproducibility. The explanation and description of the nine interpretable rules, along with their definitions of strength, include: Establish the following nine rules, where the antecedent is a combination of two input semantics, the consequent is the output semantics, the rule strength is a positive real number, and these rules are archived: R1: Low load and high reliability → Low output; R2: Low load and medium reliability → Low output; R3: Low load and low reliability → Output is in progress; R4: Medium load and high reliability → Medium output; R5: Medium load and medium reliability → Medium output; R6: Medium load and low reliability → High output; R7: High load and high reliability → Output in progress; R8: High load and medium reliability → High output; R9: High load and low reliability → High output.

[0041] Among them, the rule strength is used to express prior confidence and subsequent correction magnitude. The value range is from the open interval zero to the closed interval one. The initial value is set based on category engineering experience and is bound to the fact chain evidence when updated.

[0042] The definition of "activation level" is as follows: Activation is a measure of the validity of a single rule. It is obtained by conjunctive operation of the membership degree of the two inputs (the two inputs refer to the quantile normalized values ​​of the load fluctuation index and the reliability index entering the neural fuzzy inference model) in the corresponding semantics of the rule antecedent, and multiplied by the rule strength to form a weighted activation. The conjunctive operation is fixed as a product conjunctive operation. The activation value is in the range of zero to one.

[0043] Explain the definition of the representative values ​​of the semantic terms "low", "medium", and "high". The output semantic set is "low", "medium" and "high", and the corresponding representative values ​​are denoted as low output representative value, medium output representative value and high output representative value. The three are in the range of zero to one and satisfy a strict order relationship: low output representative value is less than medium output representative value and medium output representative value is less than high output representative value. The representative value is set when the category is initialized and archived with the version number. It is used for weighted centroid calculation during defuzzification.

[0044] Furthermore, the maintenance window determines the operation adaptability coefficient as the criterion and sets the hysteresis threshold, with the upper threshold and the lower threshold forming a threshold band; When the operational adaptability coefficient is higher than the upper threshold, the executable maintenance window is determined based on the location capacity label, project time limit label and personnel availability label in the asset knowledge graph. The maintenance work package and bill of materials are cleared and alternative solutions and cross-location transfer paths are generated. When the operating adaptability coefficient is within the threshold zone, the observation window and re-judgment mechanism are activated. By shortening the inspection cycle, limiting short-term high loads, and coordinating online diagnosis with project-wide redistribution or short-term allocation, and at the preset re-judgment time point, the dynamic trend indicator M is used to determine whether to upgrade to early maintenance. When the operational adaptability coefficient is below the lower threshold, the original maintenance plan shall be maintained. All judgments and actions are interpreted, written back to the asset knowledge graph and bound to the unified asset entry, while the fact chain is updated and the data enters the offline replay buffer.

[0045] Furthermore, the initialization of the hysteresis threshold is set based on category engineering experience. Subsequently, the false trigger rate and missed trigger rate obtained from the fact chain statistics are used as correction signals for small-step adjustments. The threshold parameters, asset knowledge graph version number, and time window label are archived together to achieve traceable reproducibility during the generation of decision suggestion sets and replay auditing.

[0046] For ease of understanding, the calculation method for the fitness coefficient is given below: ; Where X represents the original value of the load fluctuation index, a real number, which is the source quantity used for normalization input; Y represents the original value of the reliability index, a real number, which is the source quantity used for normalization input; x represents the normalized load fluctuation index, x∈[0,1], which is used as the input of the neural fuzzy inference model; y represents the normalized reliability index, y∈[0,1], which is used as the input of the neural fuzzy inference model. The function represents the empirical quantile mapping based on location ℓ and project p, which is monotonically increasing and maps the original quantity to [0,1]; τ represents the time window label, which identifies the scroll window used for normalization; ν represents the asset knowledge graph version number, which identifies the normalization and data retrieval context; ; in, Let (j,k) represent the activation level of rule (j,k). ∈[0,1], which measures the strength of the rule's validity; Let (j,k) represent the rule strength. ∈ (0,1], representing prior confidence and participating in subsequent small-step corrections; Let represent the membership function value of input 1 (normalized load fluctuation index) with respect to semantics j∈{L,M,H}. ∈[0,1]; This represents the membership function value of input two (the normalized reliability index) with respect to the semantics k∈{L,M,H}. ∈[0,1]; {L,M,H} represents the semantic set, which represents the three groups of membership functions: low, medium, and high, respectively; Note: Membership functions use triangular or trapezoidal curves, and their positions and spans are context-shaped based on asset class, location characteristics, and maintenance plan rigor (recorded with offset and scaling factor and archived). ; This represents a low-output semantic value used for defuzzing numerical characterization. The output represents the semantic value. This indicates the high-output semantic representative value. The consequent of rule (j,k) represents a value that corresponds one-to-one with "output low / medium / high". It is used to map interpretable semantics to the [0,1] scalar space and is used for centroid defuzzification. ; S represents the operational adaptability coefficient, S∈[0,1], which represents the overall level of stress resistance and reliability in the current context.

[0047] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing an enterprise asset data platform driven by artificial intelligence, characterized in that: include: Step S1: Integrate asset status update events from procurement, acceptance, use and maintenance stages through an event-driven architecture, and perform semantic annotation, entity extraction and structure transformation on the raw event data to form an asset information flow; Step S2: Perform multimodal feature embedding and similarity matching on the asset information stream, map and merge it into a unified asset entry, and update the cross-source asset mapping rules based on the incremental learning mechanism to obtain a unified asset record set; Step S3: Construct a dynamic asset knowledge graph based on a unified asset record set, establish multi-dimensional associations between assets and locations, projects, personnel and maintenance plans, and use graph reasoning algorithms to complete and update potential links; Step S4: Train and apply a graph neural network model on the asset knowledge graph to generate representations of nodes and edges, and combine structural and temporal features to identify abnormal patterns of resource conflicts and uneven utilization, and output abnormal warning results and abnormal root cause candidate set. Step S5: Using the anomaly warning results and unified asset entries as input, perform strategy evaluation and optimization. Through strategy evaluation and reinforcement learning, obtain adaptive adjustment suggestions for maintenance plans and allocation strategies, and output a set of decision suggestions for management deployment.

2. The method for constructing an AI-driven enterprise asset data platform according to claim 1, characterized in that, By using an event-driven architecture, real-time asset status update events from the procurement, acceptance, use, and maintenance stages are accessed, and semantic annotation, unstructured information extraction, and structured transformation are performed on the captured event data to form a real-time asset information stream.

3. The method for constructing an AI-driven enterprise asset data platform according to claim 1, characterized in that, When performing unified matching processing on structured asset information flows, multimodal feature embedding, similarity matching, and normalization are introduced, and incremental learning mechanisms are combined to dynamically optimize cross-source mapping rules, where: Multimodal feature embedding refers to aligning text logs, structured fields, and time series associated with a unified asset entry under a unified time raster, and mapping them through an encoder to an asset event feature representation in the same vector space; Similarity matching refers to determining whether events from different sources belong to the same unified asset item in a candidate unified asset item set based on a dual-channel order decision of semantic vector similarity and temporal consistency, and outputting a unique merging result within the matching threshold. Normalization refers to performing quantile normalization on the key numerical features and related indicators of the embedded vector according to the location dimension and the project dimension to a closed interval of zero to one. Dynamically optimizing cross-source mapping rules by combining incremental learning mechanisms refers to using the fact chain consisting of confirmed matching and rejection samples as supervision, and updating the embedding parameters, matching threshold, and rule strength online while satisfying the constraints of minimum number of evidence items and small step size.

4. The method for constructing an AI-driven enterprise asset data platform according to claim 1, characterized in that, The graph reasoning algorithm includes: A path pattern library with time-series constraints is established in the asset knowledge graph; an attribute nearest neighbor channel is constructed, embedding the text, structured fields, and time series corresponding to unified asset entries into a shared vector space, and a fixed number of nearest neighbors are selected to generate candidate links under the criteria of category consistency and quantile normalization; a context consistency gate is introduced to screen the feasibility of candidate links using location capacity labels and project time limit labels, retaining candidates that meet the capacity and time limit requirements, and transferring candidates that do not meet the requirements to cold storage with the reasons recorded; an intersection and union judgment is performed on the outputs of the path pattern channel and the attribute nearest neighbor channel, the intersection is directly stored in the database, and the union is marked as a link to be verified; a credibility score is maintained for each candidate link, which is updated according to the occurrence count and recent confirmed events using a time decay function, and links below the threshold are transferred to cold storage with evidence citations retained.

5. The method for constructing an AI-driven enterprise asset data platform according to claim 1, characterized in that, Step S4 includes: Within the version-frozen view of the asset knowledge graph, a dual-channel graph neural network with relation type awareness and event string temporal encoding is constructed. Node representations and edge representations are generated by message passing weighted according to relation type. The structural channels and temporal channels are merged at the node level and edge level with gating. At the same time, contrast constraints are applied to the same node and its neighborhood distribution in adjacent versions to stabilize cross-version representations. Local consistency scores are calculated based on a resource conflict pattern library. Location capacity labels and project time limit labels are introduced as hard constraint filters to form an anomaly candidate list. The usage allocation vectors of location and project dimensions are compared within the same time window label, and the anomaly sorting is completed by combining the overlap relationship between adjacent version migration trajectories and maintenance plans. Generate the minimum evidence subgraph and explanation field by backtracking the path within a limited number of hops.

6. The method for constructing an AI-driven enterprise asset data platform according to claim 5, characterized in that, The method for obtaining the local consistency score is as follows: Within the version-frozen view of the asset knowledge graph, node and edge representations, as well as relationship types, role labels, and time window labels, are read for target assets and relationship instances. Within a limited number of hops, neighborhood representations are aggregated by relationship type and role label to form a structural prototype. A temporal prototype is generated based on the time sequence and change interval of edge event sequences. The similarity and continuity between the target instance representation and the structural and temporal prototypes are jointly measured as a consistency measure. When the constraints of location capacity label and project time limit label are not met, a penalty is applied and a local consistency score is output. The significance of the local consistency score is to characterize the degree of fit between the target instance and its neighborhood structural pattern and temporal evolution trajectory. A lower local consistency score indicates deviation from the normal pattern and potential resource conflicts or unbalanced utilization.

7. The method for constructing an AI-driven enterprise asset data platform according to claim 6, characterized in that, It also includes the steps for determining the enterprise asset maintenance window, including: A load fluctuation index and a usage reliability index are introduced into the asset knowledge graph. The load fluctuation index is calculated by weighting the ratio of peak load to average load of the asset's historical usage curve with the frequency of operating condition disturbances. The usage reliability index is calculated by combining the average failure interval of the asset with the logarithmic ratio of usage time to maintenance interval. The load fluctuation index and the usage reliability index are input into a neural fuzzy inference model to obtain the operating fitness coefficient. When the operating fitness coefficient is within the hysteresis threshold band, only the membership snapshot and trigger rule number are recorded without performing parameter updates. Parameter updates are performed only when the coefficient is higher than the upper threshold or lower than the lower threshold.

8. The method for constructing an AI-driven enterprise asset data platform according to claim 7, characterized in that, The method for obtaining the operational fitness coefficient includes: calculating the load fluctuation index and the reliability index within the same time window label, and performing quantile normalization to a closed interval of zero to one according to the location and project dimensions; when there are insufficient location or project samples, calling the category-level quantile snapshot or the previous version snapshot of the same type of asset and marking the source; entering the normalized data into the neural fuzzy inference model, defining membership functions respectively, and performing context shaping according to asset category, location characteristics, and maintenance plan rigor; setting interpretable rules and setting rule strength, calculating membership degree and obtaining the activation degree of each rule by product conjunction, and performing centroid defuzzification after aggregating the rule conclusions to obtain the operational fitness coefficient in the range of zero to one.

9. The method for constructing an AI-driven enterprise asset data platform according to claim 8, characterized in that, The initialization of the hysteresis threshold is based on category engineering experience. Subsequently, it is adjusted step by step using the false trigger rate and missed trigger rate obtained from the fact chain statistics as correction signals. The threshold parameters, asset knowledge graph version number, and time window label are archived together to achieve traceable reproducibility during the generation of decision suggestion sets and replay auditing.

10. An AI-driven enterprise asset data platform construction system, used to implement the method described in any one of claims 1-6, characterized in that, include: The event access module uses an event-driven architecture to access asset status update events from the procurement, acceptance, use and maintenance stages, and performs semantic annotation, entity extraction and structure transformation on the accessed raw event data to form an asset information flow. The multimodal matching and merging module performs multimodal feature embedding and similarity matching on the asset information flow, maps and merges them into unified asset entries, and updates the cross-source asset mapping rules based on an incremental learning mechanism to obtain a unified asset record set. The knowledge graph construction module builds a dynamic asset knowledge graph based on a unified asset record set, establishing multi-dimensional associations between assets and locations, projects, personnel and maintenance plans, and using graph reasoning algorithms to complete and update potential links; The graph representation anomaly identification module trains and applies a graph neural network model on the asset knowledge graph to generate representations of nodes and edges, and combines structural and temporal features to identify abnormal patterns of resource conflicts and uneven utilization, outputting anomaly warning results and anomaly root cause candidate set. The strategy evaluation module takes anomaly warning results and unified asset entries as input to perform strategy evaluation and optimization, and obtains a maintenance plan through strategy evaluation and reinforcement learning.

Images