A device and material inventory and asset verification management system combined with digital twinning

Abstract

The application discloses a kind of equipment and material inventory and asset verification management system combined with digital twin, and the application relates to the technical field of equipment and material management and asset verification, comprising: global perception layer, is deployed in the area where various assets of enterprise are, for collecting the real-time data of asset, it includes the passive ultra-high frequency RFID tag attached to asset, the Beidou positioning module for indoor and outdoor positioning and Bluetooth beacon.The equipment and material inventory and asset verification management system combined with digital twin break the physical boundary of warehouse, realize the continuous tracking and state perception of assets in various scenes.Optimal execution scheme is automatically generated during inventory verification, and the execution result is returned in real time to form a closed loop.Asset management is upgraded from passive recording to active deduction and self-organization execution, which improves asset utilization and management transparency, provides accurate decision support for enterprise operation, and realizes global coverage and intelligent closed loop of asset management.

Description

Technical Field

[0001] This invention relates to the field of equipment and materials management and asset verification technology, specifically to an equipment and materials inventory and asset verification management system that combines digital twins. Background Technology

[0002] As companies expand their asset scale and the types and quantities of equipment and materials become increasingly complex, efficient and accurate inventory checks have become a key challenge in operations management. This is especially true in asset-intensive industries such as power, manufacturing, and transportation, where high-value spare parts, critical equipment, and tools are distributed across warehouses, workshops, and open storage yards, directly impacting cost control and safe production. However, most companies still rely on manual inventory checks, leading to persistent problems.

[0003] First, discrepancies exist between records and actual assets. Traditional verification relies on manual scanning of barcodes or record numbers, which is inefficient and prone to errors. Asset status and information ledgers are outdated, and year-end inventory checks take weeks and are often inaccurate. Second, the management process lacks transparency and traceability is difficult. Changes in asset status throughout their lifecycle rely on paper documents or post-event entries, making it impossible to obtain real-time information on actual location, user, and status. Lost assets or accidents are difficult to trace. Third, locating items is inefficient, impacting production continuity. Finding spare parts during emergency repairs relies on manual experience, lacking location guidance and slowing down repairs. Finally, management decisions lack data support. Management cannot access asset distribution heatmaps, idle status, and usage frequency, hindering informed purchasing and allocation decisions and increasing waste.

[0004] To address the aforementioned issues, some explorations have been undertaken in existing technologies. For example, patent document CN121526475A discloses an intelligent hydropower station material warehouse management system based on digital twin technology. This system constructs a perception and control layer, a data transmission and processing layer, and an application and display layer, utilizing RFID technology, IoT sensors, and digital twin models to achieve location and environmental monitoring within the material warehouse. However, the application scenarios of this solution are mainly limited to enclosed warehouse spaces, focusing on material flow management. It is difficult to achieve effective coverage and real-time inventory for fixed assets such as production equipment and special vehicles distributed across a wide factory area. Furthermore, its system design emphasizes guiding the inbound and outbound processes, lacking in-depth functions for core asset verification business scenarios, and thus failing to meet the enterprise's needs for dynamic asset verification and full lifecycle management across the entire asset domain.

[0005] To address this, the present invention proposes an equipment and material inventory and asset verification management system that combines digital twins. This system aims to break down the physical boundaries of warehouses, enable real-time perception and dynamic verification of cross-regional and multi-type assets, and solve the problems of limited asset management scope, low verification efficiency, and lack of data closure in existing technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a digital twin-based equipment and material inventory and asset verification management system to solve the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a digital twin-based equipment and material inventory and asset verification management system, comprising: The global perception layer is deployed in the areas where various assets of the enterprise are located to collect real-time data of the assets. It includes passive UHF RFID tags attached to the assets, Beidou positioning modules and Bluetooth beacons for indoor and outdoor positioning, and vibration sensors and current transformers for collecting asset operating status. The data fusion layer is communicatively connected to the global perception layer. It includes an asset data platform, which has a built-in multimodal data fusion engine and a dynamic verification module. The dynamic verification module performs multiple rounds of progressive verification and review on the fused data. The spatiotemporal causal engine layer communicates with the data fusion core layer. It has a built-in triple adversarial verification mechanism for performing timestamp entropy detection, spatial trajectory game verification, and behavioral causal verification on asset data streams, and triggers a closed-loop review process for data with doubts. The twin inference layer is communicatively connected to the spatiotemporal causal engine layer. It is loaded with a five-dimensional asset twin that precisely corresponds to the enterprise's physical space. The five-dimensional asset twin adds a time dimension and a state dimension to the three-dimensional spatial model to achieve real-time synchronization of asset status, behavior prediction, and strategy inference. The self-healing execution layer is communicatively connected to the twin inference layer. It includes an intelligent scheduling gateway and various execution terminals, which are used to convert the decision instructions of the twin inference layer into physical control actions and to provide feedback verification on the execution results. The real-time data collected by the global perception layer is fused by the data fusion layer and then enters the spatiotemporal causal engine layer for multiple verifications and reviews. The high-confidence data after verification drives the five-dimensional asset twin in the twin inference layer to keep synchronized with the physical asset. The self-healing execution layer receives the instructions from the twin inference layer and executes on-site operations. The data collected during the execution process is sent back to the data fusion layer to form a closed loop.

[0008] The data flow between different levels forms a complete closed-loop circulation system. Changes in the status of physical assets can be fed back to the virtual twin in real time, and decision-making instructions from virtual simulations can be accurately implemented in physical operation links, realizing dynamic management of assets.

[0009] Preferably, the spatiotemporal causal engine layer constructs an asset behavior spatiotemporal graph, where each asset node is associated with its historical trajectory sequence, environmental interaction records, and state change events. The triple adversarial verification mechanism verifies the self-consistency of real-time data streams based on the asset behavior spatiotemporal graph. The asset behavior spatiotemporal graph uses a sliding time window mechanism to construct asset behavior snapshots, enabling structured storage and association of asset-related data, providing a complete asset data reference for data self-consistency verification.

[0010] Preferably, the triple adversarial verification mechanism includes: The first layer is the timestamp entropy detection unit, which is used to calculate the timestamp distribution entropy value of each asset data stream and compare it with the historical normal mode entropy value of the asset. When the timestamp distribution entropy value shows abnormal fluctuations or the time interval shows regular deviations, it is determined that there is a data collection anomaly and the anti-spoofing verification process is triggered. The second layer is the spatial trajectory game verification unit, which is used to conduct a three-way game verification of the asset's location trajectory with RFID reading records and visual recognition records to build a location evidence chain. It requires that at least two sensing methods generate mutually corroborating location evidence for the same asset within the same spatiotemporal window. If the asset is only captured by a single sensing method and there is a logical conflict with other sensing data, the data is determined to be isolated evidence and a review task is generated. The third layer is the behavior causality verification unit, which calls the asset behavior time sequence logic rule library to perform causal consistency checks on asset status changes. If the status change event lacks a preceding causal relationship or the causal order is logically reversed, the status change is determined to be invalid and the self-healing mechanism is triggered.

[0011] The triple adversarial verification mechanism performs layered verification of asset data from three dimensions: time, space, and behavioral causality. Each verification unit can independently determine and trigger subsequent review processes, achieving comprehensive verification of asset data.

[0012] Preferably, the spatiotemporal causal engine layer also includes a closed-loop self-verification module. When there are still doubts after the data has undergone triple adversarial verification, the closed-loop self-verification module issues a targeted review instruction to the global perception layer, requiring multiple types of perception terminals in the relevant area to conduct collaborative re-collection and re-inject the newly collected data into the triple adversarial verification mechanism for secondary verification, until a high-confidence evidence chain is formed by mutual verification by at least three independent perception methods.

[0013] The closed-loop self-verification module can generate personalized review and collection instructions based on the type of data doubt. Through multiple rounds of collaborative re-collection and secondary verification, it gradually improves the evidence chain of asset data and ensures the verification effect of doubtful data.

[0014] Preferably, when the spatial trajectory game verification unit constructs the location evidence chain, it specifically performs spatiotemporal correlation analysis on the asset location information read by the RFID reader, the continuous trajectory information collected by the Beidou positioning module, and the asset image feature information extracted by the industrial vision unit. If there is a logical contradiction among the three, it is determined to be isolated evidence data and the review process is initiated.

[0015] The spatial trajectory game verification unit transforms the three types of location information into the same spatial reference system and then conducts correlation analysis. Through the mutual verification of multi-dimensional location data, logical contradictions in asset location data can be accurately identified.

[0016] Preferably, when performing causal consistency checks, the behavioral causal verification unit specifically determines whether the asset status change event is correlated in time series with the corresponding work authorization order, access gate reading record, and operator trajectory information. If any preceding correlated event is missing or the event occurrence sequence does not conform to the preset logical rules, the status change is deemed invalid. The behavioral causal verification unit completes causal consistency checks by comparing the time series of multiple types of business data, which can accurately identify invalid events in asset status changes that lack business support, ensuring the rationality of asset status changes.

[0017] Preferably, the spatiotemporal causal engine layer further includes an anti-spoofing verification module. After the timestamp entropy value detection determines that there is a data collection anomaly, the anti-spoofing verification module sends a high-frequency sampling command to the global perception layer, requiring the relevant perception terminals to encrypt the collected data within a preset time period, and then re-injects the encrypted collected data into the triple adversarial verification mechanism for verification.

[0018] The anti-spoofing verification module can match the corresponding high-frequency sampling parameters according to the type of data anomaly, and obtain more dense asset data through encrypted collection, providing more sufficient evidence for the verification of time-dimensional anomaly data.

[0019] Preferably, the twin simulation layer incorporates a behavior prediction model and a strategy simulator. The behavior prediction model, based on the real-time state evidence chain and historical change records associated with the five-dimensional asset twin, predicts the state evolution trend of the asset in future time periods and identifies potential anomaly risks. Upon receiving an inventory verification instruction, the strategy simulator automatically generates multiple execution plans based on the current spatiotemporal distribution of the asset and the coverage of sensing terminals. It then simulates the execution process of each plan in parallel within the five-dimensional asset twin, predicts execution time and potential blind spots, selects the optimal plan, decomposes the task, and distributes it to the execution terminal. The behavior prediction model can predict state change trends based on historical and real-time asset data, while the strategy simulator compares and selects execution plans through parallel simulation of multiple plans, ensuring the rationality and efficiency of the inventory verification operation plan.

[0020] Preferably, the twin simulation layer further includes an asset health assessment unit. The asset health assessment unit calculates an asset health index by combining asset operation data, idle time and environmental parameters, and displays the risk distribution in a visual form in the five-dimensional asset twin. When the idle rate of a certain type of asset exceeds a set threshold, the strategy simulator generates an allocation optimization suggestion based on historical allocation data and current demand forecasts, and outputs the decision basis after simulating the implementation effect of the scheme in the five-dimensional asset twin.

[0021] The asset health assessment unit sets indicator weights according to asset type to complete the health index calculation. The risk distribution can be presented intuitively in a visual form. The allocation optimization suggestions are output after virtual simulation, providing detailed reference for asset allocation decisions.

[0022] Preferably, the global perception layer further includes an industrial vision unit and a microelectromechanical inertial sensor. The industrial vision unit is deployed in key areas and extracts the appearance features and position offset information of the asset in real time through edge computing, generates a visual feature vector, and forms a multimodal association with the perception data. The microelectromechanical inertial sensor is attached to the moving asset and is used to collect the asset's instantaneous acceleration and attitude change data. The multimodal data fusion engine performs spatiotemporal alignment and feature fusion of the visual feature vector, inertial sensing data, and RFID positioning data.

[0023] Industrial vision units extract asset features in real time through edge computing, microelectromechanical inertial sensors can accurately collect motion data of mobile assets, and the fusion of multiple types of sensing data can enrich the dimensions of asset data and improve the comprehensiveness of asset status perception.

[0024] This invention provides a digital twin-based equipment and materials inventory and asset verification management system. It offers the following advantages: This equipment and material inventory and asset verification management system, which combines digital twins, utilizes a triple adversarial verification mechanism built into its spatiotemporal causal engine layer. The system performs timestamp entropy detection, spatial trajectory game theory verification, and behavioral causal verification on each piece of asset data, effectively filtering out abnormal noise and logically contradictory data. For questionable data, a closed-loop self-verification process is initiated, requiring multiple types of sensing terminals to collaboratively re-collect data until a high-confidence evidence chain is formed, mutually corroborating at least three independent sensing methods. This mechanism fundamentally solves the problems of unreliable data sources and long-term discrepancies between accounts and physical assets in traditional asset management. It enables the asset status driven by the digital twin model to truly update synchronously with physical reality, providing an accurate and reliable data foundation for all subsequent management and decision-making.

[0025] This digital twin-based equipment and material inventory and asset verification management system integrates multiple technologies such as RFID, BeiDou positioning, Bluetooth beacons, vibration sensing, industrial vision, and microelectromechanical inertial sensing in its full-domain perception layer. This breaks down the physical boundaries of warehouses, enabling continuous tracking and status awareness of assets in various scenarios, including production workshops, open-air storage yards, and office areas. The twin simulation layer, based on a five-dimensional asset twin, performs behavior prediction and strategy simulation. During inventory verification, it automatically generates optimal execution plans and simulates optimized allocation paths when assets are idle. A self-healing execution layer translates decision-making instructions into physical control actions, with execution results transmitted back in real time, forming a closed loop. This design upgrades asset management from passive recording to proactive simulation and self-organizing execution, improving asset utilization and management transparency, providing precise decision support for enterprise operations, and achieving full-domain coverage and intelligent closed-loop asset management. Attached Figure Description

[0026] Figure 1 This is a data flow diagram of an equipment and material inventory and asset verification management system that combines digital twins according to the present invention; Figure 2 This is a flowchart of a triple-adversarial verification mechanism for an equipment and material inventory and asset verification management system that combines digital twins, according to the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 1 and Figure 2 This invention provides a technical solution: a digital twin-based equipment and material inventory and asset verification management system, comprising: The global perception layer is deployed in the areas where various assets of the enterprise are located to collect real-time data of the assets. It includes passive UHF RFID tags attached to the assets, Beidou positioning modules and Bluetooth beacons for indoor and outdoor positioning, and vibration sensors and current transformers for collecting asset operating status. The data fusion core layer communicates and connects with the full-domain perception layer. It includes an asset data platform, which has a built-in multimodal data fusion engine and a dynamic verification module. The dynamic verification module performs multiple rounds of progressive verification and review on the fused data. The spatiotemporal causal engine layer communicates with the data fusion core layer. It has a built-in triple adversarial verification mechanism to perform timestamp entropy detection, spatial trajectory game verification, and behavioral causal verification on asset data streams, and trigger a closed-loop review process for data with doubts. The twin inference layer communicates with the spatiotemporal causal engine layer. It is loaded with a five-dimensional asset twin that precisely corresponds to the enterprise's physical space. The five-dimensional asset twin adds a time dimension and a state dimension to the three-dimensional spatial model to achieve real-time synchronization of asset status, behavior prediction and strategy inference. The self-healing execution layer communicates with the twin simulation layer. It includes an intelligent scheduling gateway and various execution terminals, which are used to convert the decision instructions of the twin simulation layer into physical control actions and to provide feedback verification on the execution results. The intelligent scheduling gateway is an integrated hardware and software industrial gateway that combines command parsing, protocol conversion, task scheduling, status feedback, and anomaly warning. Its core hardware is equipped with an industrial-grade embedded processor, and includes an industrial IoT protocol conversion module and task scheduling algorithm. After receiving decision commands from the twin-inference layer, the gateway first parses the commands, extracts the core task information, and breaks down the tasks according to the type and function of the execution terminals. Then, the protocol conversion module converts the unified command format into a communication protocol suitable for each execution terminal. Communication with fixed execution terminals uses the Modbus-RTU protocol, while communication with mobile execution terminals and automated guided vehicles (AGVs) uses the same protocol. The gateway communicates with the execution terminals via Bluetooth 5.0 and ZigBee protocols, and with the handheld terminals of on-site personnel via 4G / 5G wireless communication protocols. After issuing tasks, the gateway monitors the execution status of each execution terminal in real time and judges the execution progress by the heartbeat packets and status data returned by each terminal. When abnormal situations such as execution terminals going offline or task execution timeout occur, the gateway will immediately trigger an abnormality warning and send the abnormality information back to the twin simulation layer. At the same time, it will reschedule the execution terminals or adjust the task issuance strategy according to the actual situation. After the task is completed, the gateway summarizes and organizes the execution results of all execution terminals and sends them back to the data fusion core layer to complete the data closure.

[0029] Among them, the real-time data collected by the global perception layer is fused through the data fusion layer and then enters the spatiotemporal causal engine layer for multiple verifications and reviews. The high-confidence data after verification drives the five-dimensional asset twin in the twin inference layer to keep synchronized with the physical asset. The self-healing execution layer receives the instructions from the twin inference layer and executes on-site operations. The data collected during the execution process is sent back to the data fusion layer to form a closed loop.

[0030] It should be further explained that the system includes a global perception layer, a data fusion layer, a spatiotemporal causal engine layer, a twin inference layer, and a self-healing execution layer.

[0031] The system employs a combination of wired and wireless communication between its various architecture layers. Between the global perception layer and the data fusion core layer, 5G and LoRa wireless communication protocols are used for outdoor, widely distributed perception nodes, while industrial Ethernet wired communication is used for indoor, fixed-deployment perception nodes and industrial vision units. High-speed, stable wired communication is achieved between the data fusion core layer and the spatiotemporal causal engine layer, and between the spatiotemporal causal engine layer and the twin inference layer, using the TCP / IP protocol via industrial Ethernet. Between the twin inference layer and the self-healing execution layer, the intelligent scheduling gateway communicates with fixed execution terminals via industrial Ethernet, and with mobile execution terminals and field operation terminals via Bluetooth 5.0 and ZigBee wireless communication protocols. All inter-layer data transmission is encapsulated and interacted based on the MQTT IoT communication protocol, ensuring standardized and compatible data transmission.

[0032] The global perception layer is deployed in various areas where the enterprise's assets are located, including production workshops, warehouses, open storage yards and office areas. Passive UHF RFID tags are attached to them as digital identity carriers for assets. At the same time, Beidou positioning modules and Bluetooth beacons are integrated to form a seamless indoor and outdoor positioning network. Vibration sensors and current transformers are additionally configured on key equipment to collect operating status parameters. All perception terminals form a dynamic topology through a self-organizing network protocol.

[0033] The sensing terminals in the global perception layer construct a dynamic topology network through the LoRaWAN self-organizing network protocol. It adopts a hybrid star and mesh topology structure, with IoT gateways deployed in various areas of the enterprise as aggregation nodes and various sensing terminals as sensing nodes. The sensing nodes can automatically discover surrounding gateways and other sensing nodes. When the communication of a node is interrupted, it can automatically switch to a nearby gateway or node for data forwarding, realizing the dynamic reconstruction and self-healing of the self-organizing network. It is suitable for the construction needs of sensing networks in different physical areas such as enterprise warehouses, workshops, and open-air storage yards. The communication distance of the sensing nodes is planned to be 100 meters indoors and 500 meters outdoors to ensure blind-spot-free coverage of the global perception.

[0034] All conventional parameters, preset thresholds, and set thresholds in this system are set according to the general standards of industrial asset-intensive industries and actual application scenarios. Specifically, the conventional sampling frequencies for various sensing nodes in the full-domain perception layer are: Beidou positioning module 1 time / second, Bluetooth beacon 1 time / 2 seconds, vibration sensor 10 times / second, current transformer 5 times / second, industrial vision unit 1 frame / second, and MEMS inertial sensor 20 times / second. The conventional transmission power of the RFID reader is 30dBm indoors and 33dBm outdoors. The preset spatial distance threshold for spatial trajectory game theory verification is 5 meters, and the preset time threshold is 30 seconds. The potential anomaly risk probability threshold in the behavior prediction model is 80%, meaning that a potential anomaly risk is identified when the predicted probability of the asset's future state type is higher than 80%. The idle rate threshold for various assets in the asset health assessment is 60%, meaning that an idle warning is triggered when the proportion of assets with excessive idle time in a certain asset category reaches 60%. The acceleration threshold for the MEMS inertial sensor is 0.5 m / s². 2 The attitude change threshold is 5° / s. If the value exceeds the threshold, the asset is determined to have moved or changed its attitude.

[0035] The data fusion core layer communicates with the global perception layer. Its asset data platform has a built-in multimodal data fusion engine and dynamic verification module. The multimodal data fusion engine corresponds to the multi-source data fusion processing unit in the field of industrial IoT. It is an integrated hardware and software processing module that integrates spatiotemporal alignment module, feature fusion module, data preprocessing module and data association module. The engine is installed in the industrial server of the asset data platform. It can adapt to multi-source heterogeneous data processing of various industrial sensors such as RFID, Beidou positioning, industrial vision, and inertial sensing. It has the ability to receive data in real time, process data at high speed and output data in a standardized manner. It is the core carrier for realizing data fusion of multiple sensing methods. The multimodal data fusion engine performs timestamp alignment and spatial coordinate matching on the received asset identity information, location coordinates, operating parameters, and environmental temperature and humidity data to construct a real-time snapshot of the asset status. The dynamic verification module performs threshold screening, cross-source cross-validation, and historical pattern verification on the fused data to remove abnormal noise data and identify logical contradictions.

[0036] The spatiotemporal causal engine layer communicates with the data fusion core layer and has a built-in triple adversarial verification mechanism. The spatiotemporal causal engine layer corresponds to the causal analysis and data verification platform in the field of industrial big data. It is a software system deployed on the enterprise industrial cloud platform. It integrates the triple adversarial verification mechanism, closed-loop self-verification module, anti-deception verification module and asset behavior spatiotemporal graph management module. It has the core functions of real-time reception of asset data, spatiotemporal causal verification, abnormal data identification, doubtful data review and high-confidence evidence chain construction. It is a key layer to ensure the authenticity and reliability of asset data. This layer can be seamlessly connected with the enterprise's existing industrial big data platform to realize data interconnection and sharing. The first layer is timestamp entropy detection, which calculates the timestamp distribution entropy value of each asset data stream and compares it with the historical normal mode entropy value of the asset. When abnormal fluctuations or regular deviations occur, it is determined that there is an abnormal data collection and triggers anti-spoofing verification. The second layer is spatial trajectory game verification, which conducts a three-way game verification between the asset positioning trajectory and RFID reading records and visual recognition records. The construction of the location evidence chain requires at least two sensing methods to mutually verify each other within the same spatiotemporal window. If the asset is captured by only a single sensing method and there is a logical conflict with other data, it is determined to be isolated evidence and a review task is generated. The third layer is behavioral causality verification, which calls the asset behavior temporal logic rule library to check the causal consistency of state changes. If the state change event lacks a preceding causal relationship or the causal order is logically reversed, it is determined to be invalid and a self-healing mechanism is triggered.

[0037] If any data still has doubts after triple verification, the spatiotemporal causal engine layer issues a targeted review instruction to the global perception layer, requiring multiple types of sensing terminals in the relevant area to re-collect the data collaboratively, and then re-inject the newly collected data into the verification mechanism for secondary verification, until a high-confidence evidence chain is formed that is mutually corroborated by at least three independent sensing methods.

[0038] The twin simulation layer communicates with the spatiotemporal causal engine layer. It is loaded with a five-dimensional asset twin that precisely corresponds to the enterprise's physical space. This twin adds a time dimension and a state dimension to the three-dimensional spatial model. It dynamically binds the real-time status evidence chain of the asset, historical change records, technical parameter files and maintenance work order data. It has a built-in behavior prediction model and strategy simulator. The behavior prediction model infers the future status evolution trend of the asset based on the current state and historical patterns and identifies potential abnormal risks. After receiving the inventory and verification instruction, the strategy simulator automatically generates multiple execution plans based on the current spatiotemporal distribution of the asset and the coverage of the sensing terminal. After simulating the execution process of each plan in parallel in the five-dimensional asset twin, the optimal plan is selected and the task is decomposed and issued.

[0039] The self-healing execution layer communicates with the twin simulation layer, which includes an intelligent scheduling gateway and various execution terminals. When the system detects abnormal asset status or isolated data, it receives a review instruction from the spatiotemporal causal engine layer, and schedules the corresponding area perception terminals and execution terminals to work together to complete on-site review and status correction. When the twin simulation layer generates an allocation optimization plan and passes the approval, it automatically decomposes the plan into specific execution instructions, schedules automated guided vehicles to pick up goods at designated locations and guides operators to complete the warehouse transfer operation. During the operation, data is continuously collected and sent back to the data fusion layer for verification to ensure that the execution results are consistent with the simulation expectations.

[0040] The self-healing execution layer and the operators achieve standardized interaction through the operation terminal. The self-healing execution layer pushes standardized work orders to the operators' handheld terminals and industrial tablets through the intelligent scheduling gateway. The work order content includes the unique asset identifier, work type, verification / operation area coordinates, operation requirements, and completion time limit. After receiving the work order on the operation terminal, the operator confirms it. During the operation, the operation terminal will collect the operator's location trajectory and operation record in real time and send it back to the self-healing execution layer. After the operation is completed, the operator submits the operation results and on-site image data on the terminal. The self-healing execution layer verifies the operation results. If the verification is successful, the task is completed. If the verification fails, a second operation prompt is pushed. At the same time, the entire operation process data is sent back to the data fusion layer for recording.

[0041] The spatiotemporal causal engine layer constructs a spatiotemporal graph of asset behavior. Each asset node is associated with its historical trajectory sequence, environmental interaction records, and state change events. A triple adversarial verification mechanism verifies the self-consistency of real-time data streams based on the spatiotemporal graph of asset behavior.

[0042] It should be further explained that the spatiotemporal graph of asset behavior constructed by the spatiotemporal causal engine layer is a dynamic knowledge network based on graph data structure. Each asset node corresponds to a unique digital identity of the asset. The information associated with the node includes the trajectory sequence formed by the asset on a continuous time axis. This trajectory sequence is formed by spatiotemporally aligning the coordinate points continuously collected by the Beidou positioning module with the Bluetooth beacon positioning data. It is also associated with the record of the interaction between the asset and the surrounding environment. The environmental interaction record includes the reader number and reading timestamp generated when the RFID reader is read, the timestamp of the start and stop events of adjacent equipment sensed by the vibration sensor, and the position change events of surrounding assets captured by the industrial vision unit. It is also associated with the status change events that occur in the asset itself. The status change events include warehousing events, outbound events, relocation events, maintenance events, and scrapping events. Each event is bound to the event occurrence time, event type code, and the work order number or authorized personnel identifier that triggered the event.

[0043] The spatiotemporal map of asset behavior is constructed using a sliding time window mechanism, which divides continuous time into multiple time windows. Within each time window, the trajectory sequence of asset nodes is compressed and sampled to form a trajectory feature vector. Environmental interaction records are clustered to form an interaction event cluster. State change events are sorted chronologically to form an event chain. The trajectory feature vector, interaction event cluster, and event chain are associated and mapped to form a snapshot of asset behavior within that time window.

[0044] The asset behavior snapshot is constructed using a key-value pair structured data structure. The core key values ​​include the asset's unique identifier, the start and end timestamps of the time window, the trajectory feature vector, the interaction event cluster, the state change event chain, and the data confidence score. The asset behavior snapshot is stored in JSON format, categorized and stored according to the asset's unique identifier and time window, and stored in the distributed database of the spatiotemporal causal engine layer. When constructing the asset behavior spatiotemporal graph, the asset behavior snapshots of each asset are linked and associated in units of time windows. After completing a triple adversarial verification, the high-confidence data that passes the verification is updated to the latest asset behavior snapshot of the corresponding asset. The updated asset behavior snapshot increment is then integrated into the asset behavior spatiotemporal graph to achieve continuous evolution and self-learning of the graph.

[0045] The triple adversarial verification mechanism, when verifying the self-consistency of real-time data streams, first maps the real-time data stream to the corresponding asset node in the asset behavior spatiotemporal graph, extracts the historical behavior snapshot of the asset node in the current time window as a benchmark, and compares the timestamp distribution characteristics of the real-time data stream with the timestamp distribution characteristics of the historical behavior snapshot. If a regular deviation in time intervals not present in the historical behavior snapshot appears in the real-time data stream, it is judged as an anomaly in the time dimension and triggers anti-spoofing verification. The spatial trajectory game verification unit compares the positioning trajectory in the real-time data stream with the asset behavior spatiotemporal graph. The historical trajectory sequence of the asset node in the graph is compared for trajectory similarity. If the real-time trajectory deviates significantly from the historical trajectory in terms of spatial form and cannot form a spatiotemporal association with the surrounding environment, it is judged as an anomaly in spatial dimension and a review task is generated. The behavior causal verification unit compares the state change event in the real-time data stream with the event chain of the asset node in the asset behavior spatiotemporal graph for causal order. If the real-time state change event is missing a preceding or subsequent event in the event chain, or if the order of the event does not match the inherent order of the same type of event in the historical event chain, it is judged as an anomaly in causal dimension and a self-healing mechanism is triggered.

[0046] After each triple adversarial verification, the spatiotemporal graph of asset behavior incrementally updates the historical behavior snapshots of the corresponding asset nodes in the graph based on the high-confidence data that passed the verification. The newly generated trajectory feature vectors, interaction event clusters and event chains are incorporated into the graph, enabling the graph to have a self-learning ability for continuous evolution.

[0047] The triple adversarial verification mechanism includes: The first layer is the timestamp entropy detection unit, which is used to calculate the timestamp distribution entropy value of each asset data stream and compare it with the historical normal mode entropy value of the asset. When the timestamp distribution entropy value shows abnormal fluctuations or the time interval shows regular deviations, it is determined that there is a data collection anomaly and the anti-spoofing verification process is triggered. The second layer is the spatial trajectory game verification unit, which is used to conduct a three-way game verification of the asset's location trajectory with RFID reading records and visual recognition records to build a location evidence chain. It requires that at least two sensing methods generate mutually corroborating location evidence for the same asset within the same spatiotemporal window. If the asset is only captured by a single sensing method and there is a logical conflict with other sensing data, the data is determined to be isolated evidence and a review task is generated. The third layer is the behavior causality verification unit, which calls the asset behavior time sequence logic rule library to perform causal consistency checks on asset status changes. If the status change event lacks a preceding causal relationship or the causal order is logically reversed, the status change is determined to be invalid and the self-healing mechanism is triggered.

[0048] It should be further explained that the timestamp entropy detection unit performs the following operations: For each asset data stream, a continuous timestamp sequence is extracted in units of a fixed-duration time window, and the distribution entropy value of the time interval in the sequence is calculated. The time interval distribution entropy value is calculated by dividing the interval value of adjacent timestamps within the time window into multiple intervals, counting the frequency of occurrence in each interval, and then substituting it into the entropy value calculation formula to obtain the timestamp distribution entropy value of the time window. The timestamp entropy value is calculated using a fixed-length time window. First, a continuous timestamp sequence of the asset data stream within the time window is extracted, and the time interval between adjacent timestamps is calculated. Then, based on the regular time interval of asset data collection, the time interval is divided into several intervals in an equidistant manner, and the frequency of time intervals in each interval is counted. Based on the frequency distribution of each interval, the timestamp distribution entropy value is calculated. The calculated real-time entropy value is compared with the baseline distribution interval of the asset's entropy value. If the real-time entropy value exceeds the interval range or shows a regular deviation, it is determined that there is a data collection anomaly and the anti-spoofing verification process is triggered.

[0049] The timestamp entropy detection unit retrieves entropy records of the asset node from the spatiotemporal causal engine layer for multiple historical time windows under the same historical period and normal operating conditions, forming the entropy baseline distribution range of the asset node. The real-time calculated time window entropy value is compared with the entropy baseline distribution range. If the real-time entropy value exceeds the upper or lower limit of the baseline distribution range, or if the real-time entropy value shows a monotonically increasing or decreasing trend in multiple consecutive time windows, it is determined that there is a time dimension anomaly in the current data stream.

[0050] The determination of the entropy benchmark distribution interval is based on the historical operating data of the asset under normal working conditions. For each type of asset, the timestamp entropy value data of the asset data stream under normal use, normal inventory, and normal circulation conditions for 180 consecutive days is collected. The collected historical entropy value data is statistically analyzed, and after removing outliers, the mean and standard deviation of the historical entropy value are calculated. The range of the mean ± 2 times the standard deviation is determined as the entropy benchmark distribution interval of the asset type. When the timestamp entropy value of the asset's real-time data stream exceeds this interval, or when the entropy value of 5 consecutive time windows shows a monotonically increasing or decreasing regular change, it is determined that there is an anomaly in the timestamp collection of the asset's data stream.

[0051] After determining that there is an anomaly in the time dimension, the timestamp entropy detection unit sends an anti-spoofing verification command to the global perception layer, requiring the RFID readers, Beidou positioning modules and industrial vision units around the corresponding assets to increase the data collection frequency to three times the normal frequency in the next time window, and to independently encapsulate the encrypted data and re-inject it into the spatiotemporal causal engine layer for secondary verification.

[0052] The spatial trajectory game verification unit performs the following operations: It acquires three sources of location information for the same asset within the same time window: a continuous trajectory coordinate sequence output by the BeiDou positioning module, antenna position coordinates recorded by the RFID reader, and asset image coordinates extracted by the industrial vision unit through image recognition. After uniformly converting these three coordinates to the same spatial reference system, it constructs the asset's trajectory line, discrete point set, and image positioning point set within the time window. The spatial trajectory game verification unit calculates the spatial proximity between the trajectory line and the discrete point set. If more than half of the points in the discrete point set have a shortest distance exceeding a preset threshold from the trajectory line, it determines that there is a contradiction between the RFID reading location and the positioning trajectory. Simultaneously, it calculates the spatial matching degree between the trajectory line and the image positioning point set. If more than half of the points in the image positioning point set cannot find a spatially corresponding point near the timestamp on the trajectory line, it determines that there is a contradiction between the visual recognition location and the positioning trajectory.

[0053] In the spatial trajectory game verification, the determination of spatial proximity and spatial matching degree are both based on a unified spatial coordinate system and preset thresholds. The spatial proximity determination targets the continuous BeiDou positioning trajectory of the asset and the discrete RFID reading point set. First, the shortest straight-line distance from each RFID reading discrete point to the BeiDou positioning trajectory is calculated. Then, the average of the shortest straight-line distances of all discrete points is calculated. If the average is within the preset spatial distance threshold, the spatial proximity is considered to meet the requirements; if the average exceeds the threshold, a spatial logical contradiction is considered to exist. The spatial matching degree determination targets the continuous BeiDou positioning trajectory of the asset. For the tracking and industrial visual recognition point set, the industrial visual recognition point set is first projected onto the Beidou positioning trajectory. The proportion of the number of recognition points that overlap with the trajectory after projection is calculated to the total number of recognition points. If the proportion is higher than 80%, the spatial matching degree between the two is determined to meet the requirements. If the proportion is lower than 80% or there are no overlapping points, it is determined that there is a spatial logical contradiction between the two. When there is a logical contradiction between the Beidou positioning trajectory and the RFID reading discrete point set, the industrial visual recognition point set, or one of them, and it cannot be reasonably explained by factors such as equipment error or scene occlusion, the location data of the asset is determined to be isolated data and the review process is initiated.

[0054] When two contradictions occur simultaneously or one of the contradictions occurs multiple times in a row, the spatial trajectory game verification unit determines that the location information of the asset within the time window is isolated evidence data and generates a review task to be sent to the self-healing execution layer. The review task carries the asset identifier, the suspicious time window, and the coordinates of the spatial area that needs to be checked.

[0055] The behavioral causality verification unit performs the following operations: It retrieves the state change rule chain corresponding to the asset type from the asset behavior temporal logic rule base. The state change rule chain defines the allowed transition directions between various states and the preceding event types that must be associated with each transition in the form of a directed graph. The behavioral temporal logic rule base is constructed according to the type of enterprise assets, divided into four major categories: fixed assets, mobile equipment, spare parts, and tools. Each major category is further subdivided according to the asset's function and usage scenario. The rule base is stored in a structured data table format, with fields including asset type code, state change type code, and preceding event type. The rule base, which includes requirements for asset type coding, the number of preceding events, the order of events, and the valid judgment conditions for state changes, is deployed in the database of the spatiotemporal causal engine layer. It uses a relational database for data management. When the behavioral causal verification unit of the spatiotemporal causal engine layer performs causal consistency checks on asset state changes, it matches the corresponding asset type code based on the asset's unique identifier and combines it with the asset's state change type code. It then retrieves the corresponding verification rules from the rule base in real time through database retrieval statements to complete the causal verification of asset state changes. The rule base allows managers to add, delete, modify, and update rules according to the company's asset management system and actual operational needs. The inbound status can be switched to the outbound status, but it must be associated with the outbound gate reading event and the outbound work order issuance event. The in-stock status can be switched to the transfer status, but it must be associated with the transfer work order generation event and the operator check-in event. The in-use status can be switched to the maintenance status, but it must be associated with the fault repair work order submission event and the shutdown operation confirmation event.

[0056] The behavior causal verification unit captures the sequence of state change events that occur in the asset within the current time window, compares each transition pair consisting of adjacent state change events in the sequence with the state change rule chain, checks whether each transition pair belongs to the allowed transition direction in the rule chain, and checks whether the set of preceding events for each transition pair exists completely in the environmental interaction record or work order record within the time window.

[0057] If there is a prohibited transition direction in the rule chain, or if the transition direction is allowed but the preceding event is missing, or if the preceding event exists but the event occurs later than the state change time, the behavior causality verification unit determines that the state change is invalid and issues a state rollback instruction to the self-healing execution layer. The state rollback instruction includes the asset identifier, the timestamp of the invalid state change event, and the previous valid state to which it should be restored.

[0058] The spatiotemporal causal engine layer also includes a closed-loop self-verification module. When there are still doubts after the data has undergone triple adversarial verification, the closed-loop self-verification module issues a targeted review instruction to the global perception layer, requiring multiple types of perception terminals in the relevant areas to conduct collaborative re-collection and re-inject the newly collected data into the triple adversarial verification mechanism for secondary verification, until a high-confidence evidence chain is formed that is mutually verified by at least three independent perception methods.

[0059] It should be further explained that the closed-loop self-verification module is deployed inside the spatiotemporal causal engine layer and is connected to the timestamp entropy detection unit, the spatial trajectory game verification unit, and the behavioral causal verification unit, respectively. When any verification unit outputs a judgment result indicating that the data is suspicious, the closed-loop self-verification module first reads the asset identifier, the type of suspicious data, and the time window of the suspicious data. Then, it queries the spatial coordinate area of ​​the asset within a preset range around the time window of the suspicious data from the spatiotemporal map of asset behavior, as well as the list of all sensing terminals currently active in the area. The list of sensing terminals includes fixed RFID readers and their antenna coverage area, the historical operation location of handheld RFID readers, the signal strength coverage area of ​​Beidou positioning base stations, the broadcast range of Bluetooth beacons, the field of view area of ​​industrial vision units, and the mobile devices where microelectromechanical inertial sensors are located.

[0060] In this system, the closed-loop self-verification module and the sensing nodes of the full-domain perception layer adopt standardized command interaction rules. The verification commands and anti-spoofing verification commands issued by the closed-loop self-verification module both adopt a unified structured command format. The command content includes the unique asset identifier, the code of the suspicious point type, the sampling parameter requirements, the data collection time range, and the data return priority. The commands are sent to the target sensing nodes in a broadcast + targeted manner via the MQTT protocol. After receiving the command, the sensing node completes the data collection according to the sampling parameter requirements in the command and encapsulates the collected data in a fixed format of "collection timestamp-node unique identifier-asset identifier-data content-data confidence level". It then prioritizes the return of the data to the closed-loop self-verification module through the original communication link. If the sensing node fails to complete the data collection within the preset time or the collected data is abnormal, it will immediately send an abnormal prompt message back to the closed-loop self-verification module.

[0061] The closed-loop self-verification module generates targeted verification instructions based on the type of suspicious event. If the suspicious event is a time-dimensional anomaly, the verification instruction requires the RFID readers in the corresponding area to repeatedly scan at the highest power during the extended period before and after the suspicious event's time window, and simultaneously requires the industrial vision unit to acquire video streams at the highest frame rate and extract asset image features during this period. If the suspicious event is a spatial-dimensional isolated evidence, the verification instruction requires RFID readers at three or more different locations around the asset to simultaneously perform directional readings, activate the synchronization signals of all Bluetooth beacons in the area, and simultaneously dispatch at least two automated guided vehicles with visual recognition capabilities to move to the vicinity of the suspicious spatial area to acquire asset images from different angles. If the suspicious event is a causal-dimensional invalid state change, the verification instruction requires retrieving the operation logs of all execution terminals associated with the asset before and after the suspicious event's time window, including PDA scanning records, work order approval timestamps, and access control records, and requires the positioning tags worn by relevant operators to perform secondary trajectory acquisition within the designated area.

[0062] After the verification instruction is issued to the full-domain perception layer, each perception terminal performs collaborative re-collection according to the instruction requirements. The newly collected data, along with the collection timestamp, node identifier, and collection parameters, is uniformly transmitted back to the spatiotemporal causal engine layer. The closed-loop self-verification module re-injects this new data into the triple adversarial verification mechanism, and each verification unit performs secondary verification.

[0063] After the secondary verification is completed, the closed-loop self-verification module collects the output results of each verification unit. If the new data can pass the three verifications at the same time and form a mutual verification relationship of at least three independent sensing methods, specifically, the same asset has location information from RFID reading, Beidou positioning and visual recognition within the same time window and the spatial coordinate deviation of the three is less than a preset threshold, or there is trajectory information from RFID reading, Bluetooth ranging and inertial sensing and the time series matching degree of the three exceeds a preset threshold, then the closed-loop self-verification module marks the state of the asset within the suspicious time window as a credible evidence chain and updates it to the spatiotemporal map of asset behavior.

[0064] If some verifications still fail or there are insufficient three verification methods after the second verification, the closed-loop self-verification module will adjust the review instructions again based on the new data, change the combination of sensing terminals or expand the spatial range, and repeatedly issue collaborative re-collection instructions until the conditions for mutual verification by at least three independent sensing methods are met, or if verification still cannot be formed after reaching the preset maximum number of review times, the closed-loop self-verification module will transfer the suspicious asset data to the manual verification queue and highlight the location of the suspicious asset with a special identifier in the twin simulation layer for management personnel to intervene and handle.

[0065] When constructing the location evidence chain, the spatial trajectory game verification unit specifically performs spatiotemporal correlation analysis on the asset location information read by the RFID reader, the continuous trajectory information collected by the Beidou positioning module, and the asset image feature information extracted by the industrial vision unit. If there is a logical contradiction among the three, it is determined to be isolated evidence data and the review process is initiated.

[0066] It should be further explained that when the spatial trajectory game verification unit performs the three-party game verification, it first obtains three types of raw data of the same asset within the same time window from the data fusion core layer. The first type is RFID reader reading data, including the antenna physical coordinates, reading timestamps and signal strength values ​​recorded by the fixed reader when the asset passes through its antenna coverage area. The second type is continuous trajectory data output by the Beidou positioning module, including longitude, latitude and elevation coordinates collected at a fixed frequency and the corresponding timestamp sequence. The third type is asset image feature data extracted by the industrial vision unit, including image frames collected by high-definition network cameras deployed in key areas when triggering shooting, the pixel coordinates of the center point of the asset detection box in the image, and the actual spatial coordinates of the asset obtained by converting the camera calibration parameters.

[0067] The spatial trajectory game verification unit inputs the three types of data into the spatiotemporal alignment module. This module uses the timestamp of the Beidou positioning module as the reference axis, matches the timestamp of the RFID reading event with the reference axis, and finds the positioning point with the closest timestamp as the spatial reference of the reading event. At the same time, it matches the timestamp of the visual recognition event with the reference axis and finds the positioning point with the closest timestamp as the spatial reference of the visual event, thus completing the alignment of the three types of data in the same spatiotemporal coordinate system.

[0068] After alignment, the spatial trajectory game verification unit constructs a chain of evidence for the asset's location within the time window. The chain of evidence consists of multiple evidence nodes connected in chronological order. Each evidence node includes an evidence type identifier, collection time, spatial coordinates, and an initial confidence value. The initial confidence value is assigned according to the following rules: if the evidence comes from the Beidou positioning module and the positioning status is a fixed solution, the confidence value is assigned to the highest level; if it comes from RFID reading and the signal strength exceeds a preset threshold, the confidence value is assigned to a medium-high level; if it comes from visual recognition and the detection box confidence value exceeds a preset threshold, the confidence value is assigned to a medium-low level.

[0069] After the evidence chain is constructed, the spatial trajectory game verification unit performs spatiotemporal consistency verification on adjacent evidence nodes in the chain. The specific verification method is to calculate the time difference and spatial distance between adjacent nodes. If the time difference is less than a preset time threshold and the spatial distance is less than a preset distance threshold, it is determined that the two evidence nodes corroborate each other and a corroboration connection is established in the evidence chain. If the time difference between adjacent nodes meets the threshold but the spatial distance exceeds the distance threshold, the reasons for the excess are further analyzed, including checking whether there is a rapid asset movement operation within the time window, whether there is a blind spot covered by RFID reader antenna, and whether there is visual obstruction causing coordinate conversion error. If the spatial distance deviation still cannot be explained after analysis, it is determined that there is a logical contradiction between the two evidence nodes.

[0070] When three or more evidence nodes in a chain of evidence are mutually corroborating, and the evidence at both ends of all the corroborating links comes from at least two different types of perception methods, the spatial trajectory game verification unit determines that the location information of the asset within that time window is a credible chain of evidence.

[0071] If there is an isolated node in the evidence chain that cannot be corroborated by the nodes before and after it, or if there are multiple unexplained spatial distance deviations between adjacent nodes in the evidence chain, or if the initial confidence values ​​of more than half of the nodes in the evidence chain are lower than the preset level, then the location information of the asset within the time window is determined to be isolated evidence data. At the same time, the suspicious asset identifier, the suspicious time window, and the list of evidence nodes with logical contradictions are output, and the isolated evidence data determination result is sent to the closed-loop self-verification module to trigger the review process.

[0072] When the behavioral causal verification unit performs causal consistency checks, it specifically determines whether the asset status change event is associated with the corresponding work authorization order, access gate reading record and operator trajectory information in the time sequence. If any of the preceding associated events is missing or the order of events does not conform to the preset logic rules, the status change is determined to be invalid.

[0073] It should be further explained that when the behavioral causal verification unit performs causal consistency checks, it first obtains the status change events of the asset that occurred within the suspicious time window from the data fusion layer. The status change events include inbound events, outbound events, relocation events, maintenance events, and scrapping events. Each status change event contains an event type code, an event occurrence timestamp, and the business document number that triggered the event.

[0074] The behavior causal verification unit retrieves the set of preceding events corresponding to the status change of the type from the asset behavior time sequence logic rule base according to the event type code. The set of preceding events includes the work authorization order type that must be associated, the access gate reading record type that must be associated, and the trajectory characteristics of the workers that must be associated.

[0075] The behavioral causality verification unit then retrieves data from the work order management module of the enterprise resource planning system from the middleware server interface, and obtains all relevant work order records for the asset within a preset time period before and after the suspicious time window. The work order records include the warehouse entry slip corresponding to the purchase order, the material requisition slip corresponding to the maintenance work order, the warehouse transfer slip corresponding to the relocation application slip, and the scrapping slip corresponding to the scrapping approval slip. Each work order record includes the work order number, work order type, creation timestamp, approval completion timestamp, and the designated operator identifier.

[0076] Simultaneously, the behavior causality verification unit extracts the access gate reading records of the asset within a preset time period before and after the suspicious time window from the historical data of the full-domain perception layer. The reading records include the timestamp and antenna position recorded when the fixed RFID reader at the warehouse entrance and exit reads the asset tag, the timestamp and lane number recorded when the buried RFID reader in the loading and unloading area reads the asset tag, and the Beidou positioning trajectory matching record triggered when the asset passes through the vehicle gate.

[0077] The behavioral causal verification unit also extracts the operator trajectory information associated with the asset within a preset time period before and after the suspicious time window from the spatiotemporal graph of the asset behavior of the spatiotemporal causal engine layer. The operator trajectory information includes the continuous movement trajectory of the warehouse manager wearing a positioning tag within the corresponding time period, the location record of the handheld PDA device, and the task path record of the automated guided vehicle.

[0078] After data extraction is completed, the behavioral causal verification unit performs a causal order matching test. Specifically, the test method is to use the timestamp of the status change event as a benchmark and trace back within a preset time period to see if there are any work order records, identification records, and personnel trajectories that meet the rules set of the preceding events. If a work order record exists but the approval completion timestamp of the work order is later than the timestamp of the status change event, it is determined that the causal order is reversed. If a work order record exists and the approval completion timestamp is earlier than the timestamp of the status change event, but the corresponding access gate identification record is missing or the timestamp of the identification record is later than the timestamp of the status change event, it is determined that the execution evidence is missing. If a work order record and identification record exist but the operator's trajectory shows that the operator was located in another area and had no connecting movement path when the status change event occurred, it is determined that the personnel evidence is contradictory.

[0079] When any of the above determinations occurs, the behavior causality verification unit outputs a determination result that the state change event is invalid, and packages the invalid determination along with all associated work order records, reading records and personnel trajectory data and sends them to the self-healing execution layer. The self-healing execution layer generates a state rollback instruction based on the invalid determination result. The state rollback instruction includes the asset identifier, the timestamp of the invalid state change event and the previous valid state obtained from the asset behavior spatiotemporal graph. At the same time, the state of the five-dimensional asset twin in the twin inference layer is rolled back to the previous valid state, and the asset is highlighted in a specific color in the twin model to prompt the management personnel to check.

[0080] If all preceding events exist and their temporal order conforms to the rules in the causal sequence matching test, the behavior causal verification unit outputs the determination result that the state change event is valid, and updates the event chain of the asset node in the asset behavior spatiotemporal graph as a new historical record.

[0081] The spatiotemporal causal engine layer also includes an anti-spoofing verification module. After the timestamp entropy value detection determines that there is an abnormal data collection, the anti-spoofing verification module sends a high-frequency sampling command to the global perception layer, requiring the relevant perception terminals to encrypt the collected data within a preset time period and re-inject the encrypted collected data into the triple adversarial verification mechanism for verification.

[0082] It should be further explained that the anti-spoofing verification module is deployed inside the spatiotemporal causal engine layer and maintains real-time data interaction with the timestamp entropy detection unit. When the timestamp entropy detection unit determines that there is a time dimension anomaly in a certain asset data stream and outputs an anti-spoofing verification trigger signal, the anti-spoofing verification module first reads the asset identifier, suspicious time window, and anomaly type code contained in the trigger signal. The anomaly type code specifically includes time interval entropy value exceeding limit code, continuous window entropy value trend anomaly code, and regularity deviation code.

[0083] The anti-spoofing verification module matches the corresponding high-frequency sampling parameters from the preset sampling strategy library based on the anomaly type code. The sampling strategy library is constructed using a mapping table structure that maps anomaly type codes to high-frequency sampling parameters. The fields of the mapping table include anomaly type code, anomaly description, RFID reader / writer sampling parameters, Beidou positioning module sampling parameters, industrial vision unit sampling parameters, Bluetooth beacon working parameters, and sampling time range. The anomaly type codes include three categories: time interval entropy value exceeding limit code, continuous window entropy value trend anomaly code, and regular deviation code. The sampling strategy library is deployed in the asset data platform and interacts with the anti-spoofing verification module in the spatiotemporal causal engine layer in real time. After receiving the anomaly trigger signal from the timestamp entropy value detection unit, the anti-spoofing verification module extracts the anomaly type code from the signal and retrieves the corresponding high-frequency sampling parameters from the sampling strategy library through precise matching. If the anomaly type code does not match the corresponding parameter in the library, the preset default high-frequency sampling parameters will be automatically called. The sampling strategy library supports dynamic adjustment and supplementation of parameters according to the updates of sensing devices and changes in application scenarios. If the anomaly type code is a time interval entropy value exceeding the limit code, the matching high-frequency sampling parameters are as follows: within a time period extending two window lengths before and after the suspicious time window, increase the RFID reader's transmission power to 1.5 times the normal power and shorten the scanning cycle to one-third of the normal cycle, increase the sampling frequency of the Beidou positioning module to four times the normal frequency, and increase the capture frame rate of the industrial vision unit to five times the normal frame rate. If the anomaly type code is a continuous window entropy value trend anomaly code, the matching high-frequency sampling parameters are as follows: within the subsequent three consecutive time windows, decrease the RFID reader's scanning cycle window by window, increase the Beidou positioning module's sampling frequency window by window, and increase the industrial vision unit's image resolution window by window. If the anomaly type code is a regular deviation code, the matching high-frequency sampling parameters are as follows: within a time period extending three window lengths before and after the suspicious time window, activate all Bluetooth beacons within a 50-meter radius around the asset to synchronously broadcast positioning signals, and simultaneously dispatch all RFID-enabled automated guided vehicles in the area to perform a carpet scan by moving back and forth along a fixed path during this time period.

[0084] The anti-spoofing verification module encapsulates the matched high-frequency sampling parameters into anti-spoofing verification instructions, which are then sent to all sensing terminals associated with the asset in the global perception layer through the data fusion layer. The instructions clearly indicate the sampling start and end time, sampling frequency or period, device power or sensitivity settings, and data feedback priority.

[0085] After receiving the instruction, each sensing terminal performs encrypted data collection according to the parameter requirements within the specified time period. The collected data is separately marked as anti-spoofing verification data stream, and is uploaded to the spatiotemporal causal engine layer with the batch identifier of this verification attached.

[0086] The anti-spoofing verification module re-injects these encrypted collected data into the triple adversarial verification mechanism, while temporarily reducing the confidence weight of the asset in the regular data stream, and using the verification result of the encrypted collected data as the final judgment basis for the asset within this time window.

[0087] If the encrypted data is verified by triple adversarial verification and forms mutual verification with other verifiable data in the regular data stream through at least two independent sensing methods, the anti-spoofing verification module marks the data within the time window as trustworthy and updates it to the asset behavior spatiotemporal graph. At the same time, it marks the original data in the corresponding time window in the regular data stream as contaminated data and removes it from the graph.

[0088] If the encrypted data still fails to pass verification, the anti-spoofing verification module will adjust the high-frequency sampling parameters again, expand the sampling time and space range or add sensing terminal types, and repeatedly issue anti-spoofing verification instructions until credible evidence is obtained or the preset maximum number of verifications is reached. If it still fails to pass after reaching the maximum number of verifications, the data of the asset within the suspicious time window will be marked as pending manual verification and the management personnel will be prompted to intervene in the twin simulation layer.

[0089] The twin simulation layer incorporates a behavior prediction model and a strategy simulator. The behavior prediction model, based on the real-time state evidence chain and historical change records associated with the five-dimensional asset twin, predicts the state evolution trend of the asset in the future and identifies potential abnormal risks. After receiving the inventory and verification instruction, the strategy simulator automatically generates multiple execution plans based on the current spatiotemporal distribution of the asset and the coverage of the sensing terminal. It then simulates the execution process of each plan in parallel in the five-dimensional asset twin, predicts the execution time and potential blind spots, selects the optimal plan, decomposes the task, and sends it to the execution terminal.

[0090] It should be further explained that the behavior prediction model built into the twin inference layer is deployed on the back end of the five-dimensional asset twin and keeps real-time data synchronized with the spatiotemporal causal engine layer. The behavior prediction model takes the five-dimensional twin of each asset as the input unit. The five-dimensional twin dynamically binds the real-time status evidence chain of the asset after passing the triple adversarial verification, the historical change records stored in the asset behavior spatiotemporal graph, the technical parameter files synchronized in the enterprise resource planning system, and the historical maintenance data recorded in the maintenance work order module.

[0091] The behavior prediction model employs a time-series prediction algorithm to independently construct a state evolution trend line for each asset. The input features of the time-series prediction algorithm include the asset's location movement frequency, RFID read counts, vibration sensor activation duration, and the number of associated work orders within multiple past time windows. The behavior prediction model is constructed using the LSTM time-series prediction algorithm. The model's network structure includes an input layer, hidden layers, and an output layer. The feature dimensions of the input layer are differentiated according to the asset type. The core input features are the asset's location movement frequency, RFID read counts, vibration sensor activation duration, number of associated work orders, and changes in operating parameters within multiple past time windows. The hidden layer has three layers, and the number of neurons in each layer is adjusted according to the complexity of the asset type. The outgoing layer represents the possible state types and corresponding predicted probabilities that assets may enter within a predetermined time window. The model is trained using the company's asset historical operation data, state change records, and environmental interaction data from the past three years as the training set. After data cleaning, missing value imputation, and normalization, the training set data is divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used to complete the basic training of the model, the validation set is used to fine-tune the model's hyperparameters, and the test set is used to verify the model's prediction accuracy. The model is deployed on the backend of the five-dimensional asset twin in the twin inference layer, maintaining real-time data synchronization with the spatiotemporal causal engine layer. It receives high-confidence perception data of assets in real time and inputs it into the model to complete the inference of the future state evolution trend of assets and the identification of potential abnormal risks.

[0092] The output is the possible state types that the asset may enter within a preset number of time windows in the future, as well as the probability value corresponding to each state type. When the probability value exceeds a preset threshold, it is judged as a potential abnormal risk. Potential abnormal risks include idle risk, displacement deviation risk, maintenance overdue risk, and unauthorized movement risk.

[0093] The behavioral prediction model attaches potential abnormal risks identified in the five-dimensional asset twin as risk labels, and distinguishes risk levels with different color levels in the global view of the twin inference layer.

[0094] The strategy simulator and behavior prediction model built into the twin simulation layer are deployed in parallel. After receiving the inventory and verification instruction, the strategy simulator first obtains the spatiotemporal distribution data of all current assets from the data fusion layer. The spatiotemporal distribution data includes the latest location coordinates of each asset, the type of the area it is located in, and the type and number of active sensing terminals in the area. At the same time, the strategy simulator obtains the execution records of historical inventory tasks from the asset behavior spatiotemporal map. The historical inventory records include the path planning, time consumption statistics, and distribution of missed assets of past inventory tasks.

[0095] The strategy simulator automatically generates multiple inventory execution plans based on the current spatiotemporal distribution of assets and the coverage of sensing terminals. Each plan includes a zoning strategy, a path planning strategy, and a sensing terminal scheduling strategy. The zoning strategy divides all assets into multiple inventory sub-areas according to spatial proximity. The path planning strategy plans the movement route of the execution terminals in each sub-area according to the shortest path principle. The sensing terminal scheduling strategy determines which fixed readers to activate, which automated guided vehicles to schedule, and which handheld terminals to coordinate during the inventory process.

[0096] The strategy simulator loads multiple generated schemes into a five-dimensional asset twin for parallel simulation and deduction. During the deduction process, each scheme is virtually executed in the twin model. The simulation execution terminal moves according to the planned path, the sensing terminal collects data according to the scheduling strategy, and the asset status is updated as expected. The strategy simulator records the expected time, the expected number of assets traversed, the expected number of blind spot assets, and the expected resource utilization rate of the sensing terminal for each scheme during the simulation execution process.

[0097] After the simulation is completed, the strategy simulator will comprehensively rank the simulation results of each scheme. The ranking criteria include the shortest expected time, the fewest blind spot assets, and the balanced resource utilization. The scheme with the best comprehensive ranking will be selected as the actual execution scheme.

[0098] The strategy simulator decomposes the selected optimal solution into specific task assignments. These tasks include the movement path sequence of each execution terminal, the activation period of each fixed reader, the scanning area coordinates of each automated guided vehicle, and the list of assets that each handheld terminal needs to check. The decomposed tasks are then distributed to the corresponding execution terminals through the self-healing execution layer.

[0099] The twin simulation layer also includes an asset health assessment unit. The asset health assessment unit calculates the asset health index by combining asset operation data, idle time and environmental parameters, and displays the risk distribution in a visual form in the five-dimensional asset twin. When the idle rate of a certain type of asset exceeds the set threshold, the strategy simulator generates an allocation optimization suggestion based on historical allocation data and current demand forecasts, and outputs the decision basis after simulating the implementation effect of the scheme in the five-dimensional asset twin.

[0100] It should be further explained that the asset health assessment unit built into the twin simulation layer maintains data interaction with the behavior prediction model. The asset health assessment unit regularly obtains real-time operating data of each asset from the data fusion layer. The operating data includes the start and stop frequency collected by the vibration sensor, the load current curve collected by the current transformer, the ambient temperature change collected by the temperature sensor, and the displacement frequency collected by the Beidou positioning module. At the same time, it extracts the idle time record of each asset from the spatiotemporal map of asset behavior. The idle time refers to the cumulative time that the asset has not moved, been RFID-read, or been associated with any work order within a continuous time window. At the same time, it synchronizes the environmental parameters of the area where each asset is located from the enterprise resource planning system. The environmental parameters include warehouse temperature and humidity, weather conditions in the open storage yard, and dust concentration in the production workshop.

[0101] The asset health assessment unit normalizes the above three types of data and inputs them into the health index calculation model. The health index calculation model outputs the health index score of each asset using a weighted summation method. The weight coefficients are set according to the asset type. The vibration frequency weight of rotating equipment is higher than the displacement frequency weight, the ambient temperature weight of precision instruments is higher than the start-stop frequency weight, and the idle time weight of standby equipment is higher than the load current weight.

[0102] In the health index calculation model of the asset health assessment unit, the collected asset operation data, idle time, and environmental parameters are first normalized. The maximum-minimum method is used to convert the original data of different dimensions and numerical ranges to a unified numerical range of 0-1, eliminating calculation biases caused by differences in data dimensions. Then, weight coefficients for each indicator are set according to the asset type. Specifically, the vibration frequency of rotating equipment has a weight of 0.3, load current has a weight of 0.25, idle time has a weight of 0.15, and environmental parameters such as ambient temperature, humidity, and dust concentration have a weight of 0.3. For precision instruments, ambient temperature has a weight of 0.35, and idle time has a weight of 0. 0.25, start-stop frequency weight is 0.2, displacement frequency weight is 0.2, idle time of standby equipment weight is 0.4, environmental parameters weight is 0.25, load current weight is 0.2, vibration frequency weight is 0.15. The weight coefficients for other types of assets are set according to the functional attributes and usage requirements, referring to the above three types of assets. Finally, the normalized index values ​​are multiplied by their corresponding weight coefficients by a weighted summation method, and the summation result is magnified by 100 times to obtain an asset health index score of 0-100. The higher the score, the better the asset health. When the score is below 60, the asset health is judged to be poor, and a maintenance prompt needs to be triggered.

[0103] After the health index score is generated, it is displayed in a visual form in the five-dimensional asset twin. The specific display method includes assigning a color level mapping to each asset node in the three-dimensional model. The health index score changes from high to low, corresponding to a color level that changes from cool to warm. At the same time, an asset health heat map is generated in the global view. The heat map is overlaid on the warehouse or factory floor plan, and the color depth of the blocks reflects the average health index of the assets in that area.

[0104] When the idle rate of a certain type of asset exceeds a preset threshold, the idle rate refers to the proportion of the number of assets in that type of asset that have been idle for more than a specified number of days to the total number of assets in that type. The asset health assessment unit sends an idle warning signal to the strategy simulator. The warning signal carries the type code of the asset, the list of idle assets, and the current distribution area of ​​the idle assets.

[0105] After receiving an idle warning signal, the strategy simulator retrieves the historical transfer records of this type of asset from the spatiotemporal map of asset behavior. The historical transfer records include the region from which the asset was moved, the region from which it was moved, the transfer time, and the change in usage frequency after the transfer in past transfer tasks. At the same time, it obtains the current production plan and maintenance work order demand forecast from the enterprise resource planning system. The demand forecast includes the expected demand for this type of asset in each region within a preset time period, the demand time window, and the demand priority.

[0106] The strategy simulator generates multiple allocation optimization suggestions based on the distribution of idle assets and demand forecasts. Each suggestion includes a list of assets to be allocated, the target relocation area, the allocation schedule, and the expected release of storage space. The list of assets to be allocated is selected from the list of idle assets. The selection criteria include that the asset health index is not lower than the minimum value required by the relocation area, the asset type matches the demand forecast, and the transportation distance between the current location of the asset and the target area.

[0107] The strategy simulator loads each allocation optimization suggestion into a five-dimensional asset twin for simulation. During the simulation, the twin model virtually executes the allocation operation according to the scheme content, simulating the changes in shelf vacancy after the asset is removed from its original location, the changes in storage space occupancy rate after the asset is moved to the target area, and the changes in asset idle rate in each area after the allocation is completed. The strategy simulator records the asset utilization rate improvement, the total amount of warehouse space released, and the estimated transportation cost after each scheme is simulated and executed.

[0108] After the simulation is completed, the strategy simulator compares and analyzes the simulation results of each scheme. The scheme with the highest increase in asset utilization, the most warehouse space release, or the lowest transportation cost is selected as the recommended scheme. At the same time, a comparison chart of the implementation effect of the scheme is generated. The chart shows the changes in asset distribution balance, regional idle rate difference, and demand satisfaction rate between the current state and after the implementation of the scheme. Finally, the recommended scheme and the comparison chart are output to the monitoring screen and management terminal for management personnel approval as the basis for decision-making.

[0109] The global perception layer also includes industrial vision units and microelectromechanical inertial sensors. The industrial vision units are deployed in key areas and extract asset appearance features and position offset information in real time through edge computing, generate visual feature vectors, and form multimodal associations with the perception data. Microelectromechanical inertial sensors are attached to mobile assets to collect instantaneous acceleration and attitude change data. A multimodal data fusion engine performs spatiotemporal alignment and feature fusion of visual feature vectors, inertial sensing data, and RFID positioning data.

[0110] The multimodal data fusion engine integrates visual feature vectors, inertial sensing data, and RFID positioning data through two core steps: spatiotemporal alignment and feature fusion. Spatiotemporal alignment uses the global clock of the global perception layer as a reference, precisely matching the timestamps of data collected by industrial vision units, microelectromechanical inertial sensors, and RFID readers at the millisecond level. This ensures that data about the same asset collected by different sensing methods remains consistent in the time dimension. At the same time, the spatial coordinates of all sensing data are uniformly converted to the national 2000 geodetic coordinate system to eliminate spatial coordinate deviations between different positioning methods and devices. Feature fusion uses a feature stitching method, combining the asset appearance feature vector extracted by the industrial vision unit, the asset motion feature vector collected by the microelectromechanical inertial sensor, and the asset identity and location feature vector collected by the RFID reader according to a fixed dimension to form an initial multimodal feature vector. Then, a feature filtering algorithm removes redundant and invalid features from the initial feature vector, retaining the core features that can effectively represent the asset status, forming the final fused feature data. This data will be associated with other sensing data of the asset for subsequent spatiotemporal causal verification and asset status analysis.

[0111] Dead reckoning is used in this system to supplement the positioning of mobile assets. When the BeiDou positioning module experiences signal blockage or positioning failure, the system uses instantaneous acceleration and angular velocity data collected by the microelectromechanical inertial sensors attached to the mobile asset. Combined with the asset's latest valid BeiDou positioning coordinates as the initial position, the system continuously calculates the asset's direction of motion, speed of motion, and duration of motion to obtain the asset's relative displacement and real-time position during the period of positioning signal failure. Finally, the calculated position data is fused with Bluetooth beacon positioning data to achieve continuous positioning and tracking of the asset. This method is suitable for scenarios with poor BeiDou positioning signals, such as warehouse shelves blocking the view or workshop equipment obstructing the view.

[0112] It should be further explained that the industrial vision unit deployed in the full-domain perception layer includes multiple high-definition network cameras and edge computing nodes deployed on the field side. The high-definition network cameras are installed at key locations such as warehouse entrances and exits, intersections of shelf aisles, loading and unloading areas, and open-air storage yards. The field of view of each camera covers adjacent shelves or stacking areas. The cameras are equipped with image sensors and video encoding chips to collect real-time video streams at a fixed frame rate and push the video streams to edge computing nodes.

[0113] The edge computing node has a built-in visual feature extraction algorithm. The visual feature extraction algorithm uses an object detection network to identify and locate assets in each frame of the image, and outputs the detection box coordinates, confidence score and appearance feature vector of each asset. The appearance feature vector is formed by concatenating color histogram, texture features and contour descriptor. At the same time, the edge computing node converts the pixel coordinates of the detection box center point into the actual spatial coordinates of the asset according to the camera calibration parameters, and calculates the position offset and offset direction of the asset between the current frame and the previous frame.

[0114] Microelectromechanical inertial sensors are attached to forklifts, automated guided vehicles, and mobile equipment. The inertial sensors have built-in three-axis accelerometers and three-axis gyroscopes to collect instantaneous acceleration and angular velocity data of the assets at a fixed sampling frequency. The inertial sensors send the collected data to a nearby IoT gateway via Bluetooth or ZigBee protocol. The IoT gateway timestamps the data and then uploads it to the data fusion layer.

[0115] After receiving the visual feature vectors and position offset information output by the industrial vision unit, as well as the acceleration and angular velocity data output by the inertial sensor, the multimodal data fusion engine first performs spatiotemporal alignment processing. The spatiotemporal alignment processing uses the global clock of the global perception layer as a reference to match the timestamps of the visual data with the timestamps of the inertial data, and finds the visual frame and inertial sampling point with the smallest time difference as the multimodal observation pair at the same moment. At the same time, the spatial coordinates of the asset converted from the visual data are aligned with the relative displacements calculated from the dead reckoning of the inertial data to eliminate the deviations introduced by different coordinate systems.

[0116] After spatiotemporal alignment is completed, the multimodal data fusion engine performs feature fusion processing. The feature fusion processing combines visual feature vectors, position offset information, and acceleration amplitude and attitude angle changes from inertial sensing data into multimodal feature vectors. The visual part of the multimodal feature vector is used for asset identification and appearance status judgment, the position offset information is used to determine whether the asset has shifted or vibrated, and the inertial sensing data is used to determine the asset's motion mode and attitude changes.

[0117] The multimodal data fusion engine associates the fused multimodal feature vector with the RFID tag identification data corresponding to the asset. The RFID tag identification data includes the tag's unique identifier, the reading timestamp, and the reader antenna position coordinates. The association method is to bind RFID reading events falling within the same time window with the multimodal feature vector, forming a complete perception record of the asset within that time window.

[0118] After the complete perception record is generated, the multimodal data fusion engine sends it to the spatiotemporal causal engine layer. In the subsequent triple adversarial verification process, the spatiotemporal causal engine layer uses the multimodal feature vector as one of the independent perception methods to participate in spatial trajectory game verification and behavioral causal verification. When there is a lack of RFID reading data or signal obstruction of the positioning module, the visual position information and inertial trajectory data in the multimodal feature vector can serve as supplementary evidence to maintain continuous tracking of the asset status. When there is a deviation in the asset position output by different perception methods, the visual feature matching degree and inertial attitude consistency in the multimodal feature vector can be used to determine whether the deviation is caused by sensor error or actual asset movement, thereby improving the accuracy and robustness of the verification results.

[0119] A self-verifying and trustworthy system for asset data has been constructed: through a triple adversarial verification mechanism built into the spatiotemporal causal engine layer, the system performs timestamp entropy value detection, spatial trajectory game verification, and behavioral causal verification on each piece of asset data, effectively filtering abnormal noise and logically contradictory data. For questionable data, a closed-loop self-verification process is initiated, requiring multiple types of sensing terminals to collaboratively re-collect data until a high-confidence evidence chain mutually corroborating at least three independent sensing methods is formed. This mechanism fundamentally solves the problems of unreliable data sources and long-term discrepancies between accounts and physical assets in traditional asset management, enabling the asset status driven by the digital twin model to truly have the ability to update synchronously with physical reality, providing an accurate and reliable data foundation for all subsequent management and decision-making.

[0120] The comprehensive perception layer integrates multiple technologies such as RFID, BeiDou positioning, Bluetooth beacons, vibration sensing, industrial vision, and microelectromechanical inertial sensing to break down the physical boundaries of the warehouse and achieve continuous tracking and status awareness of assets in various scenarios, including production workshops, open-air storage yards, and office areas. The twin simulation layer, based on a five-dimensional asset twin, performs behavior prediction and strategy simulation. It automatically generates optimal execution plans during inventory checks and simulates optimized allocation paths when assets are idle. Through a self-healing execution layer, decision-making instructions are transformed into physical control actions, with execution results transmitted back in real time to form a closed loop. This design upgrades asset management from passive recording to proactive simulation and self-organizing execution, improving asset utilization and management transparency, providing precise decision support for enterprise operations, and achieving comprehensive coverage and intelligent closed-loop asset management.

[0121] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital twin-based equipment and materials inventory and asset verification management system, characterized in that, include: The global perception layer is deployed in the areas where various assets of the enterprise are located to collect real-time data of the assets. It includes passive UHF RFID tags attached to the assets, Beidou positioning modules and Bluetooth beacons for indoor and outdoor positioning, and vibration sensors and current transformers for collecting asset operating status. The data fusion layer is communicatively connected to the global perception layer. It includes an asset data platform, which has a built-in multimodal data fusion engine and a dynamic verification module. The dynamic verification module performs multiple rounds of progressive verification and review on the fused data. The spatiotemporal causal engine layer communicates with the data fusion core layer. It has a built-in triple adversarial verification mechanism for performing timestamp entropy detection, spatial trajectory game verification, and behavioral causal verification on asset data streams, and triggers a closed-loop review process for data with doubts. The twin inference layer is communicatively connected to the spatiotemporal causal engine layer. It is loaded with a five-dimensional asset twin that precisely corresponds to the enterprise's physical space. The five-dimensional asset twin adds a time dimension and a state dimension to the three-dimensional spatial model to achieve real-time synchronization of asset status, behavior prediction, and strategy inference. The self-healing execution layer is communicatively connected to the twin inference layer. It includes an intelligent scheduling gateway and various execution terminals, which are used to convert the decision instructions of the twin inference layer into physical control actions and to provide feedback verification on the execution results. The real-time data collected by the global perception layer is fused by the data fusion layer and then enters the spatiotemporal causal engine layer for multiple verifications and reviews. The high-confidence data after verification drives the five-dimensional asset twin in the twin inference layer to keep synchronized with the physical asset. The self-healing execution layer receives the instructions from the twin inference layer and executes on-site operations. The data collected during the execution process is sent back to the data fusion layer to form a closed loop.

2. The equipment and material inventory and asset verification management system combining digital twins according to claim 1, characterized in that: The spatiotemporal causal engine layer constructs an asset behavior spatiotemporal graph, where each asset node is associated with its historical trajectory sequence, environmental interaction records, and state change events. The triple adversarial verification mechanism performs self-consistency verification on the real-time data stream based on the asset behavior spatiotemporal graph.

3. The equipment and material inventory and asset verification management system combining digital twins according to claim 2, characterized in that: The triple adversarial verification mechanism includes: The first layer is the timestamp entropy detection unit, which is used to calculate the timestamp distribution entropy value of each asset data stream and compare it with the historical normal mode entropy value of the asset. When the timestamp distribution entropy value shows abnormal fluctuations or the time interval shows regular deviations, it is determined that there is a data collection anomaly and the anti-spoofing verification process is triggered. The second layer is the spatial trajectory game verification unit, which is used to conduct a three-way game verification of the asset's location trajectory with RFID reading records and visual recognition records to build a location evidence chain. It requires that at least two sensing methods generate mutually corroborating location evidence for the same asset within the same spatiotemporal window. If the asset is only captured by a single sensing method and there is a logical conflict with other sensing data, the data is determined to be isolated evidence and a review task is generated. The third layer is the behavior causality verification unit, which calls the asset behavior time sequence logic rule library to perform causal consistency checks on asset status changes. If the status change event lacks a preceding causal relationship or the causal order is logically reversed, the status change is determined to be invalid and the self-healing mechanism is triggered.

4. The equipment and material inventory and asset verification management system combining digital twins according to claim 3, characterized in that: The spatiotemporal causal engine layer also includes a closed-loop self-verification module. When there are still doubts after the data has undergone triple adversarial verification, the closed-loop self-verification module issues a targeted review instruction to the global perception layer, requiring multiple types of perception terminals in the relevant area to re-collect data collaboratively and re-inject the newly collected data into the triple adversarial verification mechanism for secondary verification, until a high-confidence evidence chain is formed that is mutually verified by at least three independent perception methods.

5. The equipment and material inventory and asset verification management system combining digital twins according to claim 4, characterized in that: When constructing the location evidence chain, the spatial trajectory game verification unit specifically performs spatiotemporal correlation analysis on the asset location information read by the RFID reader, the continuous trajectory information collected by the Beidou positioning module, and the asset image feature information extracted by the industrial vision unit. If there is a logical contradiction among the three, it is determined to be isolated evidence data and the review process is initiated.

6. The equipment and material inventory and asset verification management system combining digital twins according to claim 4, characterized in that: When the behavior causal verification unit performs causal consistency checks, it specifically determines whether the asset status change event is associated with the corresponding work authorization order, access gate reading record and operator trajectory information in the time sequence. If any of the preceding associated events is missing or the order of events does not conform to the preset logic rules, the status change is determined to be invalid.

7. The equipment and material inventory and asset verification management system combining digital twins according to claim 2, characterized in that: The spatiotemporal causal engine layer also includes an anti-spoofing verification module. After the timestamp entropy value detection determines that there is a data collection anomaly, the anti-spoofing verification module sends a high-frequency sampling command to the global perception layer, requiring the relevant perception terminals to encrypt the collected data within a preset time period and re-inject the encrypted collected data into the triple adversarial verification mechanism for verification.

8. The equipment and material inventory and asset verification management system combining digital twins according to claim 1, characterized in that: The twin inference layer has a built-in behavior prediction model and strategy simulator. The behavior prediction model is based on the real-time state evidence chain and historical change records associated with the five-dimensional asset twin to infer the state evolution trend of the asset in the future period and identify potential abnormal risks. After receiving the inventory check instruction, the strategy simulator automatically generates multiple execution plans based on the current spatiotemporal distribution of assets and the coverage of sensing terminals. It then simulates the execution process of each plan in parallel in the five-dimensional asset twin, predicts the execution time and potential blind spots, selects the optimal plan, decomposes the task, and sends it to the execution terminal.

9. The equipment and material inventory and asset verification management system combining digital twins according to claim 8, characterized in that: The twin simulation layer also includes an asset health assessment unit. This unit calculates an asset health index by combining asset operation data, idle time, and environmental parameters, and displays the risk distribution in a visual form within the five-dimensional asset twin. When the idle rate of a certain type of asset exceeds a set threshold, the strategy simulator generates an allocation optimization suggestion based on historical allocation data and current demand forecasts. After simulating the implementation effect of the suggestion in the five-dimensional asset twin, it outputs the decision basis.

10. The equipment and material inventory and asset verification management system combining digital twins according to claim 1, characterized in that: The global perception layer also includes industrial vision units and microelectromechanical inertial sensors. The industrial vision units are deployed in key areas and extract asset appearance features and position offset information in real time through edge computing, generating visual feature vectors and forming multimodal associations with the perception data. The microelectromechanical inertial sensors are attached to the moving assets to collect the assets' instantaneous acceleration and attitude change data. The multimodal data fusion engine performs spatiotemporal alignment and feature fusion of visual feature vectors, inertial sensing data, and RFID positioning data.

Images