Supply chain risk early warning evaluation method and system based on digital twin platform

Abstract

The present application relates to the field of wisdom logistics and digital twin, and provides a supply chain risk early warning evaluation method based on a digital twin platform, comprising: acquiring multi-modal sensor data to construct a dynamic graph sequence to represent physical entities and interaction relationships; in response to a target entity identification signal loss, calling a dynamic graph sequence input track extrapolation model of a time window before the loss to generate a candidate state track set, the model being a generative model introducing physical prior constraints to correct the inferred direction; acquiring real multi-modal observation data corresponding to space-time and cross-verification with predicted sensor observation features, calculating evidence support degree, and determining a physical law violation index based on the degree of continuous deviation; calculating an abnormality measurement index according to the evidence support degree, and triggering an early warning when a preset early warning condition is met. The present application realizes physical law driven continuous state reasoning without oscillation, improves abnormality detection from signal loss to physical law violation, and greatly enhances the hidden risk discovery capability.

Description

Technical Field

[0001] This application relates to the fields of smart logistics and digital twins, and in particular to supply chain risk early warning and assessment methods and systems based on digital twin platforms. Background Technology

[0002] Currently, in the fields of smart logistics and digital twin supply chains, a combined architecture of an IoT sensing layer and a digital twin platform mapping layer is typically used to achieve real-time monitoring and risk warning of goods in transit or in storage. In the physical world, RFID readers, cameras, and other sensing devices collect cargo identification and location information, which is then uploaded to a cloud-based or edge-based digital twin platform. Within this platform, a corresponding digital twin is created for each physical entity, allowing for state updates based on an event-driven architecture. Risk warnings are generated when state transitions trigger preset rules. Object state management is generally based on a finite state machine, defining state transition conditions such as "entering an area," "leaving an area," and "timeout." When an RFID tag is read, an object "creation" event is triggered; when the same tag is not read within a fixed time period, a "destruction" or "departure" event is triggered. To reduce interference from occasional missed reads, sliding time window statistics or simple delayed cancellation strategies are often used.

[0003] However, in complex real-world physical environments such as ports and large distribution centers, due to the multipath reflection and obstruction effects of dense shelving or metal containers, RFID radio frequency signals or visual recognition results often exhibit repeated jumps within a short period of time: "successful identification—complete loss—successful identification again." Faced with this instability of the sensing signal, the event-driven architecture of existing technologies treats each sensor report as an isolated, decisive atomic event—"if it is read, it exists; if it is not read, it disappears"—completely abandoning the continuous context of the object's own movement and existence. Once the identification signal is briefly lost due to environmental interference, the system cannot continuously infer the true motion state of the object based on physical laws, and can only default to the simplest static assumption: "the object disappears at its last known location." This hard-decision approach, lacking prior modeling capabilities of the physical world, leads the system to make erroneous state inferences that contradict the true continuous movement of objects in the physical world. Summary of the Invention

[0004] To address the problem of errors in object state reasoning caused by the lack of prior physical modeling capabilities in existing technologies, this application proposes a supply chain risk early warning and assessment method based on a digital twin platform. This method deeply integrates generative models with multimodal cross-validation mechanisms to construct a differentiable closed-loop reasoning engine driven by physical priors, thereby achieving oscillation-free continuous state reasoning and accurate early warning of hidden risks.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, the supply chain risk early warning and assessment method based on a digital twin platform includes: acquiring multimodal sensor data and constructing a dynamic graph sequence, which is used to characterize the physical entities and their interactions within the detected supply chain scenario. In response to the loss of the target entity's identification signal, the dynamic graph sequence corresponding to the time window before the loss is used as a condition and input into a trajectory extrapolation model to generate a candidate state trajectory set for the target entity. The trajectory extrapolation model is configured as a generative model that generates future state sequences based on historical states, and physical prior constraints are introduced during the generation process to correct the inference direction. Real multimodal observation data corresponding to the candidate state trajectory set in time and space are acquired and cross-validated with the predicted sensor observation features within the candidate state trajectory set. Evidence support is calculated using an anomaly metric function, and a physical law violation index is determined based on the degree of persistent deviation from the evidence support. Based on the evidence support, an anomaly metric index for the target entity is calculated, and a supply chain risk early warning is triggered when the anomaly metric index meets preset early warning conditions.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the trajectory extrapolation model is a conditional diffusion model, including a backbone denoising network and a parallel differentiable physics simulator. The process of generating a candidate state trajectory set for the target entity specifically includes: using the dynamic graph sequence corresponding to the time window before loss as a condition, performing a reverse diffusion process starting from random noise; within each denoising time step, the backbone denoising network predicts the initial denoising direction based on the current noise state and conditions, and simultaneously outputs the predicted original physical state corresponding to the current noise state, which includes the target entity's position, velocity, and acceleration. The differentiable physics simulator receives the predicted original physical state, calculates the corresponding physically reasonable state according to Newton's laws of motion, and generates a physical gradient correction term to correct the initial denoising direction. After adjusting the initial denoising direction with the physical gradient correction term, the current noise state is updated, and the process proceeds to the next denoising time step, iterating until at least one future state trajectory that meets the physical constraints is reconstructed, forming a candidate state trajectory set.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the process of calculating the physical gradient correction term using a differentiable physics simulator specifically includes: resolving the current position, velocity, and acceleration of the target entity from the predicted original physical state, and extracting the interaction force vectors of neighboring entities on the target entity from the dynamic graph sequence. The preset mass, interaction force vectors, and environmental force vectors of the target entity are input into a built-in rigid body dynamics solver, and the corrected acceleration is calculated according to Newton's second law. A physically reasonable state satisfying mechanical constraints is obtained through numerical integration. The element-wise physical residual between the physically reasonable state and the predicted original physical state is calculated, and the physical residual is mapped to the noise prediction space via an encoding network to generate the physical gradient correction term.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the process of calculating the evidence support through the anomaly metric function specifically includes: extracting predicted sensor observation features for future time points from the predicted state sequences of each candidate state trajectory in the candidate state trajectory set. These predicted sensor observation features include at least predicted spatial location and predicted visual feature vectors. Using the future time point as an index, performing a spatiotemporal attention query from the real multimodal observation data, centered on the predicted spatial location, to extract the real multimodal feature vector for the corresponding spatiotemporal point. For each future time point, calculating the multimodal similarity between the predicted sensor observation features and the real multimodal feature vector, as the single-point evidence support for that time point. Aggregating the single-point evidence support for each future time point along the time dimension to obtain the comprehensive evidence support for the candidate state trajectory, and constructing the evidence support set from the comprehensive evidence support of each candidate state trajectory, which serves as the output of the anomaly metric function.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the process of determining the physical law violation index based on the degree of persistent deviation of evidence support specifically includes: obtaining the evidence support set of candidate state trajectories within the current and historical continuous time windows; detecting persistent deviation patterns from the evidence support set for each candidate state trajectory, where the persistent deviation pattern represents the cumulative state where the evidence support of the candidate state trajectory is below a preset normal threshold over multiple consecutive time windows; and calculating the physical law violation index based on the minimum comprehensive evidence support of each candidate state trajectory and the number of windows in which the persistent deviation pattern has persisted.

[0010] In conjunction with the first aspect mentioned above, one possible implementation also includes an end-to-end optimization step: A differentiable distance metric is used as the loss function to measure the distribution difference between the predicted and true distributions. The predicted distribution consists of a set of candidate state trajectories, while the true distribution consists of real multimodal observation data. The differentiable distance metric includes at least the bulldozer distance or the difference in maximum mean. Based on the distribution difference, the loss value is calculated, and the parameters of the trajectory extrapolation model and the anomaly metric function are updated simultaneously through gradient backpropagation to achieve end-to-end joint optimization between the trajectory extrapolation model and the anomaly metric function.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves acquiring multimodal sensor data and constructing a dynamic graph sequence. Specifically, this includes: acquiring multimodal data streams based on heterogeneous sensors deployed in the supply chain scenario; parsing and mapping the multimodal data streams to a unified semantic space to obtain the entity multimodal feature vectors of the tracked entities in each time slice; using each tracked entity in the scenario as a graph node, and the entity multimodal feature vector as the node attribute of the corresponding graph node; calculating and establishing graph edges representing the interaction strength between entities based on the spatial proximity between entities within the same time slice; and constructing a scene semantic snapshot graph for each time slice. The scene semantic snapshot graphs of consecutive time slices are stacked vertically along the time axis, and self-recursive connection edges are introduced between the same graph nodes in adjacent time slices to form a dynamic graph sequence. The weights of the self-recursive connection edges are dynamically calculated by a gated recurrent unit based on the node attributes of adjacent time steps, used to encode the temporal evolution characteristics of the entity's own state.

[0012] Secondly, a supply chain risk early warning and assessment system based on a digital twin platform is provided, comprising a communication unit and a processing unit. The communication unit receives multimodal data streams collected by multiple heterogeneous sensors deployed in the supply chain scenario and transmits them to the processing unit. The processing unit, through a cross-modal semantic alignment module built into the unit, parses and maps the multimodal data streams to a unified semantic space, constructing a dynamic graph sequence to characterize the physical entities and their interactions within the monitored supply chain scenario. In response to the loss of the target entity's identification signal, the physical trajectory extrapolation and verification module built into the processing unit uses the dynamic graph sequence corresponding to the time window before the loss as a condition to generate a candidate state trajectory set for the target entity. It then obtains real multimodal observation data corresponding to the candidate state trajectory set in time and space for cross-validation and calculates the physical law violation index. The early warning judgment and optimization control module built into the processing unit calculates the abnormality measurement index of the target entity according to the physical law violation index, and generates an early warning trigger signal when the abnormality measurement index meets the preset early warning conditions. It also updates the internal parameters of the cross-modal semantic alignment module and the physical trajectory extrapolation and verification module according to the gradient of the backpropagation of the end-to-end loss function.

[0013] In conjunction with the second aspect mentioned above, in one possible implementation, the cross-modal semantic alignment module specifically includes: a multi-source feature parsing submodule, a graph construction and spatiotemporal coding submodule, and a first data cache for connecting the two. The multi-source feature parsing submodule includes a parallel-configured RFID baseband signal processor, a deep neural network accelerator, and an industrial fieldbus interface, used to synchronously extract phase and confidence joint encoding, instance-level re-identification feature vectors, and motion state encoding from the multimodal data stream, respectively, and project these three into entity multimodal feature vectors through a cascaded first multilayer perceptron. The graph construction and spatiotemporal coding submodule includes a tensor processing unit and a gated recurrent unit array. The tensor processing unit is used to construct a scene semantic snapshot graph by time slice, using entity multimodal feature vectors as node attributes and spatial proximity as edge weights. The gated recurrent unit array is used to calculate the self-recursive connection edge weights for the same graph node across time steps, stacking the scene semantic snapshot graphs along the time axis into a dynamic graph sequence. The first data cache is used to store the intermediate feature maps output by the multi-source feature parsing submodule, which are then used by the map construction and spatiotemporal coding submodules for tensor-level batch reading.

[0014] In conjunction with the second aspect mentioned above, in one possible implementation, the physical trajectory extrapolation and verification module is integrated into a field-programmable gate array (FPGA) or application-specific integrated circuit (ASIC) chip. Specifically, it includes: a conditional diffusion inverse extrapolation submodule, a spatiotemporal cross-validation and anomaly measurement submodule, and a second data cache. The conditional diffusion inverse extrapolation submodule comprises a backbone denoising network and a parallel-connected differentiable physics simulator. The backbone denoising network predicts the initial denoising direction based on the input noise and the conditional dynamic graph sequence. The differentiable physics simulator generates a physical gradient correction term based on Newton's laws of motion to guide the backbone denoising network and reconstruct a set of candidate state trajectories that conform to physical constraints. The spatiotemporal cross-validation and anomaly measurement submodule extracts predicted sensor observation features from the candidate state trajectory set and extracts real multimodal feature vectors from real multimodal observation data using a spatiotemporal attention query method. It outputs evidence support through a built-in similarity calculation engine and calculates a physical law violation index based on continuous deviation detection logic. The second data cache is used to share tensor data of the candidate state trajectory set between the conditional diffusion backward extrapolation submodule and the spatiotemporal cross-validation and anomaly measurement submodule.

[0015] This application provides a supply chain risk early warning and assessment method and system based on a digital twin platform. It enables a differential physical trajectory extrapolation engine, through an embedded differentiable physical simulator, to force the diffusion model to satisfy hard constraints such as inertia and momentum conservation when generating all possible trajectories. This fundamentally eliminates erroneous state reasoning that violates physical laws, such as "instantaneous appearance / disappearance," resulting in a smooth and continuous state of the digital twin. This ensures the confidence level of cumulative risk indicators such as long-term dwell timers, providing physically consistent trajectory extrapolations even during periods of complete occlusion. This achieves physics-driven, oscillatory, continuous state reasoning, completely eradicating the risk calculation failure problem caused by state jumps. Furthermore, through generative trajectory prediction and multimodal evidence cross-validation, the level of anomaly detection is elevated from "signal missing" to "violation of physical laws," greatly enhancing the ability to discover hidden risks. The evidence verification and anomaly measurement layer uses the "physical law violation index" as a warning basis. The index will only increase when multiple conventional assumptions that conform to physical laws are disproven by multimodal facts, thereby accurately identifying cargo movement or theft behaviors that seem to have coherent signals but are physically logically abnormal, and realizing the proactive discovery of new types of covert theft behaviors outside the rules. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the supply chain risk early warning and assessment method based on a digital twin platform provided in this application embodiment; Figure 2 A schematic diagram of the internal structure and generation process of the trajectory extrapolation model in the supply chain risk early warning and assessment method based on a digital twin platform provided in the embodiments of this application; Figure 3 The flowchart of the calculation of the physical gradient correction term by the differentiable physics simulator in the supply chain risk early warning assessment method based on the digital twin platform provided in the embodiments of this application is shown. Figure 4 In the supply chain risk early warning and assessment method based on a digital twin platform provided in the embodiments of this application, the flowchart of the anomaly measurement function for calculating the evidence support degree and the physical law violation index is shown. Figure 5 A schematic diagram of the end-to-end joint optimization mechanism in the supply chain risk early warning and assessment method based on a digital twin platform provided in the embodiments of this application; Figure 6 A flowchart illustrating the dynamic graph sequence construction process in the supply chain risk early warning and assessment method based on a digital twin platform provided in this application embodiment; Figure 7 The module architecture and hardware structure diagram of the supply chain risk early warning and assessment system based on the digital twin platform provided in the embodiments of this application are shown. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. In the description of this invention, it should be understood that the terms “first,” “second,” etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance or order.

[0019] Example 1: like Figure 1 As shown, this embodiment provides a supply chain risk early warning and assessment method based on a digital twin platform. This method constructs a complete closed-loop logic from multimodal perception to physical trajectory extrapolation, then to evidence cross-validation and early warning determination, providing an anchor point for subsequent physical priors and anomaly measurement. It should be understood that this embodiment is merely illustrative and not restrictive; those skilled in the art can make various modifications and variations without departing from the spirit of the invention.

[0020] Step 100: Acquire multimodal sensor data and construct a dynamic graph sequence. The dynamic graph sequence is used to characterize the physical entities and their interactions within the detected supply chain scenario.

[0021] In this context, the dynamic graph sequence is not a discrete message queue or an isolated set of feature vectors as in traditional techniques. Instead, it refers to a stacked structure of continuous scene semantic snapshot graphs that evolve over time, with the tracked entities as graph nodes and the spatial interaction strength between entities as graph edges. Furthermore, this sequence simultaneously encodes the temporal evolution features of the entities themselves and the spatial contextual relationships between entities.

[0022] In terms of operational methodology, multimodal data streams are acquired from heterogeneous sensors deployed in supply chain scenarios (such as RFID readers, surveillance cameras, forklift controllers, Bluetooth beacons, etc.). The parsed heterogeneous data can then be mapped to a unified semantic space to generate entity multimodal feature vectors as node attributes. Graph edges are then established based on the spatial proximity between entities within the same time slice, and finally, these edges are stacked vertically along the time axis to form a dynamic graph sequence.

[0023] This step constructs a dynamic graph sequence, uniformly projecting multimodal data onto a continuous, object-centric tensor graph neural network. This softens the hard decision boundaries regarding the presence or absence of a single signal, providing structured, differentiable conditional inputs for subsequent continuous inference defenses and completely eliminating the potential for state oscillations caused by contextual fragmentation. This avoids the problem of existing event-driven architectures treating each sensor report as an isolated atomic event. This hard decision of "existence if read, disappear if not read" completely abandons the continuous context of the object's own motion and existence, leading to atomized state processing and contextual fragmentation.

[0024] Step 200: In response to the loss of the identification signal of the target entity, the dynamic graph sequence corresponding to the time window before the loss is called as a condition and input into the trajectory extrapolation model to generate a set of candidate state trajectories of the target entity; the trajectory extrapolation model is configured as a generative model that generates future state sequences based on historical states, and physical prior constraints are introduced in the generation process to correct the inference direction.

[0025] The trajectory extrapolation model is a generative neural network architecture that generates hypotheses about future states based on historical state sequences. Its output is a set of candidate state trajectories containing high-order physical state tensors such as position, velocity, and acceleration. Physical prior constraints refer to the use of first principles of physics, such as Newton's laws of motion, as hard constraints during the inference process of the generative model. Physical gradient correction terms are calculated using a differentiable physics simulator to forcibly correct the data-driven inference direction, ensuring that the generated trajectory satisfies mechanical laws such as inertia and momentum conservation.

[0026] In some implementations, when the target entity (such as a high-value goods pallet) enters an occluded state, causing the identification signal to be lost, the system no longer immediately triggers the "destroy" or "leave" event. Instead, it uses the dynamic graph sequence within a certain time window before the loss as the conditional input, starts from random noise, and performs the reverse diffusion process of the generative model. In each step of inference, it integrates the physical gradient correction term and iteratively samples multiple future state trajectory hypotheses that conform to physical constraints to form a candidate state trajectory set.

[0027] Therefore, this step, by introducing generative extrapolation and physical prior constraints, avoids hard decisions during the signal loss vacuum period. Instead, it continuously and smoothly extrapolates the existence state based on physical laws and interactions with other entities, fundamentally eliminating erroneous state reasoning such as "instantaneous appearance / disappearance" that does not conform to physical laws, and establishing a continuous reasoning defense line. This addresses the problem that existing technologies default to executing the simplest static model when information is missing, i.e., "the object disappears in place after its last appearance." This embedded erroneous prior model contradicts the reality of continuous motion of objects in the physical world, leading to continuous erroneous state reasoning by the system.

[0028] Step 300: Obtain real multimodal observation data corresponding to the candidate state trajectory set in time and space, and cross-validate it with the predicted sensor observation features in the candidate state trajectory set. Calculate the evidence support degree through the anomaly metric function, and determine the physical law violation index based on the degree of continuous deviation of the evidence support degree.

[0029] Among them, evidence support refers to the similarity or matching metric between the predicted candidate trajectory and the subsequent real-world multimodal observation data in the spatiotemporal and modal dimensions. It reflects the degree to which the predicted hypothesis is supported by real evidence from the physical world. The physical law violation index, on the other hand, is an anomaly score calculated based on the cumulative state where evidence support is consistently below the normal threshold when all conventional trajectory hypotheses that conform to physical laws are disproven by multimodal facts. It characterizes the degree to which physically reasonable behavior cannot explain the violation of the current observation.

[0030] In some implementations, for each generated candidate state trajectory, the system extracts predicted sensor observation features (such as predicted spatial location and predicted visual feature vectors) for future time points. When real physical time advances to the corresponding time point, a spatiotemporal attention query is performed on the real multimodal observation data centered on the predicted spatial location to extract the real feature vectors and calculate the multimodal similarity as the single-point evidence support. After aggregation along the time dimension, a comprehensive evidence support score is obtained. Finally, a persistent deviation pattern where the evidence support score is below a preset normal threshold is detected within multiple consecutive time windows. When all candidate trajectories exhibit this pattern, a physical law violation index is calculated based on the minimum value of the comprehensive evidence support score and the number of persistent windows.

[0031] This step elevates the level of anomaly detection from "signal missing" to "violation of physical laws" through generative trajectory prediction and multimodal evidence cross-validation. The physical law violation index only increases when multiple conventional assumptions conforming to physical laws are falsified by multimodal facts, thus accurately identifying hidden risks where signals appear coherent but physical logic is anomalous. This solves the problems of existing technologies where signal missing at a single point in time triggers state reversal, failing to distinguish between sensor noise and genuine violations of physical laws, and completely failing to identify covert theft mimicking normal paths. Furthermore, it enables proactive detection of novel covert theft behaviors outside the established rules.

[0032] Step 400: Calculate the anomaly measurement index of the target entity based on the degree of evidence support, and trigger a supply chain risk warning when the anomaly measurement index meets the preset warning conditions.

[0033] Among them, the anomaly metric refers to a quantitative value calculated based on the evidence support set and the physical law violation index, used to ultimately determine whether a target entity poses a risk. It can be a weighted integral of the violation index or a function of the support decay rate. Triggering a supply chain risk warning means that when the anomaly metric exceeds a preset threshold or meets specific logical conditions, the system generates an alarm signal and pushes it to the front end of the digital twin platform and maintenance personnel, indicating potential cargo loss, abnormal detention, or covert theft.

[0034] In some implementations, the physical law violation index and the evidence support set are used as inputs. The specific value of the anomaly measurement index is calculated through the built-in early warning judgment logic. For example, the violation index and the minimum support are nonlinearly fused. When the value exceeds the preset safety threshold, an early warning trigger signal is immediately generated, and the abnormal entity and its trajectory hypothesis are highlighted on the digital twin 3D monitoring interface.

[0035] It should be noted that by using the abnormal measurement indicators output by the continuous reasoning defense line for early warning judgment, compared with the existing technology that relies on the atomic event-driven dwell timer, it can avoid the situation where the risk alarm is constantly reset to zero when the signal fluctuates at high frequency, thus completely eliminating the risk calculation failure problem caused by state transition. This makes the confidence of cumulative risk indicators such as long-term dwell timers extremely high, and can ensure the continuous effectiveness of risk rules even during full occlusion.

[0036] In summary, this embodiment achieves oscillatory, continuous state reasoning driven by physical laws, completely eliminating the risk calculation failure problem caused by state transitions. Simultaneously, through generative trajectory prediction and multimodal evidence cross-validation, the cognitive level of anomaly detection is elevated from "signal missing" to "violation of physical laws," greatly enhancing the ability to discover hidden risks.

[0037] Example 2: like Figure 2-3 As shown, based on step 200 of Example 1, this example further delves into the internal architecture and generation mechanism of the trajectory extrapolation model, providing algorithm-level defense depth for the core innovation point "physical prior constraints", and explaining why it is not a purely data-driven black-box generation.

[0038] It should be understood that this embodiment is illustrative only and not restrictive, and those skilled in the art can make various modifications and variations without departing from the spirit of the invention.

[0039] Step 201: Using the dynamic graph sequence corresponding to the time window before the loss as a condition, perform the reverse diffusion process starting from random noise.

[0040] Among them, the conditional diffusion model refers to a deep generative network architecture constrained by historical observation conditions, which mainly restores structured data by gradually stripping away random Gaussian noise, rather than generating data freely without conditions. The reverse diffusion process refers to an iterative sampling mechanism that starts from a pure noise state, works backward along the time step, and gradually corrects the noise state until a clear physical trajectory is restored.

[0041] In some implementations, when the target entity identification signal is lost, the system uses the dynamic graph sequence within a time window before the loss as the input model of the conditional tensor, and initializes a random noise tensor with the same dimension as the target trajectory as the starting point to start multi-step reverse diffusion.

[0042] Step 202: Within each denoising time step, the backbone denoising network predicts the initial denoising direction based on the current noise state and conditions, and synchronously outputs the predicted original physical state corresponding to the current noise state. The predicted original physical state includes the position, velocity and acceleration of the target entity.

[0043] The backbone denoising network refers to the neural network backbone built based on U-Net or Transformer architecture. Its core function is to predict the noise components contained in the current noise state to guide denoising. The prediction of the original physical state refers to the original physical state tensor, which is not contaminated by noise, directly inferred by the auxiliary decoder head while the backbone network predicts the denoising direction. This tensor explicitly contains the three-dimensional position coordinates, three-dimensional velocity vector, and three-dimensional acceleration vector of the target entity.

[0044] In some implementations, for each denoising time step, the backbone denoising network receives the current noise state and conditional dynamic graph sequence, fuses the conditional information through a cross-attention mechanism, outputs a preliminary denoising direction vector, and simultaneously outputs the predicted original physical state through a parallel decoding branch.

[0045] It should be understood that although this embodiment lists position, velocity, and acceleration as components of the predicted original physical state, in other embodiments, the state may also include higher-order kinematic parameters such as angular velocity and attitude angle, as long as the input requirements of the differentiable physics simulator are met.

[0046] Step 203: The differentiable physics simulator receives the predicted original physical state, calculates the corresponding reasonable physical state according to Newton's laws of motion, and generates a physical gradient correction term to correct the initial denoising direction.

[0047] The differentiable physics simulator is a computational graph module embedded in the generative model, built based on physical equations. All its operations support automatic differentiation, enabling it to calculate a reasonable state conforming to the laws of mechanics based on the input physical state and backpropagate gradients. The physical gradient correction term is a vector calculated by the physics simulator, with the same dimension as the initial denoising direction, used to pull the data-driven denoising direction back into the constraint space that conforms to physical laws.

[0048] In some implementations, a differentiable physics simulator receives a predicted original physical state, then calculates the physically reasonable state that the object should reach under that state based on Newton's laws of motion, and then calculates the deviation between the two and maps the deviation to a physical gradient correction term.

[0049] It should be noted that this step introduces differentiable physical equations as hard constraints, forcibly verifying the physical rationality at every step of the generation process, fundamentally suppressing anti-physics inferences. This can effectively solve the problem of pure data-driven black-box generation lacking physical common sense and producing "teleportation" trajectories that do not conform to physical laws, such as objects changing direction instantly or moving without acceleration without the action of external force.

[0050] The process by which a differentiable physics simulator calculates the physics gradient correction term specifically includes: Step 2031: Extract the current position, velocity and acceleration of the target entity from the predicted original physical state, and extract the interaction force vectors of neighboring entities on the target entity from the dynamic graph sequence.

[0051] Among them, the interaction force vector refers to the spatial force vector calculated based on the graph edge attributes in the dynamic graph sequence, which represents the spatial force exerted by neighboring entities on the target entity. It covers abstract interaction effects such as thrust, pull, and friction.

[0052] In some implementations, the current position, velocity, and acceleration of the target entity are parsed from the predicted original physical state. At the same time, the interaction force vectors of neighboring entities on the target entity can be extracted from the graph edge features of the conditional dynamic graph sequence based on spatial proximity and entity type.

[0053] Step 2032: Input the preset mass, interaction force vector and environmental force vector of the target entity into the built-in rigid body dynamics solver, calculate the corrected acceleration according to Newton's second law, and obtain the physically reasonable state that satisfies the mechanical constraints through numerical integration.

[0054] Among them, the rigid body dynamics solver refers to a differentiable numerical integrator constructed based on Newton's second law, which can calculate acceleration based on forces and mass, and obtain the future state through integration using the Euler method or the Runge-Kutta method. Meanwhile, the environmental force vector refers to the global environmental force vector, such as gravity and friction.

[0055] In some implementations, the system sums the target entity's preset mass, interaction force vector, and environmental force vector to obtain the net external force. Then, according to Newton's second law, the corrected acceleration can be calculated, and then through numerical integration, the physically reasonable state that satisfies the mechanical constraints can be obtained.

[0056] It should be understood that although this embodiment lists Newton's second law, in other embodiments, the law of conservation of momentum or the law of conservation of energy can also be introduced as supplementary constraints, as long as the requirements for differentiable calculation are met.

[0057] Step 2033: Calculate the element-wise physical residual between the physical reasonable state and the predicted original physical state, and map the physical residual to the noise prediction space through the encoding network to generate the physical gradient correction term.

[0058] Here, the element-wise physical residual refers to the difference tensor between the physically plausible state and the predicted original physical state in each dimension, such as position, velocity, and acceleration. Meanwhile, the encoding network refers to a multilayer perceptron or a small convolutional network, used to map the physical residual to a vector space with the same dimensions as the noise prediction space.

[0059] In some implementations, the element-wise physical residual between the physically reasonable state and the predicted original physical state can be directly calculated. The physical residual is then mapped to the noise prediction space via an encoding network to generate a physical gradient correction term.

[0060] It should be noted that since the physical residual and the noise prediction space are often misaligned in dimension and semantics, direct addition will lead to optimization failure. Therefore, this step uses an encoding network to perform spatial mapping, which realizes the smooth projection of physical constraints onto the gradient space of the generative model, ensuring that the physical gradient correction term can be effectively absorbed by the backbone denoising network.

[0061] Step 204: After adjusting the initial denoising direction with the physical gradient correction term, update the current noise state and enter the next denoising time step. Iterate until at least one future state trajectory that meets the physical constraints is restored, forming a candidate state trajectory set.

[0062] This step details the iterative fusion mechanism of the denoising direction and the physical gradient. In terms of operation, the system performs a weighted fusion of the initial denoising direction and the physical gradient correction term, for example, by linearly adding them using learnable weight parameters, to obtain the final denoising direction. Based on this, the current noise state is updated, and the system proceeds to the next denoising time step.

[0063] Meanwhile, the iterative process is executed repeatedly until at least one future state trajectory that meets the physical constraints is restored, and multiple trajectories constitute a candidate state trajectory set.

[0064] It should be understood that although this embodiment lists a linear addition fusion method, in other embodiments, a gating mechanism or a nonlinear fusion network can also be used to integrate the two, as long as the physical gradient correction term can effectively intervene in the denoising direction.

[0065] Based on the above technical solution, by incorporating the first principles of physics as hard constraints into the denoising process of the conditional diffusion model, the traditional "black box" sequence generation of generative models is abandoned. Pure data-driven generation may produce "instantaneous" trajectories that do not conform to physical laws. Introducing differentiable physical equations as hard constraints fundamentally suppresses anti-physics inferences, ensuring that all generated trajectory samples are kinematically reasonable. For example, trajectories that do not conform to physical laws, such as instantaneous changes of direction or movement without acceleration, are directly suppressed. This achieves physics-driven, oscillatory, continuous state reasoning, completely eliminating the risk of computational failure caused by state jumps.

[0066] Example 3: like Figure 4 As shown, based on step 300 of Example 1, this example further refines the internal calculation logic and measurement standard of the anomaly measurement function, providing a computable measurement standard for cross-validation and explaining why it is not a simple threshold comparison.

[0067] It should be understood that this embodiment is illustrative only and not restrictive, and those skilled in the art can make various modifications and variations without departing from the spirit of the invention.

[0068] The process of calculating the support of evidence using anomaly measurement functions specifically includes: Step 301: Extract the predicted sensor observation features at future time points from the predicted state sequence of each candidate state trajectory in the candidate state trajectory set. The predicted sensor observation features include at least the predicted spatial location and the predicted visual feature vector.

[0069] Among these, predicted sensor observation features refer to the set of virtual observation data generated by the trajectory extrapolation model, which represents the target entity that should be captured by various sensors at future points in time. These are not abstract state coordinates, but rather feature representations directly mapped to the sensor's perceptual domain. Predicted spatial location refers to the predicted three-dimensional coordinates in the global coordinate system. Predicted visual feature vectors refer to the deep feature embedding of the target entity's appearance from the predicted perspective, such as instance-level re-identification feature vectors.

[0070] In some implementations, the system decodes the corresponding predicted spatial location and predicted visual feature vector point by point from the candidate state trajectory set, along its predicted future time series, for each candidate trajectory.

[0071] It should be understood that although this embodiment lists predicted spatial location and predicted visual feature vector as components of predicted sensor observation features, in other embodiments, the features may also include predicted radio frequency signal phase and confidence encoding, predicted Bluetooth signal strength fingerprint, etc., as long as they meet the requirement of being able to be compared with real multimodal observation data in the same dimension.

[0072] Step 302: Using future time points as indexes, perform spatiotemporal attention queries from real multimodal observation data, centered on the predicted spatial location, to extract the real multimodal feature vectors of the corresponding spatiotemporal points.

[0073] Spatiotemporal attention query refers to a mechanism that actively retrieves evidence from real data tensors based on predicted coordinates. Its core lies in aggregating real features in the spatial and temporal neighborhoods with the predicted location as the attention center, rather than passively waiting for sensor reports. The real multimodal feature vector refers to the entity feature tensor located near the predicted location, constructed by the cross-modal semantic alignment module on the current real time slice.

[0074] In some implementations, when real physical time advances to the future time point indexed in step 301, the system acquires the current real multimodal observation data, delineates a spatiotemporal search window centered on the predicted spatial location, performs attention-weighted aggregation within the window, and extracts the real multimodal feature vector of the corresponding spatiotemporal point.

[0075] This step transforms passive event-driven processes into proactive evidence retrieval through spatiotemporal attention queries. Even when the main identifier signal of the target entity is missing, it can extract relevant evidence from the surrounding environment through spatial proximity and temporal continuity, filling the blind spots of missing information.

[0076] Step 303: For each future time point, calculate the multimodal similarity between the predicted sensor observation features and the actual multimodal feature vector, and use it as the single-point evidence support for that time point.

[0077] Among them, single-point evidence support refers to the quantitative value of the degree of agreement between the predicted hypothesis and the actual observation at a specific point in time. It can be calculated by various measures such as cosine similarity, the reciprocal of Euclidean distance, or the correlation coefficient between feature vectors.

[0078] In some implementations, the system aligns the predicted sensor observation features extracted in step 301 with the real multimodal feature vectors extracted in step 302, calculates the multimodal similarity between the two in the embedding space, and maps the similarity value to the interval between 0 and 1 as the single-point evidence support at that time point.

[0079] It should be understood that although this embodiment lists cosine similarity as a calculation method, in other embodiments, differentiable bulldozer distance or deep learning metric networks can also be used to calculate similarity, as long as they meet the requirements of reflecting the degree of feature matching and supporting gradient backpropagation.

[0080] Step 304: Aggregate the single-point evidence support of each future time point along the time dimension to obtain the comprehensive evidence support of the candidate state trajectory. Construct the evidence support set of the comprehensive evidence support of each candidate state trajectory as the output of the anomaly measurement function.

[0081] The overall evidence support refers to the time-series aggregated value of the evidence consistency of a candidate state trajectory over the entire prediction time window. It can be the mean, weighted integral, or decreasing cumulative support of individual points. The evidence support set refers to the vector or set composed of the overall evidence support of all candidate state trajectories, reflecting the system's overall evidence assessment of various physically plausible assumptions.

[0082] In some implementations, the system performs weighted aggregation of the single-point evidence support at each time point along the time dimension. For example, more recent evidence is given higher weight to obtain the comprehensive evidence support of each candidate trajectory. Then, the comprehensive values ​​of all candidate trajectories are summarized and output as a set of evidence support.

[0083] Furthermore, the process of determining the violation exponent of physical laws based on the degree of persistent deviation from the level of evidence support specifically includes: Step 305: Obtain the set of evidence support for candidate state trajectories within the current and historical continuous time windows.

[0084] In some implementations, the system not only obtains the evidence support set for the current time window, but also backtracks to obtain the evidence support sets for multiple consecutive historical time windows, forming a temporal evidence support sequence, which provides temporal contextual support for subsequent pattern detection.

[0085] Step 306: For each candidate state trajectory, detect a persistent deviation pattern from the evidence support set. The persistent deviation pattern represents the cumulative state in which the evidence support of the candidate state trajectory is lower than the preset normal threshold within multiple consecutive time windows.

[0086] The persistent divergence pattern refers to a prolonged period of low evidence support over time. This is not a sudden, single-point drop, but rather a cumulative anomaly where physically plausible assumptions consistently fail to gain genuine multimodal evidence support across multiple consecutive time windows. The preset normal threshold is a critical support value derived from historical normal operating data, distinguishing between noise fluctuations and genuine anomalies.

[0087] In some implementations, the system scans the evidence support sequence of each candidate state trajectory along the time axis to detect whether there is a situation where the support is lower than the preset normal threshold in multiple consecutive time windows. If such a situation exists, it is marked as a continuous deviation mode.

[0088] It should be understood that although this embodiment lists the detection of persistent deviation patterns based on a fixed threshold, in other embodiments, dynamic adaptive thresholds or abnormal state transition detection based on hidden Markov models can also be used, as long as the requirement of distinguishing between instantaneous noise and long-term anomalies is met.

[0089] This step detects persistent deviation patterns, requiring anomalies to be continuous over time. This effectively filters out transient signal jitter interference, shifting the focus of anomaly detection from instantaneous signal loss to persistent violations of physical laws. It solves the problem that evidence missing at a single time point is easily confused by sensor noise, leading to frequent false alarms.

[0090] Step 307: When the evidence support set of all candidate state trajectories shows a continuous deviation pattern, calculate the physical law violation index based on the minimum value of the comprehensive evidence support of each candidate state trajectory and the number of windows in which the continuous deviation pattern has continued.

[0091] The physical law violation index refers to the anomaly score calculated based on the depth and breadth of persistently low evidence support when all conventional trajectory assumptions that conform to physical laws are falsified by multimodal facts. It characterizes the degree of violation that cannot be explained by physically reasonable behavior.

[0092] In some implementations, when the system determines that all candidate state trajectories exhibit a continuous deviation pattern, the minimum value of the comprehensive evidence support of each trajectory is extracted as a measure of deviation depth, and the number of windows in which the continuous deviation pattern has persisted is counted as a measure of deviation breadth. The two are then nonlinearly fused and calculated, for example, by multiplying the reciprocal of the minimum value by the number of windows and then normalizing it to obtain the physical law violation index.

[0093] It should be noted that the violation index can only be triggered when all physically reasonable assumptions are continuously falsified. This means that even deliberately disguised theft can be identified as an anomaly that does not conform to physical laws due to the lack of coherent multimodal interaction evidence, thus raising the cognitive level of anomaly detection from signal absence to violation of physical laws.

[0094] Based on the above technical solution, a calculable metric is provided for cross-validation, explaining why it is not a simple threshold comparison. The lack of evidence at a single time point may be sensor noise, while a continuous deviation pattern can distinguish noise from a violation of true physical laws, elevating anomaly detection from signal absence to a violation of physical laws. The physical law violation index only increases when multiple conventional hypotheses conforming to physical laws are continuously falsified by multimodal facts, thereby accurately identifying hidden risks that appear coherent but have anomalous physical logic, greatly enhancing the ability to discover hidden risks.

[0095] Example 4: like Figure 5 As shown, based on Example 1, this example further delves into the training dimension, which is independent of the inference process, to provide a parameter update mechanism for binding the functionally limited trajectory extrapolation model and the anomaly measurement function, and to establish a closed-loop optimization defense point.

[0096] It should be understood that this embodiment is illustrative only and not restrictive, and those skilled in the art can make various modifications and variations without departing from the spirit of the invention.

[0097] The method of the present invention further includes an end-to-end optimization step: using a differentiable distance metric as a loss function to measure the distribution difference between the predicted distribution and the true distribution, wherein the predicted distribution consists of a set of candidate state trajectories and the true distribution consists of real multimodal observation data, and the differentiable distance metric includes at least bulldozer distance or maximum mean difference; based on the distribution difference, the loss value is calculated, and the parameters of the trajectory extrapolation model and the anomaly metric function are updated simultaneously through gradient backpropagation, in order to achieve end-to-end joint optimization between the trajectory extrapolation model and the anomaly metric function.

[0098] Differentiable distance metric refers to a mathematical metric tool that can measure the difference between two probability distributions and supports gradient backpropagation calculation. Its core is that it can not only output the distance between distributions, but also the partial derivatives with respect to the input parameters, thereby guiding the parameter update of the neural network.

[0099] The bulldozer distance, also known as the Wasserstein distance, is a measure of the minimum "transfer" cost required to transform one distribution into another. It can smoothly measure the difference even when the support sets of two distributions do not overlap, avoiding the gradient vanishing problem of traditional measures such as KL divergence when the distributions do not intersect.

[0100] Maximum mean difference is another distribution distance metric based on kernel functions. It measures the difference by mapping samples to the reproducing kernel Hilbert space and calculating the distance between the mean vectors of the two distributions. It is particularly suitable for processing the distribution comparison of high-dimensional feature tensors.

[0101] The predicted distribution refers to the probability distribution of future states formed by a set of multiple candidate state trajectories generated by the trajectory extrapolation model. It reflects the model's overall inference assumptions about the future evolution of the physical world.

[0102] The true distribution refers to the actual state distribution constituted by the subsequent emergence of multimodal observation data, which represents the objective facts that actually occur in the physical world.

[0103] It should be understood that although this embodiment lists bulldozer distance and maximum mean difference as specific forms of differentiable distance measurement, in other embodiments, contrast loss based on deep neural networks or energy distance-based measurement methods can also be used, as long as they meet the requirements of measuring distribution differences and supporting automatic differential calculation.

[0104] In terms of operational methods, the execution logic for end-to-end optimization is as follows: The system first parameterizes the candidate state trajectory set generated by the trajectory extrapolation model into a predicted distribution, and simultaneously parameterizes the real multimodal observation data extracted by the cross-modal semantic alignment module into a true distribution. Then, it selects either the bulldozer distance or the maximum mean difference as the loss function to calculate the distribution difference between these two distributions. For example: When using bulldozer distance, the system performs optimal transmission matching between the set of state feature vectors of candidate trajectories and the set of feature vectors of actual observations, and calculates the minimum transportation cost as the distribution difference value. When using the maximum mean difference, the system maps the two types of features to a high-dimensional space using a kernel function and calculates the mean difference.

[0105] Based on the calculated distribution differences, the loss value can be further calculated, and gradient backpropagation can be performed through the automatic differentiation engine built into the deep learning framework. Since the trajectory extrapolation model and the anomaly metric function form a tight data flow connection during the inference phase, the gradient signal can be fed back from the output of the anomaly metric function, through the comparison logic of evidence support calculation and spatiotemporal attention query, all the way to the backbone denoising network and differentiable physics simulator of the trajectory extrapolation model, thereby simultaneously updating the internal parameters of the two models and achieving end-to-end joint optimization.

[0106] This allows this embodiment to completely break down information barriers between cascaded models by introducing end-to-end joint optimization. Driven by gradient backpropagation, the trajectory extrapolation model automatically adjusts its generation strategy, tending to output trajectory features with high discriminative power that are easier to retrieve and falsify in real multimodal data. Simultaneously, the anomaly metric function automatically learns the optimal evidence comparison pattern and similarity threshold, eliminating the need for manually set hard rules. This closed-loop optimization mechanism enables extrapolation and verification to form a mutually reinforcing co-evolution, constructing a fully differentiable closed-loop system from multimodal perception to trajectory prediction to verification. This fundamentally eliminates the accumulation of cascaded errors and dependence on manual rules, significantly improving the system's adaptability in different supply chain scenarios.

[0107] Example 5: like Figure 6 As shown, based on step 100 of embodiment 1, this embodiment further sinks the internal construction logic of the dynamic graph sequence to provide structured representation sinking for the front-end input multimodal sensor data, explaining why it is not a discrete message queue.

[0108] It should be understood that this embodiment is merely illustrative and not restrictive, and those skilled in the art can make various modifications and variations without departing from the spirit of the invention. Step 501: Based on heterogeneous sensors deployed in the supply chain scenario, acquire multimodal data streams, parse and map the multimodal data streams to a unified semantic space to obtain the entity multimodal feature vectors of the tracked entity in each time slice.

[0109] The unified semantic space refers to a high-dimensional, continuous vector embedding space. Its core function is to nonlinearly project heterogeneous sensor data from different modalities with different physical dimensions and data structures, such as the phase and confidence of RFID radio frequency signals, instance segmentation masks and re-identification feature vectors of visual cameras, odometer and forklift load status codes of forklift or AGV controllers, and signal strength fingerprints of Bluetooth beacons, onto the same metric system. This allows data from different modalities to perform similarity calculations and feature fusion in the same mathematical space.

[0110] Entity multimodal feature vectors refer to the comprehensive state representation tensors formed by splicing or fusing the perception data of all available modalities of each tracked entity in a scene, such as high-value goods pallets, forklifts, and supervisors, within each tiny time slice. They not only contain the entity's identity information but also encode the entity's instantaneous motion state and appearance features.

[0111] In some implementations, the system continuously acquires multimodal data streams based on heterogeneous sensors deployed in the supply chain scenario. For each time slice, the original features of each modality are extracted by a parallel feature parsing network. Then, these heterogeneous features are forcibly aligned by a cascaded multilayer perceptron or a cross-modal attention fusion network and projected onto a unified semantic space to generate the entity multimodal feature vector of each tracked entity in that time slice.

[0112] It should be understood that although this embodiment lists RFID, vision, forklift controller, and Bluetooth beacon as specific forms of heterogeneous sensors, other embodiments may also include point cloud features of lidar, ultrasonic ranging features, or environmental features of temperature and humidity sensors, as long as they meet the requirement of providing multi-dimensional state representation of entities.

[0113] Step 502: Using each tracked entity in the scene as a graph node, the entity's multimodal feature vector as the node attribute of the corresponding graph node, and based on the spatial proximity between entities within the same time slice, calculate and establish graph edges that characterize the interaction strength between entities, and construct a scene semantic snapshot graph for each time slice.

[0114] It should be noted that the scene semantic snapshot graph refers to a complete mathematical abstraction of all entities and their interactions within a specific time slice of a supply chain scenario. It is not a simple discrete message queue, but a dense connection network organized in the form of a graph data structure, containing node attributes and edge weights.

[0115] Spatial proximity refers to the distance between two entities in physical three-dimensional space within the same time slice. It can be calculated using Euclidean distance, Manhattan distance, or based on the overlap of sensor sensing range.

[0116] Interaction intensity refers to the potential physical interaction between entities due to spatial proximity, and is a quantitative representation of the magnitude of the force. For example, the interaction intensity is extremely high when a forklift and a pallet are very close, while the interaction intensity is extremely low when pallets are stationary at a distance.

[0117] In some implementations, each tracked entity within the scene is treated as a graph node, and the entity's multimodal feature vector is directly assigned as the node attribute of the corresponding graph node. Subsequently, the spatial proximity between all entity pairs within the same time slice is calculated. For example, by calculating the Euclidean distance between the 3D coordinates of the nodes and using a Gaussian kernel function or a threshold truncation function, the distance is mapped to an edge weight between 0 and 1. This establishes a graph edge characterizing the interaction strength between entities, thereby constructing a complete scene semantic snapshot for each time slice.

[0118] It should be understood that although this embodiment lists the calculation of interaction strength based on spatial proximity, in other embodiments, edge weights can also be calculated by combining dimensions such as the consistency of the movement direction of entities and the frequency of historical interactions, as long as the requirement of reflecting the real physical interaction relationship between entities is met.

[0119] This step embeds isolated entity states into a dense interaction network by constructing a scene semantic snapshot graph, so that the state of each entity is tightly bound to the context of its neighbors, completely eliminating the risk of state oscillation caused by context fragmentation.

[0120] Step 503: Stack the scene semantic snapshots of consecutive time slices vertically along the time axis, and introduce self-recursive connection edges between the same graph nodes of adjacent time slices to form a dynamic graph sequence. The weights of the self-recursive connection edges are calculated by the gated recurrent unit based on the node attributes of adjacent time steps to form the hidden state. The hidden state is then mapped to scalar weight values ​​through a fully connected layer and a nonlinear activation function to encode the temporal evolution features of the entity's own state.

[0121] Among them, self-recursive connection edges refer to special edge structures that connect graph nodes corresponding to the same entity in adjacent time slices. Their core function is to explicitly establish temporal dependency channels in the topology of the graph network, so that the node state at the current moment can directly receive and integrate node information from historical moments.

[0122] Gated recurrent units (GRUs) are lightweight recurrent neural network structures that dynamically control the proportion of historical information flowing into the current state through the collaborative mechanism of input gates, forget gates, and output gates, thereby effectively filtering noise and retaining long-term dependencies.

[0123] In some implementations, scene semantic snapshots of consecutive time slices are stacked vertically along the time axis, and self-recursive connection edges are introduced between the same graph nodes in adjacent time slices. For each entity node, the gated loop unit receives the node attribute vector from the previous time step and the input feature vector from the current time step, calculates the weight of the self-recursive connection edges through internal gated logic, and outputs the updated value of the current node attributes that incorporates historical evolution information. Finally, the snapshots of all time slices are integrated with the self-recursive connection edges to form a dynamic graph sequence that evolves over time.

[0124] It should be understood that although this embodiment lists a gated recurrent unit as a specific mechanism for calculating the weights of self-recursive connection edges, in other embodiments, a long short-term memory network or a temporal attention mechanism can be used instead, as long as the requirement of dynamically encoding the temporal evolution characteristics of the entity's own state is met.

[0125] This step introduces self-recursive connection edges and gated loop units to explicitly model temporal dependencies in the topology of the graph network. This enables the dynamic graph sequence to encode not only the intensity of spatial interactions between entities, but also the temporal evolution characteristics of the entities' own states. This provides structured and differentiable conditional inputs for the trajectory extrapolation model, fundamentally eliminating inference errors caused by context fragmentation.

[0126] In summary, based on the above technical solutions, by constructing a unified semantic space and scene semantic snapshot graph, the semantic barriers between modalities and the isolated state between entities are broken down, enabling multi-source data to mutually verify and complement each other within the same mathematical space. Simultaneously, by introducing self-recursive connection edges and gated loop units, temporal dependencies are explicitly modeled in the graph network topology, giving the system continuous context awareness capabilities and fundamentally eliminating inference errors caused by context fragmentation. This provides structured representation sinking for the pre-input multimodal sensor data, explaining why it is not a discrete message queue. It also allows the dynamic graph sequence to simultaneously encode the spatial interaction strength between entities and its own temporal evolution, providing structured and differentiable conditional input for trajectory extrapolation, completely eliminating context fragmentation. This embodiment...

[0127] Example 6: like Figure 7 As shown, this embodiment provides a supply chain risk early warning and assessment system based on a digital twin platform. This system is the hardware and software collaborative execution entity of the methods in the aforementioned embodiments 1 to 5, and constructs a complete closed-loop architecture from multimodal perception to physical trajectory extrapolation to evidence cross-verification and early warning determination.

[0128] It should be understood that this embodiment is illustrative only and not restrictive, and those skilled in the art can make various modifications and variations without departing from the spirit of the invention.

[0129] The supply chain risk early warning and assessment system specifically includes a communication unit and a processing unit.

[0130] Among them, the communication unit refers to the set of front-end interfaces for the system to interact with the physical world. It includes not only traditional network adapters and wireless receiving modules, but also fieldbus protocol parsing engines that directly interface with various industrial IoT sensing devices. These engines are responsible for converting the raw analog or digital signals collected by heterogeneous sensors into standardized data frames and transmitting them to the processing unit.

[0131] It is used to receive multimodal data streams collected by multiple heterogeneous sensors deployed in the supply chain scenario and transmit the multimodal data streams to the processing unit.

[0132] In terms of operation, the communication unit establishes a connection with heterogeneous sensors such as RFID readers, surveillance cameras, forklift controllers, and Bluetooth beacons deployed in ports, distribution centers, and other scenarios via wired or wireless means, and continuously receives multimodal data streams such as radio frequency signal frames, video compression streams, and motion status messages generated by these sensors.

[0133] The communication unit performs preliminary protocol parsing and timestamp alignment on these asynchronous and differently formatted data streams, and then pushes the preprocessed multimodal data streams in batches to the input buffer of the processing unit via a high-speed internal bus or network link.

[0134] The processing unit refers to the core computing and logic control center of the system. It can be an industrial control computer deployed at the edge, a high-performance server cluster, or a system-on-a-chip integrated on a dedicated chip. It carries the algorithm execution and state management logic of the three core modules.

[0135] During its use, the multimodal data stream can be directly parsed and mapped to a unified semantic space through the cross-modal semantic alignment module built into the processing unit, and a dynamic graph sequence can be constructed to characterize the physical entities and interaction relationships within the detected supply chain scenario.

[0136] Specifically, the cross-modal semantic alignment module refers to the first-level core engine within the processing unit, which is responsible for eliminating modal heterogeneity and constructing structured scene representations. It corresponds to the logical execution entity of steps 100 and 101 to 103 in the aforementioned method embodiment.

[0137] In terms of data flow, this module receives multimodal data streams transmitted by the communication unit, projects various heterogeneous data onto a unified semantic space through an internal feature parsing network to generate entity multimodal feature vectors, and constructs a dynamic graph sequence containing scene semantic snapshot graphs and self-recursive connection edges based on spatial proximity and gated loop units, providing structured and differentiable conditional inputs for subsequent physical extrapolation.

[0138] In response to the loss of the identification signal of the target entity, the physical trajectory extrapolation and verification module built into the processing unit calls the dynamic graph sequence corresponding to the time window before the loss as a condition to generate a set of candidate state trajectories of the target entity, and obtains real multimodal observation data corresponding to the spatiotemporal of the candidate state trajectory set for cross-validation to calculate the physical law violation index.

[0139] Among them, the physical trajectory extrapolation and verification module refers to the second-level core engine within the processing unit, which is responsible for continuous state reasoning and evidence verification. It corresponds to steps 200 to 300 in the aforementioned method embodiments and the logical execution body of embodiments 2 and 3.

[0140] In terms of data flow and control logic, this module is activated when a marker indicating the loss of target entity recognition signal appears in the dynamic graph sequence output by the cross-modal semantic alignment module. Its internal trajectory extrapolation model, using the dynamic graph sequence before the loss as a condition, performs a reverse diffusion process with physical prior constraints, generating multiple candidate state trajectory sets that conform to physical laws. Subsequently, as real physical time progresses, this module acquires the latest real multimodal observation data generated by the cross-modal semantic alignment module, calculates the evidence support for each candidate trajectory through spatiotemporal attention queries and anomaly measurement functions, and calculates the physical law violation index based on persistent deviation pattern detection, thus completing the closed-loop reasoning from generative hypotheses to multimodal evidence cross-validation.

[0141] Secondly, the warning judgment and optimization control module built into the processing unit calculates the abnormality metric of the target entity based on the physical law violation index, and generates a warning trigger signal when the abnormality metric meets the preset warning conditions. It also updates the internal parameters of the cross-modal semantic alignment module and the physical trajectory extrapolation and verification module based on the gradient of the backpropagation of the end-to-end loss function.

[0142] Among them, the early warning judgment and optimization control module refers to the third-level core engine within the processing unit, which is responsible for risk decision-making and system self-evolution. It corresponds to step 400 in the aforementioned method embodiment and the logical execution subject in embodiment 4.

[0143] In terms of control logic, this module receives the physical law violation index and evidence support set output by the physical trajectory extrapolation and verification module. Through built-in threshold judgment or nonlinear fusion logic, it calculates the specific value of the anomaly metric. Once this value meets preset warning conditions, such as exceeding a safety threshold or exhibiting a specific surge pattern, the module immediately generates a warning trigger signal and pushes alarm instructions to the digital twin platform front-end and operation and maintenance terminal via the communication unit. Simultaneously, when the system is in the training or incremental fine-tuning phase, this module uses a differentiable distance metric to calculate the loss value between the predicted distribution and the true distribution. Through a gradient backpropagation mechanism, it synchronously distributes the updated gradient to the cross-modal semantic alignment module and the physical trajectory extrapolation and verification module, driving them to perform end-to-end joint parameter updates, thus achieving continuous self-optimization of the system.

[0144] Starting from existing technical problems, current supply chain risk warning systems are generally based on event-driven architecture. This architecture lacks the ability to model the physical world of continuous object movement in advance, and the functional modules are only loosely coupled and cascaded through discrete atomic events, lacking closed-loop reasoning capabilities. Once a single credential signal is lost, the system falls into an evidence vacuum, relying solely on an embedded static model for hard decisions, resulting in high-frequency oscillations in the state of the digital twin and complete inability to detect covert theft behaviors that mimic normal paths. In contrast, this system, through close collaboration between the communication unit and the processing unit, and by constructing a collaborative architecture within the processing unit consisting of three core modules—a cross-modal semantic alignment module, a physical trajectory extrapolation and verification module, and an early warning judgment and optimization control module—completely breaks the atomic processing logic of traditional event-driven architectures. The three modules sequentially complete the entire data flow and control loop from multimodal perception alignment and physical prior-driven generative trajectory extrapolation to multimodal evidence cross-verification and end-to-end optimization, constructing a differentiable world state trajectory extrapolation and evidence verification closed-loop reasoning engine driven by physical laws. This closed-loop engine no longer makes hard decisions during periods of signal absence. Instead, it continuously and smoothly extrapolates the existing state based on physical laws, softening the decision boundaries. This allows it to move beyond relying on the presence or absence of a single signal during evidence verification. Instead, it uses continuous deviation pattern detection to elevate the anomaly perception from signal absence to a violation of physical laws, thus achieving oscillating, continuous state reasoning and accurate detection of hidden risks.

[0145] Through the aforementioned system architecture, the supply chain risk early warning and assessment system based on a digital twin platform in this embodiment not only achieves oscillatory, continuous state reasoning driven by physical laws, completely eliminating the risk calculation failure problem caused by state transitions, but also elevates the level of anomaly detection from signal loss to violation of physical laws through generative trajectory prediction and multimodal evidence cross-validation, greatly enhancing the ability to discover hidden risks. Finally, a fully differentiable end-to-end closed-loop optimization system is obtained, reducing reliance on manual rules and possessing strong scenario adaptability.

[0146] Example 7: like Figure 7 As shown, based on Embodiment 6, this embodiment further de-emphasizes the internal hardware structure of the cross-modal semantic alignment module, providing specific hardware execution entity bindings for functional modules, thus meeting the defense requirements of binary binding between structure and function. It should be understood that this embodiment is merely illustrative and not restrictive; those skilled in the art can make various modifications and variations without departing from the spirit of the invention.

[0147] The cross-modal semantic alignment module specifically includes: a multi-source feature parsing submodule, a graph construction and spatiotemporal coding submodule, and a first data cache for connecting the two.

[0148] Among them, the multi-source feature parsing submodule refers to a set of hardware acceleration units specifically responsible for extracting depth features of each modality in parallel from the original heterogeneous sensor data stream. Its core lies in realizing the synchronous parsing of radio frequency, vision and motion control data through a dedicated processor, rather than relying on the serial polling of a general-purpose processor.

[0149] The graph construction and spatiotemporal coding submodule refers to the computational array responsible for quickly organizing the parsed feature vectors into a graph structure and calculating temporal dependencies. It undertakes the tasks of topology construction and weight update for the transformation from isolated features to structured dynamic graph sequences.

[0150] The first data cache refers to an on-chip storage medium located between two sub-modules, characterized by high bandwidth and low latency. Its design purpose is to eliminate data transfer latency between sub-modules and support tensor-level batch data throughput.

[0151] An RFID baseband signal processor is a hardware circuit dedicated to demodulating and decoding radio frequency signals. It can directly extract low-level information such as tag identification, signal phase and received signal strength indication from the analog signal received by the antenna, and digitize it into confidence joint coding.

[0152] Deep neural network accelerators are dedicated computing chips that optimize instruction sets for convolution operations and matrix multiplication, such as tensor processing units or neural network inference application-specific integrated circuits. They are responsible for high-speed execution of visual feature extraction models and output instance-level re-identification feature vectors.

[0153] An industrial fieldbus interface refers to a hardwired data transceiver controller that conforms to industrial communication protocol standards (such as CAN bus, EtherCAT, or PROFINET) and is used to directly read motion status codes and load status codes reported by forklift or AGV controllers.

[0154] In terms of operation method and data flow path, the multi-source feature analysis submodule includes a parallel-configured RFID baseband signal processor, a deep neural network accelerator, and an industrial fieldbus interface. These are used to synchronously extract phase and confidence joint encoding, instance-level re-identification feature vectors, and motion state encoding from the multimodal data stream, respectively. The three are then projected into entity multimodal feature vectors through a cascaded first multilayer perceptron. These three hardware units adopt a parallel bus architecture in their physical layout. Each unit independently receives the raw data stream from its corresponding sensor and synchronously performs feature extraction. The extracted three heterogeneous feature vectors are then fed into the cascaded first multilayer perceptron.

[0155] The first multilayer perceptron uses a configurable weight matrix to force the alignment of features from different dimensions and project them onto a unified semantic space, outputting dense entity multimodal feature vectors. These vectors are not directly transmitted to the next stage via the network, but are instead written into the first data cache.

[0156] The graph construction and spatiotemporal coding submodule includes a tensor processing unit and a gated loop unit array. The tensor processing unit is used to construct a scene semantic snapshot graph by time-series slices, using entity multimodal feature vectors as node attributes and spatial proximity as edge weights. The gated loop unit array is used to calculate the weights of self-recursive connection edges for the same graph node across time steps, and stack the scene semantic snapshot graphs along the time axis into a dynamic graph sequence.

[0157] The tensor processing unit and the gated loop unit array are interconnected via an internal high-speed bus. They read all entity multimodal feature vectors of the current time slice in batches from the first data cache. Based on spatial coordinates, the tensor processing unit quickly calculates the Euclidean distance between entities and maps it to edge weights, thus completing the topology construction of the semantic snapshot graph of the current slice scene.

[0158] The gated cyclic cell array synchronously reads the node attributes from the cache at the previous moment, calculates the weights of the self-recursive connection edges through internal gate control logic, updates the current node state, and finally writes the updated graph structure back to the cache or directly outputs it as a dynamic graph sequence.

[0159] The first data cache stores intermediate feature maps output by the multi-source feature parsing submodule, enabling tensor-level batch reading by the graph construction and spatiotemporal coding submodules. This cache employs a dual-port static random access memory or a high-bandwidth memory stack architecture, supporting simultaneous continuous writing by the multi-source feature parsing submodule and parallel reading of large tensors by the graph construction submodule. This eliminates the need for intermediate data transfer via the system memory of a general-purpose processor, completely removing the intermediate data handling steps.

[0160] It should be understood that although this embodiment lists an RFID baseband signal processor, a deep neural network accelerator, and an industrial fieldbus interface as specific hardware for parallel parsing, other embodiments may also include a LiDAR point cloud parsing chip or an ultrasonic signal preprocessing circuit, as long as the requirement of parallel and synchronous extraction of multimodal features is met. Similarly, in addition to static random access memory, the first data cache may also adopt a combination of on-chip dynamic random access memory and a direct memory access engine, as long as the low latency requirement for tensor-level batch reads is met.

[0161] Through the aforementioned hardware architecture, this embodiment provides a specific hardware execution entity binding for the cross-modal semantic alignment module, explaining why it is not a software implementation using a general-purpose processor. General-purpose CPUs suffer from latency bottlenecks when processing multi-source heterogeneous data, while the use of parallel hardware accelerators and dedicated high-speed caches enables microsecond-level feature parsing and graph construction, meeting the real-time requirements of digital twins. This embodiment eliminates the time skew in modal parsing through a parallel architecture of an RFID baseband signal processor, a deep neural network accelerator, and an industrial fieldbus interface; it also eliminates intermediate data transfer delays through a tensor-level batch read mechanism in the first data cache, thereby ensuring the real-time continuous output of dynamic graph sequences and providing highly timely structured input for subsequent physical trajectory extrapolation and evidence verification.

[0162] Example 8: like Figure 7 As shown, based on Example 6, this example further degrades the internal hardware structure of the physical trajectory extrapolation and verification module, providing FPGA / ASIC-level hardware binding for the core inference module, and degrading the core algorithm to the chip level, thereby enhancing the visibility of infringement evidence collection and the ability to prevent circumvention.

[0163] It should be understood that this embodiment is illustrative only and not restrictive, and those skilled in the art can make various modifications and variations without departing from the spirit of the invention.

[0164] The physical trajectory extrapolation and verification module is integrated into a field-programmable gate array or application-specific integrated circuit chip, and specifically includes: a conditional diffusion inverse extrapolation submodule, a spatiotemporal cross-validation and anomaly measurement submodule, and a second data cache.

[0165] Among them, Field Programmable Gate Array (FPGA) refers to a semi-custom logic circuit that contains a large number of configurable logic blocks and programmable interconnects. It can be reprogrammed by the user after manufacturing according to the algorithm requirements, realizing the dynamic reconstruction of the parallel computing architecture. It is particularly suitable for carrying diffusion denoising calculations that require frequent iterations and high parallelism.

[0166] Application-Specific Integrated Circuits (ASICs) are non-reconfigurable hard-wired chips designed and manufactured specifically for a particular algorithm. The layout and interconnection of their internal logic gates are fixed during the manufacturing stage, resulting in extremely high computing speed and extremely low power consumption, making them suitable for mature algorithms deployed on a large scale.

[0167] The conditional diffusion reverse extrapolation submodule refers to the computing engine mapped within the chip hardware logic, which is responsible for executing the reverse diffusion sampling process with physical prior constraints. It corresponds to the logic execution entity in steps 200 to 204 of the aforementioned method embodiment.

[0168] The spatiotemporal cross-validation and anomaly measurement submodule refers to the computing engine mapped within the chip hardware logic, which is responsible for performing spatiotemporal attention queries and continuous divergence pattern detection. It corresponds to the logical execution entity in steps 300 to 307 of the aforementioned method embodiment.

[0169] The second data cache refers to an on-chip storage medium located between the two sub-modules, which has extremely high bandwidth and extremely low access latency. Its design purpose is to eliminate bus congestion and latency caused by external memory interaction between sub-modules and to support tensor-level zero-copy data sharing.

[0170] It should be understood that although this embodiment lists FPGA and ASIC as integration carriers, in other embodiments, edge smart chips or system-on-a-chips containing multi-core vector processors can also be used, as long as they can meet the requirements of providing hardware-level parallel computing and low-latency data sharing.

[0171] In terms of operation methods and data flow paths, the integration of submodules in FPGAs or ASICs presents a highly parallel internal architecture. The conditional diffusion inverse extrapolation submodule includes a backbone denoising network and a parallel-connected differentiable physics simulator. The backbone denoising network is used to predict the initial denoising direction based on the input noise and the conditional dynamic graph sequence. The differentiable physics simulator is used to generate physical gradient correction terms based on Newton's laws of motion, which guide the backbone denoising network to reconstruct a set of candidate state trajectories that meet physical constraints.

[0172] In terms of hardware layout, the backbone denoising network is mapped as a set of cascaded matrix multiplication acceleration units and activation function lookup table arrays within an FPGA or ASIC, responsible for high-speed execution of cross-attention calculations and denoising direction predictions in the U-Net or Transformer architecture. The differentiable physics simulator is mapped as a set of fixed-point / floating-point hybrid arithmetic pipelines and rigid body dynamics solution microcode engines connected in parallel with it, responsible for synchronously receiving the predicted raw physical states output by the backbone network and, within the same clock cycle, calculating the physically valid state and physical gradient correction terms according to Newton's second law.

[0173] This parallel-connected hardware architecture enables data-driven denoising direction prediction and physics-equation-driven gradient correction to be completed synchronously within the same denoising time step, without sequential waiting. Subsequently, the two are fused and the current noise state is updated through a hardware addition tree. After iterative sampling, the high-order tensor data of the generated candidate state trajectory set is not output through the chip's external bus, but is directly written to the second data cache.

[0174] The spatiotemporal cross-validation and anomaly measurement submodule extracts predicted sensor observation features from the candidate state trajectory set and extracts real multimodal feature vectors from real multimodal observation data using a spatiotemporal attention query approach. It outputs evidence support through a built-in similarity calculation engine and calculates a physical law violation index based on persistent deviation detection logic. This submodule is internally configured with a dedicated spatiotemporal attention query hardware accelerator and a cosine similarity calculation array. It directly reads candidate trajectory tensors from the second data cache via an internal high-speed interconnect bus and simultaneously receives real multimodal observation data from the cross-modal semantic alignment module. It calculates single-point evidence support and comprehensive evidence support at each time point in parallel on-chip and quickly detects persistent deviation patterns through a built-in threshold comparison register and window counter, ultimately outputting the physical law violation index.

[0175] The second data cache is used to share tensor data of the candidate state trajectory set between the conditional diffusion inverse extrapolation submodule and the spatiotemporal cross-validation and anomaly measurement submodule. This cache adopts a dual-port static random access memory or a high-bandwidth memory stack architecture, supporting simultaneous continuous writing by the conditional diffusion inverse extrapolation submodule and parallel reading of large blocks of tensors by the spatiotemporal cross-validation and anomaly measurement submodule, realizing zero-latency transfer of candidate trajectory data between the two core inference engines.

[0176] Through the aforementioned hardware architecture, this embodiment provides FPGA / ASIC-level hardware bonding for the physical trajectory extrapolation and verification module, explaining why the core inference cannot be placed in the cloud. Performing trajectory extrapolation in the cloud suffers from communication latency and makes it difficult to guarantee real-time performance. Integrating the core inference engine into an edge FPGA or ASIC leverages hardware parallelism to accelerate diffusion denoising and physical correction iterations, achieving millisecond-level risk response. Simultaneously, the parallel connection architecture between the backbone denoising network and the differentiable physical simulator eliminates the computational bottleneck of serial waiting; the tensor data sharing mechanism of the second data cache eliminates data transfer latency between extrapolation and verification, thus ensuring a rapid closed loop at the edge for the entire continuous inference defense line from signal loss to triggering an early warning, greatly enhancing the system's real-time blocking capability against high-speed data transfer and covert theft.

[0177] Example 9: Based on Examples 1 to 8, this example combines a specific port distribution center application scenario to demonstrate the operational effect of the technical solution of the present invention in a real physical environment, mapping the abstract algorithm logic to concrete physical entities and business processes, and intuitively presenting the beneficial effects of eliminating state oscillations and detecting hidden theft.

[0178] It should be understood that this embodiment is illustrative only and not restrictive. Those skilled in the art can apply this solution to other similar complex obstruction scenarios such as customs-supervised warehouses and large cold chain warehouses without departing from the spirit of the invention.

[0179] The application scenario is set as a large port container yard, with high-value cargo pallets being tracked. Heterogeneous sensors deployed in this scenario include RFID readers, surveillance cameras, and forklift controllers. During normal operation, forklifts move pallets within the yard. RFID readers continuously read the phase and confidence levels of pallet tags, cameras capture instance-level re-identification feature vectors of the pallets and forklifts, and the forklift controller reports its odometer readings and forklift load status codes. The system uses a cross-modal semantic alignment module to parse and map these heterogeneous data streams into a unified semantic space, constructing a dynamic graph sequence containing the interaction relationships between the forklift and the pallet.

[0180] Step 9001, Signal Loss and Dynamic Graph Sequence Condition Input: When a forklift picks up a pallet and drives into a densely packed racking area or between metal containers, the RFID radio frequency signal and the visual recognition signal are simultaneously lost at high frequency due to multipath reflection and physical obstruction effects. At this time, the system no longer triggers atomic events of departure or destruction immediately as in existing technologies. Instead, in response to the loss of the target pallet's identification signal, it calls the dynamic graph sequence corresponding to the time window before the loss as a condition. This dynamic graph sequence contains the interaction graph edges at the moment the pallet is picked up by the forklift. These edges characterize the forklift's handling relationship with the pallet with extremely high spatial proximity and interaction strength.

[0181] Step 9002, Extrapolation of candidate trajectories under physical prior constraints: The system inputs the above dynamic graph sequence containing interaction relationships into the trajectory extrapolation model to generate a set of candidate state trajectories for the target tray.

[0182] Specifically, the conditional diffusion back extrapolation submodule uses the dynamic graph sequence as a condition to perform a back diffusion process starting from random noise. Within each denoising time step, the backbone denoising network predicts the initial denoising direction and the original physical state, which includes the pallet's position, velocity, and acceleration. Simultaneously, a parallel-connected differentiable physics simulator receives this predicted original physical state and extracts the interaction force vectors of neighboring forklift entities on the pallet from the dynamic graph sequence. According to Newton's second law, the simulator inputs the pallet's preset mass, interaction force vectors, and environmental gravity vectors into the rigid body dynamics solver, calculates the corrected acceleration, and obtains a physically reasonable state that satisfies mechanical constraints through numerical integration. This generates a physical gradient correction term to guide the backbone denoising network.

[0183] This mechanism forces all trajectory samples generated by the model to satisfy inertial constraints, thereby extrapolating multiple candidate trajectories that conform to physical laws and indicate that the forklift will continue to move the pallet. For example, trajectory hypothesis A is that the forklift moves straight along the main aisle, and trajectory hypothesis B is that the forklift turns right and enters the temporary storage area. These candidate trajectories are all kinematically reasonable, completely suppressing erroneous inferences such as "the pallet disappears instantly" or "remains stationary."

[0184] Step 9003, Spatiotemporal attention query and evidence support calculation: As real physical time progresses, the system acquires real multimodal observation data corresponding to the candidate state trajectory set in spatiotemporal time, and performs cross-validation with the predicted sensor observation features in the candidate state trajectory set.

[0185] For trajectory hypothesis A, the system extracts its predicted spatial location and predicted visual feature vector at a future time point. Using this predicted location as the center, it performs a spatiotemporal attention query on real multimodal observation data to extract the corresponding real multimodal feature vector at that spatiotemporal point. However, since the pallet was actually transported out along an abnormal path by a disguised theft vehicle, the real camera did not capture the visual features of the forklift and pallet at the predicted passage location of hypothesis A, and the real RFID reader did not read the tag signal. The system calculates the multimodal similarity between the predicted and real features and finds that trajectory hypothesis A has extremely low single-point evidence support. Similarly, trajectory hypothesis B and all other conventional handling hypotheses that conform to physical laws lack visual and RFID evidence support.

[0186] Step 9004, Continuous Deviation Detection and Physical Law Violation Index Determination: The system aggregates the single-point evidence support at each time point along the time dimension and finds that the comprehensive evidence support of all candidate state trajectories is lower than the preset normal threshold in multiple consecutive time windows, showing a continuous deviation pattern.

[0187] This persistent deviation is not a transient sensor noise fluctuation, but rather a cumulative state characterized by the continuous falsification of all physically plausible conventional transport assumptions by multimodal facts. When all candidate trajectories exhibit this pattern, the system calculates a physical law violation index based on the minimum value of the comprehensive evidence support and the number of windows through which the persistent deviation has persisted. This index rises sharply, directly elevating the anomaly detection level from "signal missing" to "physical law violation," indicating that only an event defying conventional physical logic—"transported by an abnormal vehicle along an abnormal path"—can explain the current observational vacuum.

[0188] Step 9005, Warning Trigger: The warning judgment and optimization control module calculates the abnormal measurement index of the target pallet based on the violation index of physical laws. When the index meets the preset warning conditions, a supply chain risk warning is immediately triggered. The abnormal state and extrapolation trajectory of the pallet are highlighted on the front end of the digital twin platform, and a hidden theft alarm is pushed to the operation and maintenance personnel.

[0189] The specific demonstration of the port scenario above visually demonstrates the beneficial effects of this solution in eliminating state oscillations and detecting covert theft. Existing technologies, when both RFID and visual signals are lost simultaneously, cause the digital twin's state to oscillate and disappear at high frequencies, the dwell timer repeatedly resets to zero, rendering the alarm ineffective, and it is completely unable to identify covert theft mimicking normal paths. In contrast, this invention, through extrapolation of physical prior constraints, maintains continuous state reasoning conforming to the laws of mechanics during signal vacuum periods, completely eliminating state jumps; through cross-validation of multimodal evidence and continuous deviation detection, it accurately identifies covert risks that appear to have coherent signals but exhibit abnormal physical logic, achieving proactive detection of new types of theft beyond the established rules.

[0190] Comparative Example 1: To more clearly illustrate the inventiveness of this invention, the same port container yard scenario and signal loss conditions as in Example 9 are set up below to compare and analyze the typical system behavior of existing technologies based on atomic events and hard decisions using a single RFID signal. It should be understood that this comparative example is merely illustrative and not restrictive, intended to demonstrate through a negative example that the lack of physical prior modeling and multimodal cross-validation inevitably leads to state oscillations and risky failures, thereby highlighting the beneficial effects of this invention.

[0191] The application scenario is set as a large port container yard, with the tracked entity being high-value cargo pallets, and the deployed sensors being only RFID readers. When a forklift picks up a pallet and drives into a densely packed racking area, causing the RFID signal to be lost due to multipath reflection and obstruction, the behavior logic and data flow of the existing technology system are as follows: Step SC01: Atomized Event Triggering and Disappearance of State Oscillation. Existing technology systems are based on an event-driven architecture, treating each sensor report as an isolated, decisive atomic event, following a hard-decision logic of "existence upon reading, disappearance upon not reading." When the RFID signal is lost, the system immediately triggers an "away" or "destroy" event, the digital twin's state oscillation disappears, and the corresponding cargo pallet icon vanishes instantly from the 3D monitoring interface. This approach completely abandons the continuous context of the object's own movement and existence, lacks fault tolerance for transient sensing noise, and leads to atomized state processing and a disconnect from the context.

[0192] Step SC02, brief reread and state oscillation occur: Due to the multipath reflection effect of radio frequency signals between metal containers, the RFID reader may recapture the weak reflected signal of the pallet tag after a short time interval. At this time, the existing technology system immediately triggers the "entry" or "create" event, and the previously disappeared digital twin's state oscillation reappears, with the pallet icon flashing again on the interface. This high-frequency repeated transition of "successful identification - complete loss - successful identification again" directly triggers a lifecycle event storm for the corresponding pallet's twin within the digital twin platform.

[0193] Step SC03, Repeated Zeroing of the Dwell Timer and Alarm Failure: The existing system relies on a long-duration dwell timer based on state counts to trigger risk warnings. However, due to frequent oscillations in the state of the digital twin, the dwell timer is reset to zero whenever an "entry" event is triggered, failing to accumulate to the set alarm threshold (e.g., dwell time exceeding tens of minutes). The root cause of this failure in risk indicator calculation lies in the strong reliance on a single identifier, which plunges the system into a complete evidence vacuum when the signal is lost. It can only rely on the simplest embedded static model for reasoning, i.e., the default assumption that "the object disappears in the same spot after its last appearance." This erroneous prior model contradicts the reality of continuous motion of objects in the physical world, leading the system to make continuous erroneous state inferences and risk judgments.

[0194] Step SC04, Blind Spot in Identifying Covert Theft: When covert theft occurs, mimicking normal paths—for example, a stolen vehicle transporting a pallet along an unusual route—existing systems are completely unable to identify it. This is because the RFID signal is also lost during the theft; the system cannot distinguish between "signal loss due to normal obstruction" and "signal loss due to abnormal theft" based solely on the presence or absence of a single signal. Even if the stolen vehicle briefly passes through a reader's coverage area, causing the signal to reappear momentarily, the system will only misinterpret it as normal flow and reset the timer, failing to elevate the level of anomaly detection from "signal loss" to "violation of physical laws," thus creating a significant blind spot in identifying covert risks.

[0195] The above comparative examples clearly demonstrate that the inherent flaw of existing technologies—the "read-is-existent, not-read-is-disappeared" hard-decision approach—stems essentially from a lack of prior modeling capabilities for the physical world of continuous object motion and a lack of multimodal cross-validation mechanisms to fill the evidentiary vacuum when a single information source fails. This atomized event-based hard-decision inevitably leads to state oscillations and reasoning errors, ultimately causing the complete failure of risk warning mechanisms. This invention, by introducing physical prior extrapolation and multimodal cross-validation, continuously and smoothly extrapolates the existence state based on physical laws during signal vacuum periods, softening the decision boundary. Furthermore, through continuous deviation pattern detection, it elevates the anomaly perception from signal absence to violation of physical laws, fundamentally solving the crux of existing technologies and demonstrating ingenuity.

[0196] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention, such as adjustments to the trajectory extrapolation model architecture, equivalent replacements for the anomaly metric function calculation logic, or different integration methods for the hardware execution entity, should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A supply chain risk early warning and assessment method based on a digital twin platform, characterized in that, include: Acquire multimodal sensor data and construct a dynamic graph sequence, which is used to characterize the physical entities and their interactions within the detected supply chain scenario; In response to the loss of the identification signal of the target entity, the dynamic graph sequence corresponding to the time window before the loss is called as a condition and input into the trajectory extrapolation model to generate a set of candidate state trajectories of the target entity; the trajectory extrapolation model is configured as a generative model that generates future state sequences based on historical states, and physical prior constraints are introduced in the generation process to correct the inference direction. Acquire real multimodal observation data corresponding to the candidate state trajectory set in time and space, and cross-validate it with the predicted sensor observation features in the candidate state trajectory set. Calculate the evidence support degree through an anomaly metric function, and determine the physical law violation index based on the degree of continuous deviation of the evidence support degree. Based on the evidence support level, calculate the anomaly measurement index of the target entity, and trigger a supply chain risk warning when the anomaly measurement index meets the preset warning conditions.

2. The supply chain risk early warning and assessment method based on a digital twin platform according to claim 1, characterized in that, The trajectory extrapolation model is a conditional diffusion model, including a backbone denoising network and a parallel differentiable physics simulator. The process of generating a candidate state trajectory set for the target entity specifically includes: Using the dynamic graph sequence corresponding to the time window before the loss as a condition, a reverse diffusion process is executed starting from random noise; Within each denoising time step, the backbone denoising network predicts the initial denoising direction based on the current noise state and the conditions, and synchronously outputs the predicted original physical state corresponding to the current noise state. The predicted original physical state includes the position, velocity and acceleration of the target entity. The differentiable physics simulator receives the predicted original physical state, calculates the corresponding reasonable physical state according to Newton's laws of motion, and generates a physical gradient correction term for correcting the initial denoising direction. After adjusting the initial denoising direction with the physical gradient correction term, the current noise state is updated, and the next denoising time step is entered. The iteration continues until at least one future state trajectory that meets the physical constraints is restored, forming the candidate state trajectory set.

3. The supply chain risk early warning and assessment method based on a digital twin platform according to claim 2, characterized in that, The process by which the differentiable physics simulator calculates the physical gradient correction term specifically includes: The current position, velocity, and acceleration of the target entity are parsed from the predicted original physical state, and the interaction force vectors of neighboring entities on the target entity are extracted from the dynamic graph sequence. The preset mass of the target entity, the interaction force vector, and the environmental force vector are input into the built-in rigid body dynamics solver. The corrected acceleration is calculated according to Newton's second law, and the physically reasonable state that satisfies the mechanical constraints is obtained through numerical integration. Calculate the element-wise physical residual between the physical reasonable state and the predicted original physical state, and map the physical residual to the noise prediction space through an encoding network to generate the physical gradient correction term.

4. The supply chain risk early warning and assessment method based on a digital twin platform according to claim 1, characterized in that, The process of calculating the evidence support using the anomaly metric function specifically includes: From the predicted state sequence of each candidate state trajectory in the candidate state trajectory set, the predicted sensor observation features at future time points are extracted. The predicted sensor observation features include at least the predicted spatial location and the predicted visual feature vector. Using the future time point as an index, a spatiotemporal attention query is performed from the real multimodal observation data, centered on the predicted spatial location, to extract the real multimodal feature vector of the corresponding spatiotemporal point; For each future time point, the multimodal similarity between the predicted sensor observation features and the real multimodal feature vector is calculated as the single-point evidence support for that time point; The single-point evidence support of each future time point is aggregated along the time dimension to obtain the comprehensive evidence support of the candidate state trajectory. The comprehensive evidence support of each candidate state trajectory is used to form an evidence support set, which is then used as the output of the anomaly measurement function.

5. The supply chain risk early warning and assessment method based on a digital twin platform according to claim 4, characterized in that, The process of determining the physical law violation index based on the degree of persistent deviation from the level of evidence support specifically includes: Obtain the set of evidence support for the candidate state trajectory within the current and historical continuous time windows; For each candidate state trajectory, a persistent divergence pattern is detected from the evidence support set. The persistent divergence pattern represents the cumulative state in which the evidence support of the candidate state trajectory is lower than a preset normal threshold within multiple consecutive time windows. When the evidence support set of all candidate state trajectories presents the persistent deviation pattern, the physical law violation index is calculated based on the minimum of the comprehensive evidence support of each candidate state trajectory and the number of windows in which the persistent deviation pattern has continued.

6. The supply chain risk early warning and assessment method based on a digital twin platform according to claim 1, characterized in that, It also includes end-to-end optimization steps: Differentiable distance metric is used as a loss function to measure the distribution difference between the predicted distribution and the true distribution. The predicted distribution is composed of the candidate state trajectory set, and the true distribution is composed of the true multimodal observation data. The differentiable distance metric includes at least bulldozer distance or maximum mean difference. Based on the distribution differences, the loss value is calculated, and the parameters of the trajectory extrapolation model and the anomaly metric function are updated simultaneously through gradient backpropagation to achieve end-to-end joint optimization between the trajectory extrapolation model and the anomaly metric function.

7. The supply chain risk early warning and assessment method based on a digital twin platform according to claim 1, characterized in that, The process of acquiring multimodal sensor data and constructing a dynamic graph sequence specifically includes: Based on the heterogeneous sensors deployed in the supply chain scenario, multimodal data streams are acquired, and the multimodal data streams are parsed and mapped to a unified semantic space to obtain the entity multimodal feature vectors of the tracked entity in each time slice. Each tracked entity in the scene is taken as a graph node, and the multimodal feature vector of the entity is the node attribute of the corresponding graph node. Based on the spatial proximity between entities in the same time slice, graph edges representing the interaction strength between entities are calculated and established, and a scene semantic snapshot graph is constructed for each time slice. The scene semantic snapshot image is stacked vertically along the time axis for consecutive time slices, and self-recursive connection edges are introduced between the same graph nodes of adjacent time slices to form the dynamic graph sequence. The weight of the self-recursive connection edge is calculated by the gated loop unit based on the node attributes of adjacent time steps to form the hidden state, and the hidden state is mapped to scalar weight values ​​through a fully connected layer and a nonlinear activation function to encode the temporal evolution features of the entity's own state.

8. A supply chain risk early warning and assessment system based on a digital twin platform, characterized in that: The supply chain risk early warning and assessment system, which is applied to the digital twin platform-based supply chain risk early warning and assessment method according to any one of claims 1-7, specifically includes: a communication unit and a processing unit; The communication unit is used to receive multimodal data streams collected by multiple heterogeneous sensors deployed in the supply chain scenario, and transmit the multimodal data streams to the processing unit; The processing unit parses and maps the multimodal data stream to a unified semantic space through a cross-modal semantic alignment module built into the processing unit, and constructs a dynamic graph sequence to represent each physical entity and its interaction relationship. In response to the loss of the identification signal of the target entity, the physical trajectory extrapolation and verification module built into the processing unit calls the dynamic graph sequence corresponding to the time window before the loss as a condition to generate a set of candidate state trajectories of the target entity, and obtains real multimodal observation data corresponding to the spatiotemporal of the candidate state trajectory set for cross-validation, and calculates the physical law violation index. The warning judgment and optimization control module built into the processing unit calculates the abnormality metric of the target entity according to the physical law violation index, and generates a warning trigger signal when the abnormality metric meets the preset warning conditions. The internal parameters of the cross-modal semantic alignment module and the physical trajectory extrapolation and verification module are updated according to the gradient of the backpropagation of the end-to-end loss function.

9. The supply chain risk early warning and assessment system based on a digital twin platform according to claim 8, characterized in that, The cross-modal semantic alignment module specifically includes: a multi-source feature parsing submodule, a graph construction and spatiotemporal coding submodule, and a first data cache for connecting the two; The multi-source feature analysis submodule includes a parallel RFID baseband signal processor, a deep neural network accelerator, and an industrial fieldbus interface, which are used to synchronously extract phase and confidence joint encoding, instance-level re-identification feature vector, and motion state encoding from the multimodal data stream, respectively, and project the three into entity multimodal feature vectors through a cascaded first multilayer perceptron; The graph construction and spatiotemporal coding submodule includes a tensor processing unit and a gated recurrent unit array. The tensor processing unit is used to construct a scene semantic snapshot graph by time-series slices using the entity multimodal feature vector as node attributes and spatial proximity as edge weights. The gated recurrent unit array is used to map the hidden state of the same graph node across time steps to scalar weight values ​​through a fully connected layer and a nonlinear activation function, which are used as weights for self-recursive connection edges. The scene semantic snapshot graphs are then stacked along the time axis to form the dynamic graph sequence. The first data cache is used to store the intermediate feature maps output by the multi-source feature parsing submodule, for tensor-level batch reading by the graph construction and spatiotemporal coding submodule.

10. The supply chain risk early warning and assessment system based on a digital twin platform according to claim 8, characterized in that, The physical trajectory extrapolation and verification module is integrated into a field-programmable gate array or an application-specific integrated circuit chip, and specifically includes: a conditional diffusion inverse extrapolation submodule, a spatiotemporal cross-validation and anomaly measurement submodule, and a second data cache; The conditional diffusion inverse extrapolation submodule includes a backbone denoising network and a parallel-connected differentiable physics simulator. The backbone denoising network is used to predict the initial denoising direction based on the input noise and the conditional dynamic graph sequence. The differentiable physics simulator is used to generate a physical gradient correction term based on Newton's laws of motion, which guides the backbone denoising network to restore the candidate state trajectory set that meets the physical constraints. The spatiotemporal cross-validation and anomaly measurement submodule is used to extract predicted sensor observation features from the candidate state trajectory set, extract real multimodal feature vectors from the real multimodal observation data using a spatiotemporal attention query method, output evidence support through the built-in similarity calculation engine, and calculate the physical law violation index based on the continuous deviation detection logic. The second data cache is used to share the tensor data of the candidate state trajectory set between the conditional diffusion inverse extrapolation submodule and the spatiotemporal cross-validation and anomaly measurement submodule.

Images