A method and system for a consumer product anti-counterfeiting inspection big data model combining artificial intelligence and behavior graph

Abstract

The application discloses a kind of fusion artificial intelligence and behavior atlas's consumer goods anti-fake inspection big data model method and system, belong to data processing technical field, this method is by obtaining production, logistics, sales, consumption and so on multi-source behavior data, constructs time-space coupling's commodity circulation hypergraph, and adopts hyperbolic space embedding and is hierarchized representation;Based on hypergraph structure, by differentiable inductive logic programming automatic learning explainable first-order logic inspection rule, construct the evolving rule base;The inspection process is modeled as the partially observable stochastic game with the counterfeiter, and the active inspection decision instruction is generated by multi-agent deep reinforcement learning;Inspection result is fed back to rule base, realizes the online incremental update of rule and strategy;The application realizes the active identification and dynamic confrontation to complex gang counterfeiting mode, improves the intelligent level and explainability of anti-fake inspection.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for a big data model for anti-counterfeiting inspection of consumer products that integrates artificial intelligence and behavioral graphs. Background Technology

[0002] With the globalization of the commodity circulation system and the rapid development of e-commerce, the inspection and supervision of consumer goods anti-counterfeiting is facing increasingly severe challenges. Traditional anti-counterfeiting technologies mainly rely on physical anti-counterfeiting labels, query codes and other means. However, the production and sale of counterfeit goods are increasingly characterized by gangs, cross-regional operations and high technology. Simply relying on physical anti-counterfeiting or consumers' active verification is no longer enough to effectively curb the circulation of counterfeit goods.

[0003] In recent years, artificial intelligence technology has been gradually applied to the field of anti-counterfeiting. Among the existing technologies, there are already methods for distinguishing authenticity based on image recognition. For example, deep learning models are used to extract the appearance features of goods and perform cluster analysis to distinguish the authenticity of goods. However, these methods only focus on the physical characteristics of the goods themselves and lack the fusion analysis of multi-source behavioral data during the circulation of goods, making it difficult to identify group counterfeiting patterns involving complex relationships. On the other hand, anti-counterfeiting technology based on behavior analysis has gradually attracted attention. Existing technologies have disclosed anomaly identification methods based on dynamic behavior graphs. By constructing behavior graphs and monitoring new behavior flow data to update the topology, the identification of fake information on the network can be achieved. However, these methods are mainly aimed at the field of information dissemination and have not yet involved the fusion modeling of behavior data from multiple links such as production, logistics, sales and consumption in the field of commodity circulation. Moreover, their graph structure is mainly based on ordinary graphs, which are difficult to express the complex relationship between multiple commodities and multiple entities. In terms of rule learning, existing technologies have limitations of both symbolic and neural methods. Symbolic inductive logic programming (ILP) models can learn first-order logic rules in a data-efficient way, but they lack robustness to noisy data and are difficult to adapt to incomplete or inconsistent data in real-world scenarios. Neural symbolic ILP models use neural networks to learn logic programs in a differentiable way, which improves the robustness of the model. However, most existing methods require strong language bias to learn logic programs, which reduces the usability and flexibility of the model and makes it difficult to apply to both small-scale datasets and large-scale knowledge graph scenarios. In terms of game theory decision-making, existing technologies disclose anomaly detection methods based on product lineage, which predict abnormal states by constructing a product lineage table and using an autoencoder. However, such methods belong to offline batch processing mode, which cannot respond to the dynamic evolution of counterfeiting methods in real time, and are based only on contact point records within the supply chain, lacking proactive estimation and game theory confrontation of counterfeiters' strategies. In addition, existing anti-counterfeiting systems generally suffer from data silos, making it difficult to share data between different brands and channels. This makes it difficult to detect cross-brand and cross-category counterfeiting patterns in a timely manner. At the same time, the black-box nature of deep learning models makes inspection decisions lack interpretability, making it difficult to meet the requirements of administrative law enforcement for traceable and interpretable judgment basis. In summary, existing technologies have the following shortcomings: They rely on a single data source, lacking integrated modeling of multi-source behavioral data such as production, logistics, sales, and consumption; their graph structures are simple, making it difficult to express complex relationships between multiple entities; rule learning depends on manual pre-setting or strong language bias, lacking automatic evolution capabilities; their auditing model is passively reactive, lacking proactive game-playing against counterfeiters' strategies; and their decision-making process is opaque, failing to meet interpretability requirements. Therefore, there is an urgent need for an anti-counterfeiting auditing method that can integrate multi-source behavioral data, construct complex relationship graphs, and achieve automatic rule learning and proactive game-playing decision-making. Summary of the Invention

[0004] To overcome the problems of the prior art, this invention discloses a method and system for a big data model for anti-counterfeiting inspection of consumer products that integrates artificial intelligence and behavioral graphs.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A big data model method for anti-counterfeiting inspection of consumer products that integrates artificial intelligence and behavioral graphs includes the following steps: S1, acquire multi-source behavioral data of the target consumer product throughout its entire supply chain lifecycle. This multi-source behavioral data includes production data, logistics data, sales data, and consumer interaction data. S2, based on multi-source behavioral data, constructs a spatiotemporally coupled commodity circulation hypergraph with commodities, subjects, and events as nodes and behavioral relationships as edges; Among them, the hypergraph uses hyperbolic space embedding to represent nodes and hyperedges, capturing the hierarchical structure and complex association patterns of commodity circulation behavior; S3, based on the commodity circulation hypergraph, automatically learns first-order logic form audit rules from historical audit cases through differentiable inductive logic programming, and builds an evolvable rule base; Among them, the rule learning process uses hyperbolic space embedding as input features to realize the conditional awareness and dynamic adjustment of rules; S4 models the inspection process as a partially observable stochastic game between the inspector and the counterfeiter. Based on the current market state features and rule base extracted from the commodity circulation hypergraph, the game strategy is solved through multi-agent deep reinforcement learning, generating proactive inspection decision instructions for different commodity categories, circulation channels or regions. Among them, the multi-agent system includes multiple collaborative agents responsible for channel merchant behavior analysis, consumer verification behavior analysis, and logistics trajectory analysis, respectively. S5 responds to proactive audit decision instructions, performs audit operations on target products, and feeds back the audit results to the rule base to update the evolvable rule base.

[0006] Preferably, step S2 further includes: assigning a trusted timestamp and a trusted geographic location code to the multi-source behavioral data, wherein the trusted timestamp is generated based on the standard time of the National Time Service Center, and the trusted geographic location code is generated based on the positioning data of the BeiDou Navigation Satellite System or the Global Positioning System; By using trusted timestamps and trusted geolocation codes as spatiotemporal attributes of event nodes in the commodity circulation hypergraph, a hypergraph structure with a spatiotemporal trusted foundation is constructed.

[0007] Preferably, the learning of audit rules through differentiable inductive logic programming in step S3 includes: Deep neural networks are used to extract deep features from multi-source behavioral data. By combining deep features with domain prior knowledge through differentiable inductive logic programming, first-order logic rules with confidence weights are learned. The rule base is incrementally updated in response to new audit cases or newly detected fraud patterns.

[0008] Preferably, solving the game strategy in step S4 includes: Based on the commodity circulation hypergraph, feature vectors of the current market status are extracted. The feature vectors include circulation distribution characteristics of different categories of commodities, statistical characteristics of historical audit results, and emotional characteristics of consumer feedback. By using generative adversarial imitation learning, the strategy space and behavioral preferences of fraudsters are estimated based on historical fraud case data; A graph theory-based counterfactual regret minimization algorithm is used to solve for the approximate Nash equilibrium strategy of partially observable stochastic games. Based on an approximate Nash equilibrium strategy, proactive audit instructions are generated, which include key audit targets, audit channels, and audit timing.

[0009] Preferably, the construction of the spatiotemporally coupled commodity circulation hypergraph in step S2 further includes: Perceptual hash fingerprints are extracted from different modalities of multi-source behavioral data, including product images, RFID signals, logistics codes, and consumer scanning behavior sequences. The perceptual hash fingerprint is used as the original feature vector of the product node, and then mapped to a low-dimensional hyperbolic representation through hyperbolic space embedding.

[0010] Preferably, the multi-source behavioral data includes: The scanning behavior data generated when consumers scan the code to check the authenticity of products using mobile terminals includes scanning time, scanning location, scanning frequency, and scanning result feedback. Step S2, which involves constructing a spatiotemporally coupled commodity circulation hypergraph, also includes: Consumer scanning behavior data is used as consumer-end interaction event nodes, and hyper-edge connections are established with corresponding product nodes and sales channel nodes to form a holographic map of commodity circulation that includes consumer behavior dimensions.

[0011] Preferably, the multi-source behavioral data obtained in step S1 further includes: Process parameter data and quality inspection data of the product during the production process, as well as environmental monitoring data of the product during the distribution process; In step S2, process parameter data and environmental monitoring data are used as attribute features of event nodes and participate in the hypergraph construction.

[0012] Preferably, the rules in the evolvable rule base constructed in step S3 are stored in an interpretable first-order logic form and are associated with rule application conditions, rule confidence, and rule lifecycle information; The rule base supports rule transfer learning across brands and product categories.

[0013] Preferably, after the audit results are fed back to the rule base in step S5, steps S3 and S4 are also re-executed to achieve online updates of the rule base and game strategy; Online updates employ an incremental learning approach, updating model parameters only based on newly added data.

[0014] On the other hand, a big data model system for anti-counterfeiting inspection of consumer products that integrates artificial intelligence and behavioral graphs is provided, configured to execute the above method.

[0015] The beneficial effects of this invention are as follows: Compared with the prior art, the technical solution provided by the present invention has the following significant advantages: A spatiotemporally coupled commodity circulation hypergraph is constructed, which enhances the ability to express complex relationship patterns. This invention unifies the modeling of multi-source behavioral data such as production, logistics, sales, and consumption into a hypergraph structure with commodities, subjects, and events as nodes and behavioral relationships as edges. By connecting two or more nodes through hyperedges, it can effectively represent the group fraud behavior pattern involving multiple commodities and multiple subjects, overcoming the limitation of traditional ordinary graph structures in expressing complex relationships. By introducing hyperbolic space embedding, the ability to capture the hierarchical structure of commodity circulation is enhanced. This invention uses hyperbolic space to represent hypergraph nodes and hyperedges. By utilizing the tree-like structure characteristics of hyperbolic geometry, it naturally adapts to the brand-category-single product hierarchical structure and power-law distribution characteristics in commodity circulation behavior, providing more expressive input features for subsequent rule learning. This invention enables automatic learning of rules for differentiable inductive logic programming, balancing accuracy and interpretability. Through the ∂ILP framework, the invention automatically learns first-order logic forms of inspection rules from historical inspection cases. The rules are stored as explicit logical expressions with confidence weights, ensuring interpretability to meet the requirements of administrative law enforcement evidence, while also possessing automatic evolution capabilities to adapt to the iteration of fraud methods. By establishing a dynamic evolution model of offensive and defensive game, the invention achieves a leap from passive response to active deterrence. The invention models the inspection process as a partially observable stochastic game between the inspector and the counterfeiter. It estimates the counterfeiter's strategy by generating adversarial imitation learning, and uses a graph theory-based counterfactual regret minimization algorithm to solve the approximate Nash equilibrium strategy. This generates forward-looking active inspection instructions that can guide counterfeiters to abandon the optimal counterfeiting target. By constructing a holographic map that includes consumer behavior dimensions and expanding the perception range of behavior analysis, this invention incorporates consumer scanning behavior data as an independent consumer-end interaction event node into the hypergraph and establishes hyperedge connections with product nodes and sales channel nodes, making consumer feedback an important signal for rule learning and game decision-making, thus making up for the shortcomings of traditional anti-counterfeiting systems that ignore consumer behavior. By introducing cross-modal perceptual hash fingerprints, a unified representation and efficient retrieval of multi-source heterogeneous data are achieved. This invention extracts perceptual hash fingerprints from product images, radio frequency identification signals, logistics codes, and consumer scanning behavior sequences, and maps them into low-dimensional representations through hyperbolic space embedding. While preserving multimodal information, it supports rapid recall and comparison of large-scale products. Supporting cross-brand and cross-category transfer learning of rule bases reduces the cost of deployment in new fields. This invention enables the rule knowledge of existing brands or categories to be quickly transferred to new brands or categories through rule universality assessment and adaptation migration mechanism, significantly reducing the dependence on annotation data in new fields and accelerating the replication and promotion of anti-counterfeiting capabilities. Incremental learning is used to update rules and strategies online, ensuring the system's real-time response capability. After the audit results are fed back, this invention only makes incremental adjustments to the rule confidence, predicate prediction network and game strategy based on the new data, without the need for full retraining. This enables the system to keep up with the evolution of fraud methods and maintain a high efficiency in auditing. By introducing a trusted spatiotemporal benchmark, the anti-tampering capability of behavioral data is enhanced. This invention assigns a trusted timestamp based on the National Time Service Center and a trusted geographic location code based on BeiDou / GPS positioning to multi-source behavioral data as spatiotemporal attributes of event nodes, eliminating the possibility of data forgery and tampering from the source and providing a high-confidence data foundation for subsequent analysis. By integrating production quality and environmental monitoring data, the dimensions of anomaly detection are enriched. This invention incorporates process parameters, quality inspection data, and environmental monitoring data from the distribution process into event node attributes, enabling SuperGraph to perceive changes in the internal quality of goods and the external environment, providing more comprehensive data support for refined auditing. In summary, this invention, through technological innovations such as spatiotemporal hypergraph construction, hyperbolic space embedding, neural symbol rule learning, attack-defense game evolution, and incremental updates, constructs a proactive, interpretable, adaptive, and transferable big data model for anti-counterfeiting inspection of consumer products, significantly improving the ability to identify complex counterfeit behaviors and the efficiency of inspection. Attached Figure Description

[0016] Figure 1 The flowchart of a method for a big data model for anti-counterfeiting inspection of consumer products that integrates artificial intelligence and behavioral graphs is provided in Embodiment 1 of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description relating to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and systems consistent with some aspects of this application as detailed in the appended claims.

[0018] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0019] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0020] Example 1 Please refer to Figure 1 This embodiment provides a big data model method for anti-counterfeiting inspection of consumer products that integrates artificial intelligence and behavioral graphs, including the following steps: S1, acquire multi-source behavioral data; For target consumer products, such as a brand of luxury handbags, collect multi-source behavioral data generated throughout the entire lifecycle of the supply chain. Multi-source behavioral data includes production data, logistics data, sales data, and consumer interaction data. Production-side data, including unique product identifiers such as serial numbers, RFID codes, production batches, production times, production line information, and quality inspection reports, are collected in real time through enterprise resource planning systems or production management systems. Logistics data, including warehouse entry and exit records, transportation trajectories, transit nodes, carrier information, and environmental monitoring data such as temperature and humidity, are acquired through logistics management systems, GPS / BeiDou positioning devices, and IoT sensors. Sales data, including dealer information, sales orders, sales time, sales price, inventory changes, etc., are accessed through the sales terminal system or e-commerce platform interface; Consumer interaction data includes scanning behavior data generated when consumers scan product anti-counterfeiting codes through mobile terminals, specifically covering scanning time, scanning location, scanning frequency, and scanning result feedback, namely authenticity query results and user report information, etc. The aforementioned multi-source behavioral data is stored in a structured or semi-structured form on a distributed data platform, and undergoes unified data cleaning, alignment, and de-identification processing to lay the foundation for subsequent hypergraph construction.

[0021] S2, construct a spatiotemporally coupled commodity circulation hypergraph and embed it into hyperbolic space; Based on the multi-source behavioral data obtained in step S1, a spatiotemporally coupled commodity circulation hypergraph is constructed with commodities, subjects, and events as nodes and behavioral relationships as edges. Node definition: Product nodes are individual product entities, such as a single bag or suitcase, each carrying product attribute characteristics, such as category, batch, production time, etc. The main nodes are the various entities involved in the circulation of goods, including manufacturers, logistics providers, distributors, and consumers. Event nodes are specific behavioral events that occur, such as production completion, warehousing, outbound, sales, and barcode scanning. Each event node is associated with a timestamp and geographic location information. Edge and hyperedge definition: A regular edge connects two nodes, representing a direct behavioral relationship, such as product A being handled by entity B; Hyperedges are used to connect two or more nodes and represent complex association behavior patterns. For example, a hyperedge can connect multiple products in the same batch, the scanning behavior of multiple consumers in the same time period, or a collaborative transfer involving multiple entities, thereby capturing gang-related and cross-regional counterfeiting behavior patterns. After constructing the hypergraph, hyperbolic space embedding is used to represent nodes and hyperedges. Hyperbolic space, such as the Poincaré sphere model, has the continuous analogy property of tree structure, which can naturally capture the hierarchical structure in commodity circulation behavior, such as the brand-category-single product hierarchy, and the power law distribution characteristics, such as a few best-selling products occupying a large number of scanning behaviors. In the specific implementation, a graph neural network based on hyperbolic geometry, such as HGCN or HGNN, is adopted. The hypergraph adjacency matrix and the initial features of the nodes are used as inputs. The low-dimensional embedding vectors of each node and hyperedge are learned through Riemann optimization in hyperbolic space. This enables the distance relationship in the embedding space to reflect the structural similarity and hierarchical relationship in the original behavioral graph. The hyperbolic embedding vector will serve as the basic feature for subsequent rule learning and state extraction. S3, learning differentiable inductive logic programming rules based on hypergraphs; Based on the commodity circulation hypergraph and its hyperbolic embedding generated in step S2, the first-order logic form of the audit rules is automatically learned from historical audit cases through differentiable inductive logic programming ∂ILP, and an evolvable rule base is constructed. Using deep neural networks, such as Transformer, deep features are extracted from multi-source behavioral data and concatenated with the hyperbolic embedding vector obtained in step S2 to form a feature representation that integrates perceptual information. The ∂ILP framework transforms the learning of logical rules into a differentiable neural network training process: Define a set of candidate predicates, such as low-price sales, cross-regional delivery, and high-frequency scanning. Learn the confidence weights of the rules through a neural network and automatically generate first-order logic rules in the form of "low-price sales ∧ new registered seller ∧ cross-regional delivery → suspected counterfeit goods" using gradient-based optimization. Each rule is accompanied by a confidence score and stored in the rule base. The rule base has the ability to evolve. When new audit cases are added or new fraud patterns are detected, the system can update the rule weights based on the new data increments, and even discover new predicate combinations to achieve automatic rule evolution. Since rule learning takes hyperbolic space embedding as input, and hyperbolic embedding itself contains the hierarchical structure of behavioral graphs, the learned rules naturally have condition awareness capabilities. That is, the rules can automatically adjust the applicable threshold for different levels, such as different brands and different channels. S4, Dynamic Evolution of Offensive and Defensive Game and Proactive Audit Decision-Making; The audit process is modeled as a partially observable stochastic game between the auditor (system) and the fraudster (unknown). This game model includes the following elements: The state space, based on the commodity circulation hypergraph constructed in step S2, extracts the feature vector of the current market state, specifically including the circulation distribution characteristics of different categories of goods, such as the inventory turnover rate of each channel, abnormal fluctuation points, and the statistical characteristics of historical inspection results, such as the detection rate of each region, the categories with high incidence of counterfeit goods, and the emotional characteristics of consumer feedback, such as the reporting rate after scanning the code and negative evaluation keywords. These features are aggregated by hyperbolic embedding of hypergraph nodes and hyperedges to form a compact representation of the overall market. Participants—the inspectors and the forgers—are represented by virtual adversaries estimated through generative adversarial imitation learning. The system solves the policy through multi-agent deep reinforcement learning, where multiple agents are responsible for behavioral analysis in different dimensions. Agent A is responsible for analyzing channel partner behavior and monitoring abnormal patterns in distributors' purchasing, sales, and inventory. Agent B is responsible for analyzing consumer verification behavior and identifying abnormal clusters in the distribution of scanning time / location; Agent C is responsible for analyzing logistics trajectories and detecting abnormal detours or time delays in transportation routes. Each intelligent agent shares local observation information and decision confidence through a collaborative mechanism, and jointly outputs a global audit strategy; In terms of operational space, inspectors can choose to focus their inspections on specific product categories, channels, or regions, or issue early warning orders; counterfeiters, on the other hand, can choose to forge targets, distribution channels, and disguise methods. The payoff function is as follows: the inspector's payoff is the expected value of successfully seizing counterfeit goods minus the inspection cost, and the counterfeiter's payoff is the profit from successfully selling counterfeit goods minus the loss from being seized. By solving the approximate Nash equilibrium of the game, the optimal inspection strategy is obtained. In the solution process, the Counterfactual Regret Minimization (CFR) algorithm based on graph theory is used to iteratively optimize the game tree, and the Deep Q-Network (DQN) is combined to process the continuous state space. Finally, proactive inspection decision instructions are generated, such as suggesting that the focus of inspection on Category A products in Channel C of Region B be increased in the next week. The expected detection rate can be increased by x%, and counterfeiters can be forced to move to low-profit areas. S5, Audit Execution and Feedback Update; In response to the proactive audit decision instruction generated in step S4, the system performs specific audit operations on the target product. Audit methods may include retrieving detailed traceability information of the product for manual review, using AI authenticity verification based on image recognition technology, on-site sampling or joint law enforcement, etc. The audit results, including authenticity determination, quantity seized, and counterfeiting characteristics, are recorded and fed back to the rule base. Feedback triggers the update of the rule base. New cases are used as training samples to input into the ∂ILP framework in step S3, adjusting the confidence of existing rules or generating new rules. At the same time, the updated rule base affects the game state in step S4, enabling subsequent strategies to adapt to the evolution of counterfeiting methods in a timely manner. This forms a closed-loop system of graph perception, rule learning, strategy decision-making, and feedback updates, realizing the continuous evolution of anti-counterfeiting and inspection capabilities.

[0022] In practical deployment, spatiotemporal trust enhancement measures can be further introduced, such as attaching a trusted timestamp based on the National Time Service Center and a trusted geographic location code based on BeiDou positioning to behavioral data to improve SuperMap's anti-tampering capabilities. Meanwhile, perceptual hash fingerprints can be extracted from different modalities such as images and RFID signals in multi-source behavioral data as the original features of product nodes, and then low-dimensional representations can be obtained through hyperbolic space embedding, thereby improving the efficiency of large-scale product retrieval. Furthermore, the rule base can support rule transfer learning across brands and categories, and accelerate the cold start of new brands by leveraging existing brand auditing experience. All of the above extended features can be flexibly combined according to actual needs and all fall within the protection scope of this invention.

[0023] Furthermore, based on the multi-source behavioral data obtained in step S1 above, credible time and geographical location information is assigned to the multi-source behavioral data to enhance the spatiotemporal credibility of the commodity circulation supermap and prevent data tampering and forgery. In the process of constructing the spatiotemporally coupled commodity circulation hypergraph in step S2, the following operations are introduced: The first step is to generate a trusted timestamp. For each behavioral event collected, such as goods entering the warehouse or consumers scanning codes to query, the system does not rely on the local device time when recording the time of the event. Instead, it actively sends a request to the time synchronization server of the National Time Service Center to obtain the time signal based on the national standard time. The time signal contains digitally signed timestamp information to ensure the authority and immutability of the time. For example, when a consumer scans the anti-counterfeiting code of a product through a mobile terminal, the scanning time recorded in the scanning behavior data is not the local time of the mobile phone, but a reliable timestamp obtained and issued by the system backend from the National Time Service Center, thus eliminating the loopholes for counterfeiting caused by local time tampering. The second step is to generate a reliable geolocation code. For behavioral events involving geolocation, such as logistics transfers and consumer QR code scanning locations, the system prioritizes using the original positioning data from the BeiDou Navigation Satellite System or the Global Positioning System (GPS) when collecting location data, rather than the fuzzy location reported by the device or the location deduced from the network IP. In practice, when a mobile terminal or IoT device records a location, it simultaneously acquires ephemeris data and positioning calculation results from BeiDou / GPS satellite signals, and adds the original observation values ​​of the satellite signals, such as pseudorange and carrier phase, to form a traceable geographic location code. This encoding has the ability to resist spoofing and replay attacks, ensuring the authenticity and reliability of location information. For example, when a consumer scans the code, the geographical location does not rely solely on the latitude and longitude output by the mobile phone's GPS chip, but combines the short message communication function of the Beidou system to bind the positioning data with satellite signal characteristics, generating a reliable geographical location encoding with a digital signature. The third step is to use reliable spatiotemporal information as an attribute of the event node; The trusted timestamps and trusted geographic location codes generated in the above steps are used as the core attributes of each event node in the commodity circulation hypergraph. Specifically: For production event nodes, the time attribute is the production completion time certified by the National Time Service Center, and the location attribute is the latitude and longitude coordinates of the production plant calibrated by Beidou positioning. For logistics event nodes, the time attribute is a reliable timestamp of entry / exit or transit station, and the location attribute is the precise coordinates of the logistics node via GPS / BeiDou positioning. For a consumer scanning event node, its time attribute is a trusted timestamp that triggered the scanning behavior, and its location attribute is the consumer's real geographical location via Beidou / GPS positioning when scanning the code; These reliable spatiotemporal attributes, together with other characteristics of the nodes themselves, such as product identifiers and subject information, constitute a complete description of event nodes in the hypergraph; The fourth step is to construct a hypergraph structure with a spatiotemporal trustworthy foundation; After defining the nodes and hyperedges, the aforementioned trusted timestamps and trusted geographic location codes are used as constraints to participate in the structured storage and querying of the hypergraph. In the Hypergraph database, a time index based on a trusted timestamp and a spatial index based on a trusted geolocation code are built for each event node, supporting spatiotemporal range queries and spatiotemporal pattern mining. In subsequent steps S3 and S4, the rule engine and game model can directly call these trustworthy spatiotemporal attributes as an important basis for judging abnormal behavior. For example, when the same product is found to have a geographical location jump in a very short period of time, such as from Beijing to Shanghai within 1 minute, and the timestamp shows a continuous untrustworthy middle segment, the system can determine that the scanning behavior is suspected of being counterfeited. Specifically, the commodity circulation supermap constructed in the above manner has timestamps based on the national authoritative time service center and geographical locations based on Beidou / GPS raw positioning data, which ensures the authenticity and non-repudiation of behavioral data from the source, provides high-confidence input features for subsequent neural symbol rule learning, and lays a reliable data foundation for state estimation in attack and defense games, significantly improving the anti-counterfeiting inspection system's ability to combat data forgery. It should be noted that the generation process of trusted timestamps and trusted geolocation codes can be flexibly adapted according to the actual business scenario and the capabilities of terminal devices. For example, in scenarios where IoT devices do not have BeiDou positioning capabilities, base station-assisted positioning combined with the multi-source verification mechanism of Network Time Protocol (NTP) can be used instead, but the authority of the time reference still needs to be ensured through the time synchronization service of the National Time Service Center.

[0024] Furthermore, in step S3 above, based on the commodity circulation hypergraph and its hyperbolic space embedding constructed in step S2, the first-order logic form of the audit rules is automatically learned from historical audit cases through differentiable inductive logic programming (∂ILP), and an evolvable rule base is constructed. This embodiment further refines the rule learning process, specifically including the following operations: The first step is to use deep neural networks to extract deep features from multi-source behavioral data. Before rule learning begins, it is necessary to extract deep features that can characterize the state of commodity circulation from the original multi-source behavioral data. Multi-source behavioral data includes interactive data from the production, logistics, sales, and consumption ends. These data are diverse in form, high in dimensionality, and have complex relationships. Therefore, deep neural networks, such as multilayer perceptrons, Transformers, or graph neural networks, are used to encode features of the input data. For each product node, subject node, or event node, its original attributes, such as product category, price, scanning frequency, and geographical location, are concatenated into an initial feature vector. This vector is then input into a pre-trained deep neural network. Through multiple nonlinear transformations, the network outputs a low-dimensional, dense deep feature vector. This deep feature vector not only retains the key information of the original data but also automatically mines the high-order interaction relationships between different features through end-to-end learning. Taking consumer QR code scanning behavior as an example, the original input may include discrete and continuous features such as scanning time, geographical location, device model, and scanning result. The deep neural network converts the discrete features into vectors through the embedding layer, and then through multiple fully connected layers, finally outputting a deep feature vector that integrates spatiotemporal patterns, user behavior habits, and device credibility. These deep features will serve as the basic input for subsequent logical rule learning. The second step involves learning first-order logic rules with confidence weights through differentiable inductive logic programming. After obtaining deep features, these rules are fused with the hyperbolic space embedding vector obtained in step S2 and used as input to the differentiable inductive logic programming framework. This transforms the learning of logic rules into a differentiable neural network training process, enabling optimization using gradient descent. In the specific implementation, a set of domain-related predicates are first defined, such as low-price sales, newly registered sellers, cross-regional delivery, and high-frequency scanning. These predicates represent the basic facts or attributes that may affect the determination of the authenticity of goods. Then, a template rule base is built, which contains a large number of candidate rule structures. Each rule is composed of multiple predicates connected by logical AND ∧, such as "low-price sales ∧ newly registered sellers ∧ cross-regional delivery → suspected counterfeit goods". Among them, "suspected counterfeit goods" is the target predicate, indicating that the goods are judged to be counterfeit. Differentiable inductive logic programming assigns a trainable confidence weight to each candidate rule. The initial value can be set randomly or assigned based on prior knowledge. At the same time, a fuzzy truth value is assigned to the truth value of each predicate, which is between 0 and 1. This truth value is predicted by the deep features extracted in the first step through a neural network. For example, the truth value of low-price sales can be obtained by mapping the price features of the product through a small network. The goal of the entire rule learning process is to minimize the prediction result, namely the difference between the probability of suspected counterfeit goods inferred from the rule base and the actual labels of historical inspection cases. This is achieved by updating the predicate prediction network and the rule confidence weights simultaneously through the backpropagation algorithm. After training, rules with higher weights are considered valuable audit rules, and each rule is accompanied by a confidence score to indicate the credibility of the rule. In this process, domain prior knowledge can be incorporated in two ways: first, by initializing the rule confidence weights, expert experience can be encoded into higher initial values; second, by constraining the rule structure, such as stipulating that certain predicates must appear simultaneously, the search space can be narrowed. The third step is to incrementally update the rule base in response to new audit cases or new counterfeiting patterns. The rule base has the ability to evolve online. When the system performs an audit operation, the newly generated audit cases, including goods confirmed to be genuine and counterfeit goods seized, will be fed back to the rule learning module. At the same time, if a new counterfeiting pattern is detected, such as an abnormal combination of behaviors that has never been seen before, the system will also trigger a rule update. Incremental updates use two methods: Online fine-tuning allows for a small gradient update of the rule confidence weights for a small number of new samples, without the need to retrain the entire predicate prediction network, thus enabling a rapid response. Retraining: When a certain number of new fraudulent patterns are accumulated, or when the original rule system is detected to be invalid, the system starts periodic retraining. It uses all historical data, including new cases, to re-execute the first and second steps, generate a brand new rule base, or reconstruct the rule structure. Through this incremental update mechanism, the rule base can continuously evolve with the evolution of counterfeiting methods, always maintaining sensitivity and identification capabilities against new types of counterfeiting. At the same time, since the rules are stored in an interpretable first-order logic form, inspectors can intuitively understand the meaning of each rule, which facilitates manual review and law enforcement evidence collection. The aforementioned deep feature extraction network, hyperbolic space embedding module, and differentiable inductive logic programming framework can be trained end-to-end or independently in stages. The specific implementation method can be flexibly selected according to the actual data and computing power conditions.

[0025] Furthermore, in step S4 above, the inspection process is modeled as a partially observable stochastic game between the inspector and the counterfeiter. Based on the current market state features extracted from the commodity circulation hypergraph and the rule base constructed in step S3, the game strategy is solved through multi-agent deep reinforcement learning. This embodiment further refines the game strategy solution process, specifically including the following operations: Based on the commodity circulation hypergraph, the system extracts the current market state feature vector. In order to accurately depict the overall situation of the current market, the system first extracts multi-dimensional feature vectors from the commodity circulation hypergraph constructed in step S2 as the state input of the game model. The feature vector includes at least the following three types of information: The distribution characteristics of different product categories are based on product nodes and their associated event nodes in the hypergraph. The statistics of each product category in various distribution channels, such as offline counters, e-commerce platforms, and second-hand trading markets, include inventory turnover rate, sales volume, and average circulation time. Abnormal fluctuation points are also identified, such as a sudden increase in inventory or a sharp drop in turnover rate in a certain channel. These distribution characteristics reflect the normal flow pattern of goods in the market and potential abnormal areas. The statistical characteristics of historical inspection results summarize the inspection case data over a period of time, such as the most recent 30 days, including the detection rate of each region and channel, the categories of counterfeit goods that are frequently found, and the distribution of counterfeit methods. These statistical characteristics reveal the active areas and main methods of counterfeit activities, providing a basis for selecting current inspection priorities. Consumer feedback sentiment characteristics: Through natural language processing technology, sentiment analysis is performed on comments, reports, and discussions on social media platforms after consumers scan the code. Indicators such as the intensity of negative emotions and the frequency of keywords, such as poor quality and incorrect packaging, are extracted. Consumer feedback can often provide early warning of the emergence of new counterfeit models. After normalization and vectorization, the three types of features mentioned above are concatenated into a multi-dimensional state vector. This vector contains both objective circulation data and subjective consumer perception, and can comprehensively reflect the current market's authenticity and health. This state vector will serve as the basic input for subsequent game theory solutions.

[0026] Generative adversarial imitation learning is used to estimate the strategy space and behavioral preferences of the counterfeiter. In real-world scenarios, the counterfeiter's true strategy is unknown and will dynamically adjust with changes in the intensity of investigation. In order to simulate the counterfeiter's behavior in the game model, the system uses generative adversarial imitation learning to estimate its strategy space and behavioral preferences. Data from historically uncovered counterfeit cases is collected, including the target product categories, distribution channels, disguise methods, and time distribution chosen by the counterfeiters, serving as a model trajectory for the counterfeiters. Subsequently, a generative adversarial network framework is constructed. The generator simulates the counterfeiter's strategy and outputs the probability distribution of the counterfeit behavior decision based on the current market state, i.e., the extracted state vector. The discriminator distinguishes between the fraudulent behavior generated by the generator and real historical fraud cases, and outputs a reward signal. Through alternating adversarial training between the generator and the discriminator, the generator gradually learns to generate behavioral strategies that are difficult to distinguish from historical fraud cases. After training, the generator serves as an approximate estimate of the fraudster's strategy and can predict the most likely actions of the fraudster based on any market state. This estimation result will serve as the fraudster's strategy function in the game model and will be used for subsequent equilibrium solutions.

[0027] The system employs a graph theory-based counterfactual regret minimization (CFR) algorithm to solve for the approximate Nash equilibrium strategy. After obtaining the market state and the fraudster's strategy estimate, the system models the inspection process as a partially observable random game and uses the graph theory-based counterfactual regret minimization (CFR) algorithm to solve for the approximate Nash equilibrium strategy. The participants in the game include the inspector and the counterfeiter, which are composed of multiple agents working together. The state space is composed of extracted feature vectors. The action space of the inspector includes the selection of key categories, channels, regions and inspection intensity for inspection. The action space of the counterfeiter is determined by the estimated strategy. The payoff function is set according to the inspection cost and the value of the seized items. The CFR algorithm is used to solve the problem. The traditional CFR algorithm requires traversing the game tree, but it is difficult to apply directly in continuous states and large-scale action spaces. Therefore, an improved CFR algorithm based on graph theory is adopted, which introduces the structural information of the commodity circulation hypergraph into state aggregation and action abstraction. By utilizing the hierarchical relationship between nodes and hyperedges in the hypergraph, similar product nodes or channel nodes are clustered, and the continuous state space is discretized into a finite number of state clusters. Based on this, the counterfactual regret value on each information set is iteratively calculated, and the strategy probability is updated. After the algorithm converges, the optimal hybrid strategy of the inspector in each possible state is obtained. Multi-agent collaboration involves multiple agents on the auditing side, each responsible for analyzing channel behavior, consumer verification behavior, and logistics trajectory. Within the CFR framework, global optimization is achieved through shared value functions and collaborative exploration. Local observations by each agent, such as abnormal fluctuations in channel behavior, are fused using graph neural networks to form a consistent estimate of the overall market state.

[0028] The system generates proactive inspection instructions based on the approximate Nash equilibrium strategy. These instructions must include at least the following elements: Key targets for inspection include specific product categories, such as a particular brand of bags or a specific batch of cosmetics that require priority attention. Key audit channels: Identify which distribution channels, such as e-commerce platforms and second-hand trading markets, require increased auditing efforts. Key audit timings, specifying the time periods in which audit resources should be deployed, such as before and after promotional seasons and holidays; If the equilibrium strategy may output a recommendation to conduct key spot checks on the transactions of brand A bags on second-hand platform B in the coming week, while strengthening the monitoring of logistics transit stations in region C, the detection rate is expected to increase by about 15%, and it may force counterfeiters to turn to the low-profit channel D. This instruction will be pushed to the audit execution end, triggering the specific audit operation in step S5. Specifically, this embodiment extracts multi-dimensional state features from the commodity circulation hypergraph, combines generative adversarial learning to estimate the counterfeiter's strategy, and uses a graph theory-based counterfactual regret minimization algorithm to solve for the approximate Nash equilibrium. Finally, it generates a forward-looking proactive inspection instruction, realizing a leap from passive response to proactive deterrence and improving the efficiency and accuracy of anti-counterfeiting inspection.

[0029] Furthermore, in the process of constructing the spatiotemporally coupled commodity circulation hypergraph in step S2 above, in order to uniformly represent multi-source heterogeneous behavioral data as original features that can be embedded in the hyperbolic space subsequently, this embodiment introduces cross-modal perceptual hash fingerprint extraction technology. This process includes the following operations: Perceptual hash fingerprints are extracted for different modalities of data. Multi-source behavioral data contains data of various modalities, each with different data formats and semantics. To unify these modalities into the hypergraph node representation, unique and robust perceptual hash fingerprints are extracted for each modality of data, specifically including: Perceptual hash extraction of product image data: For the collected product appearance images, such as product photos uploaded by consumers, display images on e-commerce platforms, and quality inspection photos, the perceptual hash algorithm pHash based on discrete cosine transform is used. The image is scaled to a fixed size, such as 32×32 pixels, converted to grayscale, and then subjected to discrete cosine transform. Low-frequency coefficients are retained and a binary hash sequence is generated. This fingerprint is robust to image scaling, compression, and slight changes in lighting. It can uniquely identify the appearance features of a product. For example, the genuine image of a certain brand of bags is used to generate a 64-bit binary fingerprint "10110010..." using the pHash algorithm. The perceptual hash extraction of radio frequency identification signals involves reading the original signal characteristics of RFID tags attached to goods, including signal strength RSSI, phase, Doppler frequency shift, etc., discretizing these continuous signals, extracting statistical features such as mean, variance, and zero-crossing rate, and quantizing them into binary sequences. At the same time, the encrypted unique identifier stored in the RFID chip, such as TID, is used as a fixed part and spliced ​​with the signal features to form the perceptual hash fingerprint of the RFID mode. This fingerprint can not only uniquely identify the chip, but also characterize the physical characteristics of the signal, thus preventing tag cloning. Perceptual hash extraction of logistics codes involves first standardizing and parsing various logistics codes generated during the commodity circulation process, such as barcodes, QR codes, and express tracking numbers, to extract structured information from the codes, such as production batch, destination, and carrier. Then, this structured information is hashed and mapped, such as by using the SimHash algorithm, to generate a fixed-length logistics code fingerprint. This fingerprint can quickly match the same or similar logistics path patterns. Perceptual hash extraction of consumer scanning behavior sequence: For the behavior sequence generated when consumers scan codes to check authenticity through mobile terminals, including scanning time, geographical location, scanning frequency, scanning result feedback, etc., construct behavioral temporal features; The timestamps are normalized into relative time series, the geographic location is encoded into a grid index, and the behavior type, such as first scan, repeated scan, and report, is mapped into category features; Then, a time-series hashing algorithm, such as Temporal Hash, is used to compress the variable-length behavioral sequence into a fixed-length binary fingerprint, preserving the statistical characteristics of the behavioral pattern. For example, a consumer's abnormal scanning pattern, such as multiple scans in a short period of time or geographical location jumps, will be encoded into a specific hash fingerprint. The perceptual hash fingerprint lengths of the above four modalities can be uniformly designed to be 64 bits or 128 bits, which facilitates subsequent fusion processing.

[0030] Using the perceptual hash fingerprint as the original feature vector of the product node, after extracting the perceptual hash of each modality, the multimodal hash fingerprints corresponding to the same product are concatenated or fused to form the original feature vector of the product node. The fusion method can be to use concatenation, such as concatenating four 64-bit hashes into a 256-bit vector, or to use weighted average, assigning different weights to different modalities, and then performing binarization. This original feature vector contains information about the product in multiple dimensions, such as image appearance, physical identification, logistics trajectory and consumer behavior, and is a compact representation of the product's multi-source behavioral data; Taking a luxury handbag as an example, its original feature vector may consist of the following parts: Image hash, 64-bit, generated from a product appearance photo; RFID hash, 64 bits, generated based on tag signal characteristics; Logistics code hash, 64 bits, generated based on the entire logistics trajectory; Behavioral sequence hash, 64 bits, generated based on all consumer scan records; The final result is a 256-bit original feature vector, which serves as the initial attribute of the product node in the hypergraph.

[0031] After being embedded in hyperbolic space, the original feature vectors, although having compressed and encoded multimodal information, are still in high-dimensional Euclidean space and cannot directly express the hierarchical structure and complex relationships in commodity circulation behavior. Therefore, they need to be used as input and embedded in hyperbolic space in step S2 to be mapped into low-dimensional hyperbolic representations. In the specific implementation, the original feature vector of each product node is used as input to the hyperbolic graph neural network, such as the input layer of HGCN. This network performs feature transformation and neighbor aggregation in hyperbolic space, and finally outputs a low-dimensional embedding vector of each node in hyperbolic space, such as 16-dimensional or 32-dimensional. This hyperbolic embedding vector not only preserves the multimodal information in the original features, but also naturally captures the hierarchical relationship between products through the tree-like structure of hyperbolic geometry; Perceptual hash fingerprint extraction and hyperbolic space embedding are two-level operations. The former is responsible for extracting robust and compact features from the raw data, while the latter is responsible for mapping these features to a hyperbolic space suitable for expressing hierarchical structures. Together, they enable nodes in the hypergraph to have both lightweight and fast-retrieval hash representations and geometric structural information that can support complex reasoning. In subsequent steps S3 and S4, these low-dimensional hyperbolic representations can be directly used as input features, which ensures computational efficiency while preserving rich structural and semantic information. At the same time, the perceptual hash fingerprint itself can also be used to construct a locality-sensitive hash index, supporting rapid retrieval and comparison of massive amounts of goods.

[0032] Furthermore, in the process of constructing the spatiotemporally coupled commodity circulation hypergraph in step S2 above, in order to incorporate consumer-end interaction behavior into the overall behavior graph, this embodiment has specially processed the consumer scanning behavior data, treating it as an independent consumer-end interaction event node, and establishing hyperedge connections with relevant nodes to form a commodity circulation holographic graph containing consumer behavior dimensions. Specifically, it includes the following operations: Collect consumer scanning behavior data. In step S1, the consumer-end interaction data from the multi-source behavior data specifically includes scanning behavior data generated when consumers scan product anti-counterfeiting codes, such as barcodes, QR codes, or RFID, using mobile terminals. Each scanning event is recorded as a data record, which includes at least the following fields: The scanning time refers to the specific time point that triggers the scanning behavior. To ensure credibility, a trusted timestamp based on the National Time Service Center can be used as described in step S2. The location of the scanned QR code is the actual location of the consumer when scanning the QR code. It prioritizes the use of trusted geolocation codes based on the BeiDou Navigation Satellite System or the Global Positioning System to ensure the authenticity of the location. Scan frequency is the number of times the same consumer or the same device scans the same product. It can be dynamically calculated by combining historical records. For multiple consecutive scans, the system will record how many times the product has been scanned. The results of scanning the code include the authenticity verification results obtained by consumers after scanning the code, such as whether it is a genuine product, suspected counterfeit, or reported, as well as feedback information actively submitted by consumers, such as abnormal product appearance or suspicious purchase channels. These results and feedback will serve as important labels for subsequent rule learning and game decision-making. After being collected, this QR code scanning data undergoes cleaning, deduplication, and anonymization to form structured consumer behavior records, which serve as one of the basic inputs for subsequent hypergraph construction.

[0033] By treating consumer QR code scanning behavior data as consumer-side interaction event nodes, each independent QR code scanning behavior record is abstracted as a consumer-side interaction event node when constructing a spatiotemporally coupled commodity circulation hypergraph. The attributes of this node include at least: Node type identifier, marked as a consumption event; Time attribute: the time of scanning the code, i.e., the trusted timestamp; Spatial attributes, scan the geolocation code, and you can get a reliable geolocation code; Behavioral attributes, such as scanning frequency and scanning result feedback, may include whether a report was filed and the type of report; Associated identifiers are unique identifiers of the scanned product, such as product ID, and the sales channel identifier where the scanning occurred, such as e-commerce platform store ID or physical store code. This node is added to the hypergraph, where it stands alongside other types of nodes, such as product nodes, subject nodes, and other event nodes, together forming the node set of the commodity circulation hypergraph.

[0034] To establish hyper-edge connections, consumer-end interaction event nodes must connect with relevant product nodes and sales channel nodes to reflect the position and role of consumer behavior in the commodity circulation chain. For each consumer event node, the following operations are performed: Find the product node corresponding to the event, based on the product's unique identifier; Locate the sales channel node corresponding to the event, based on the channel information of the product when the scan occurred, such as online store ID or offline store ID; Create a superedge that simultaneously includes the consumption event node, product node, and sales channel node. This hyperedge connects more than two nodes, which is a typical hypergraph structure. Through this hyperedge, the system can capture the following complex relationships: When the same product is scanned by different consumers at different times and locations, these consumption event nodes, the product node, and their respective sales channel nodes form multiple hyperedges, which can be used to analyze the geographical distribution of consumers and changes in scanning popularity of the product. When the same consumer scans the codes of different products multiple times, a cross-product consumption behavior pattern can be formed by taking the consumer as the main node. However, this embodiment focuses on the connection between event nodes and products and channels. When multiple consumer events occur in the same sales channel, these event nodes and channel nodes can form a hyperedge containing multiple event nodes, reflecting the consumer activity and authenticity disputes in that channel; This hyperedge connection method formally integrates the consumer behavior dimension into the commodity circulation hypergraph, making it no longer merely an accessory attribute of commodity nodes, but rather an independent node and relationship participating in the construction and subsequent analysis of the graph.

[0035] A holographic map of commodity circulation incorporating consumer behavior dimensions is formed. Through the above steps, the hypermap, which originally only included production, logistics, and sales, now fully integrates consumer behavioral data such as scanning codes, inquiries, and feedback. The resulting hypermap is called a holographic map of commodity circulation incorporating consumer behavior dimensions, and its characteristics are: Comprehensive, covering behavioral data across the entire chain from the source of production to the end consumer, especially consumer interaction data that has been previously overlooked; Real-time performance: Consumer scanning behavior usually occurs in real time, enabling HyperGraph to reflect the latest market dynamics; The behavior is explainable. Abnormal patterns in consumer scanning behavior, such as high-frequency scanning, cross-regional scanning, and concentrated reports, can be directly used as key signals for subsequent steps S3 and S4. If a product receives a large number of scan requests from different consumers in the same geographical location within a short period of time, and the scan results frequently result in reports, the system can quickly identify potential counterfeit risks by combining the product's circulation path and sales channels, and submit the consumer event cluster as an abnormal subgraph to the rule engine. In step S2, the consumer scanning behavior sequence itself is also used to extract perceptual hash fingerprints, but that is a compressed representation of the entire behavior sequence for the construction of the original features of the product node. In this embodiment, each scanning event is included in the hypergraph as an independent node. The two are modeling consumer behavior at different granularities, complementing each other and enriching the expressive power of the commodity circulation hypergraph.

[0036] Furthermore, based on the multi-source behavioral data acquired in step S1 above, the data collection scope is expanded to include process parameters and quality inspection data from the production process, as well as environmental monitoring data from the distribution process. In step S2, these are used as attribute features of event nodes to participate in the hypergraph construction, enhancing the commodity distribution hypergraph's ability to perceive the intrinsic quality of commodities and the external environment. Specifically, this includes the following operations: In step S1, in addition to the regular production-side data, the following two types of fine-grained data from the production process are specifically collected: Process parameter data refers to the key process indicators generated during the production and manufacturing process of a product. Taking luxury handbags as an example, process parameters may include leather cutting pressure, sewing thread tension, gluing temperature, drying time, etc. Taking food as an example, process parameters may include sterilization temperature, filling pressure, packaging sealing index, etc. These data are collected in real time by sensors on the production equipment and the corresponding process parameter values ​​for each product or each production batch are recorded. The process parameter data reflects the production quality and process consistency of the product and is an important basis for judging whether the product is genuine, because counterfeit products often cannot completely replicate the production process parameter distribution of genuine products. Quality inspection data refers to the data generated after a product has been manufactured and has undergone quality inspection. This includes appearance inspection results, such as defect detection scores, functional test results, such as power-on tests of electronic products, dimensional measurement data, weight deviations, etc. Quality inspection data usually exists in the form of inspection reports, which include the inspection items, inspection values, and pass / fail results for each product or batch of products. Quality inspection data records the quality status of products before they leave the factory and provides a benchmark for the authenticity comparison in subsequent circulation links. The two types of data mentioned above are collected in real time during the production process and bound to unique product identifiers, such as serial numbers and RFID codes, and stored as attributes of production event nodes.

[0037] Environmental monitoring data is collected during the distribution process. After the goods enter the distribution process, this embodiment also collects environmental monitoring data on the goods during warehousing and transportation, specifically including: Temperature data is recorded by IoT temperature sensors to measure the ambient temperature of goods in warehouses and transport vehicles. It is particularly suitable for temperature-sensitive products such as food, medicine and cosmetics. Abnormal temperature may affect the quality of goods and may even become an indirect clue to determine whether a product is genuine. Genuine products usually follow strict temperature control standards, while counterfeit products may ignore temperature control during transportation. Humidity data is recorded by a humidity sensor to monitor the relative humidity of the environment and prevent products from becoming damp and spoiling. Vibration and shock data are used to record vibration acceleration and impact events during transportation of precision goods such as electronic products and handicrafts to prevent damage caused by rough handling. Light data, for light-sensitive products such as certain chemicals and works of art, records the intensity and duration of light exposure; These environmental monitoring data are collected in real time by IoT devices deployed on logistics nodes and transportation vehicles, and are bound to corresponding timestamps, geographical locations, product batches or individual product identifiers. They are stored as attributes of logistics event nodes. Anomalies in environmental monitoring data are often related to the authenticity of goods or the compliance of transportation. For example, the temperature record of genuine cold chain transportation should be continuous and stable throughout the process, while counterfeit products may lack cold chain or data.

[0038] Using process parameter data and environmental monitoring data as attribute features of event nodes, the newly added data is integrated into the corresponding event nodes when constructing the spatiotemporally coupled commodity circulation hypergraph in step S2: For each production event node, such as when a product is produced on a production line at a certain time, the associated process parameter data and quality inspection data are used as the attribute features of the event node. For example, the attribute set of a production event node may include key-value pairs such as leather cutting pressure 3.2 bar, sewing thread tension 1.5 N, and appearance inspection score 98 points. These attribute features are stored in the node in a structured form for subsequent querying and analysis. For each logistics event node, such as a product entering a warehouse at a certain time or a product being in a transport vehicle during a certain period, the associated environmental monitoring data is used as the attribute characteristics of the event node. For example, a warehouse entry event node can be accompanied by attributes such as temperature of 22℃ and humidity of 45% at the time of entry; a transport event node can be accompanied by statistical indicators such as average temperature of 23.5℃, highest temperature of 28℃, and number of impact events of 3.

[0039] When participating in the construction and subsequent application of hypergraphs, after incorporating the above-mentioned attribute features into the hypergraph, the event nodes in the hypergraph no longer only contain basic time and location information, but also contain rich production quality information and environmental compliance information. These newly added attribute features, together with the structural information of the node itself, such as connection relationships, constitute a complete node representation during the hypergraph construction process. In subsequent steps S3 and S4, these attribute features can be directly used as the basis for judging input features or rule predicates. For example, rule learning may find that if the production process parameters of a certain product deviate from the standard range and the quality inspection score is lower than the threshold, it is very likely to be a counterfeit product. When extracting the game state, the average environmental monitoring anomaly rate of a certain area can be used as one of the indicators of market health to adjust the focus of inspections.

[0040] Furthermore, in step S3 above, first-order logic rules are automatically learned from historical audit cases using differentiable inductive logic programming (∂ILP), and an evolvable rule base is constructed. This embodiment further refines the storage structure and expansion capabilities of the rule base, specifically including the following operations: Rules are stored in an interpretable first-order logic form. The learned rules are stored in a structured manner as explicit, human-readable first-order logic expressions. Each rule corresponds to a record in the rule base. The core of the rule is the logic expression itself. For example, a rule learned through ∂ILP may be expressed as: Low-price sales ∧ Newly registered seller ∧ Cross-regional delivery → Suspected counterfeit goods. This explicit storage method makes the rules not only executable by machines, but also directly understandable to business personnel and inspectors, meeting the requirements of traceability and interpretability of the basis for judgment in administrative law enforcement.

[0041] To associate rules with applicable conditions, each rule is not applicable in all scenarios, therefore it is necessary to associate it with explicit applicable conditions, which may include: Specify which brands the rules apply to, such as applying only to brand A and brand B, or excluding certain brands, such as not applying to brands. Different brands may have different product characteristics, distribution channels, and counterfeiting methods, so the rules need to be limited at the brand level. Specify which product categories the rules apply to, such as bags and cosmetics. Some rules may only target the counterfeiting characteristics of specific product categories. The rules specify which geographical regions the rules apply to, as the methods of fraud and the regulatory environment may differ in different regions. Specify which distribution channels the rules apply to, such as online e-commerce, offline stores, and second-hand platforms; Time applicability: the rules may only be valid for a specific period of time, such as during a promotional season, or may only apply to newly launched products; Applicable conditions are stored in the form of metadata associated with the rules. For example, the above rules can be associated with applicable conditions: brand ∈ {A, B, C}, category ∈ {bags}, region ∈ {nationwide}, channel ∈ {second-hand platform}. When performing audit reasoning, the system first filters out the set of applicable rules based on the current context, such as the brand and category of the product to be inspected, and then performs reasoning based on these rules.

[0042] To associate confidence with rules, each rule has a confidence weight, which represents the reliability of the rule in historical audit cases. The confidence is a real number between 0 and 1, which is automatically learned during the ∂ILP training process and can also be dynamically adjusted based on subsequent feedback. The higher the confidence of a rule, the greater its weight in reasoning and decision-making. The confidence level is stored along with the rule and supports online updates. When a new audit case is fed back to the rule base, the system can recalculate the confidence level of the rule. If a rule is frequently verified in new cases, its confidence level can be appropriately increased. If a case occurs that is contrary to the rule's prediction, the confidence level may decrease. The confidence level can be updated using the Bayesian method or online gradient updates.

[0043] To associate lifecycle information with rules, it is important to note that rules are not permanent and their lifecycles need to be managed. The lifecycle information associated with each rule should include at least the following: Creation time: The timestamp when the rule was first added to the rule base; Validity period: The expected validity period of the rule. After the validity period expires, the rule may be marked as pending verification or automatically expires. Expiration time: The time when the rule is marked as expired; Last verification time: the time when the rule was last verified or the confidence level was last updated. Number of times the rule is used; the number of audit cases in which the rule has been successfully applied. Lifecycle information is used for the maintenance and cleanup of the rule base. For rules that have not been used for a long time or whose confidence level is below the threshold, the system can periodically review or eliminate them to avoid the rule base becoming bloated and outdated. At the same time, when new fraud methods emerge, the system can prioritize creating new rules and give them a longer validity period for verification.

[0044] It supports rule transfer learning across brands and categories. The rule base not only serves a single brand or category, but also has the ability to transfer rules across brands and categories. Transfer learning utilizes the rule knowledge of existing brands or categories to accelerate the construction of rule bases for new brands or categories. The specific implementation includes the following mechanisms: For each rule, the system evaluates its general applicability across different brands and product categories. For example, the rule "low-price sales ∧ new registered seller → suspected counterfeit goods" may have high cross-brand applicability, while rules involving specific production processes, such as "abnormal leather cutting pressure → suspected counterfeit goods", may only apply to specific brands. The general applicability evaluation can be measured based on the domain specificity of the predicates in the rule, or the generalization performance of the rule can be calculated through cross-domain data testing. Rule adaptation and migration: When introducing a new brand or product category, firstly, highly general rules are retrieved from the rule base, and then fine-tuned based on a small amount of labeled data from the new domain. Adaptation methods include: Rule threshold adjustments, for example, the threshold for low-price sales in the original rule may need to be adjusted under the new brand, and the threshold can be relearned through new data; Rule confidence reassessment involves recalculating the confidence of migration rules based on validation data from the new domain. Predicate mapping: If there are different but semantically similar predicates in a new domain, the system can perform predicate replacement through ontology mapping or similarity calculation; During the incremental update of the rule base, the system can simultaneously use historical cases from multiple brands and product categories for joint training to learn a set of rules that are both domain-specific and cross-domain universal. This joint learning can uncover common patterns of counterfeit behavior among different brands and enhance the generalization ability of the rule base. For example, when the system migrates from luxury handbag brand A to brand B, it first selects general rules applicable to the handbag category from the rule base, such as rules related to channel behavior and consumer scanning. Then, it uses a small number of historical cases from brand B to fine-tune the confidence and threshold of the rules, quickly forming the initial rule base of brand B. Compared with learning from scratch, transfer learning can significantly reduce the need for labeled data and accelerate the deployment of anti-counterfeiting capabilities for new brands. The aforementioned metadata, namely applicable conditions, confidence level, and lifecycle, together with the rule ontology, constitute a complete record of the rule base. The rule base is stored using a database or graph database, supporting efficient rule retrieval, updates, and version management. In the feedback update in step S5, the addition of new cases will trigger incremental updates to the rule base, including operations such as confidence level adjustment, new rule generation, and old rule invalidation, thereby realizing the dynamic evolution of the rule base. The rule base constructed in the above manner not only has interpretable rules, but also comes with rich metadata to support scenario-based applications and lifecycle management. It also has cross-brand and cross-category transfer learning capabilities, which significantly improves the adaptability, maintainability and reusability of the rule base.

[0045] Furthermore, in step S5 above, after the audit operation is executed in response to the proactive audit decision instruction, the audit result is fed back to the rule base to update the evolvable rule base. This embodiment refines the feedback update mechanism, clarifies how it triggers the re-execution of steps S3 and S4, and adopts an incremental learning approach to achieve online updates of model parameters, specifically including the following operations: Audit result feedback and update triggering: After each audit operation is completed, the detailed result data of this audit is packaged into a feedback sample, which includes at least: The unique identifier of the goods to be inspected, the risk score or warning level generated by the rule base and game strategy before the inspection, the authenticity label determined after the actual inspection, such as genuine products, counterfeit products or exclusion of suspects, and new behavioral characteristics or abnormal patterns discovered during the inspection process, such as scanning time aggregation patterns that have not appeared before. The feedback sample is written into the audit history database and simultaneously pushed to the rule base management module. After receiving the new sample, the rule base management module automatically triggers the re-execution of steps S3 and S4. The triggering condition can be set to trigger every time a new sample is added, or to trigger in batches after a certain number of new samples are accumulated, such as 10. The specific conditions can be dynamically adjusted according to the system's real-time requirements and computing resources.

[0046] When re-executing step S3, an incremental learning approach is adopted, updating the existing rule base only based on the newly added audit result data. The specific implementation of incremental learning includes: The incremental adjustment of rule confidence involves adjusting the confidence weight of each existing rule using online gradient descent or Bayesian update methods, based on its applicability and prediction accuracy in the new samples. If a rule is successfully validated in the new samples (i.e., the rule prediction matches the audit results), its confidence is slightly increased; if the rule prediction is incorrect, the confidence decreases. The confidence update formula can be designed as follows:

[0047] Where w is the rule confidence, α is the learning rate, and y is the true label. For rule-based reasoning results, this process only involves the current rule weights and does not require retraining the deep neural network; When an abnormal pattern appears in a new sample that cannot be covered by existing rules, the system triggers local rule exploration. Based on the characteristics of the new sample, candidate new rules are generated through local search or genetic programming on the basis of the existing rule template library. The effectiveness is verified using a small number of new samples. If the verification is successful, the new rule is added to the rule library and its confidence, applicable conditions and life cycle information are initialized. Incremental fine-tuning of the predicate prediction network, the deep neural network in step S3, is used to extract deep features and predicate ground truth. It can also be updated by incremental learning, using the features and labels of new samples to perform gradient updates on the network in a small number of iterations without retraining the entire network. This method can adapt to small drifts in data distribution while avoiding catastrophic forgetting. Through the incremental updates described above, the rule base can be quickly adjusted after receiving new feedback without incurring high computational costs.

[0048] After the rule base is updated, the dynamic evolution model of the attack-defense game in step S4 also needs to be re-executed to generate an inspection strategy that adapts to the new market state. Similarly, the update process adopts an incremental learning approach. Based on the information in the new feedback samples, the current market status feature vector is updated online. For example, after a new case is detected in a certain area, the detection rate in the statistical features of the historical inspection results of that area can be adjusted in real time. The emotional features in the consumer's QR code feedback can be calculated incrementally through a sliding window, without the need to re-aggregate all historical data. Incremental adjustment of the forger's strategy estimation: In step S4, the forger's strategy model estimated by generating adversarial imitation learning can be incrementally trained using new samples. Newly discovered forgery cases are used as new demonstration trajectories to perform a small number of iterative adversarial trainings on the generator and discriminator, enabling the model to quickly adapt to the evolution of forgery methods. Incremental solution of game strategy involves re-running the graph theory-based counterfactual regret minimization algorithm after updating the state features and the opponent's strategy. However, since the state space may have changed, a complete re-solution may be time-consuming. Therefore, the incremental CFR algorithm is adopted, which retains the strategy obtained in the previous round as the initial value and only updates the local regret value and adjusts the strategy for the changed state nodes, thereby quickly converging to a new approximate equilibrium.

[0049] Through the aforementioned incremental update mechanism, the system achieves a rapid closed loop of audit execution, feedback, rule learning and update, game strategy update, and new round of audit instructions. Each new audit result can influence subsequent audit decisions in a short period of time, enabling the system to keep up with the evolution of counterfeit methods. For example, when a new cross-platform counterfeit sales model appears on the market, after the first batch is detected, the relevant characteristics are immediately used to adjust the rule confidence and strengthen the audit weight of this model in the next round of game strategy, thereby effectively curbing its spread. Incremental updates do not mean abandoning historical data. The system will still perform full retraining regularly, such as weekly or monthly, to fully optimize the rule base and game model using all historical data, correcting any biases that may accumulate from incremental updates. Full retraining and incremental updates complement each other and work together to maintain the long-term stability and adaptability of the system. Specifically, an incremental learning mechanism is introduced in the feedback stage of step S5, enabling steps S3 and S4 to be quickly re-executed based on new data. This achieves online updates of the rule base and game strategy, ensuring that the anti-counterfeiting inspection system has the ability to respond in real time and evolve continuously.

[0050] This embodiment also provides a consumer product anti-counterfeiting inspection big data model system that integrates artificial intelligence and behavioral graphs, configured to execute the above-mentioned method for implementing the consumer product anti-counterfeiting inspection big data model that integrates artificial intelligence and behavioral graphs.

[0051] Although alternative embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make further changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0052] The above specific embodiments further illustrate the purpose, technical solution and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this invention.

Claims

1. A big data model method for anti-counterfeiting inspection of consumer products that integrates artificial intelligence and behavioral graphs, characterized in that, Includes the following steps: S1, acquire multi-source behavioral data of the target consumer product throughout its entire supply chain lifecycle, including production data, logistics data, sales data, and consumer interaction data; S2, Based on the multi-source behavioral data, construct a spatiotemporally coupled commodity circulation hypergraph with commodities, subjects, and events as nodes and behavioral relationships as edges; The hypergraph uses hyperbolic space embedding to represent nodes and hyperedges, capturing the hierarchical structure and complex association patterns of commodity circulation behavior. S3, Based on the commodity circulation hypergraph, the first-order logic form of the inspection rules is automatically learned from historical inspection cases through differentiable inductive logic programming, and an evolvable rule base is constructed; The rule learning process uses the hyperbolic space embedding as input features to achieve conditional awareness and dynamic adjustment of the rules. S4, the inspection process is modeled as a partially observable random game between the inspector and the counterfeiter. Based on the current market state features extracted from the commodity circulation hypergraph and the rule base, the game strategy is solved through multi-agent deep reinforcement learning to generate proactive inspection decision instructions for different commodity categories, circulation channels or regions. The multi-agent system includes multiple collaborative agents responsible for channel merchant behavior analysis, consumer verification behavior analysis, and logistics trajectory analysis, respectively. S5, in response to the proactive audit decision instruction, perform an audit operation on the target product, and feed the audit results back to the rule base to update the evolvable rule base.

2. The method according to claim 1, characterized in that, Step S2 further includes: assigning a trusted timestamp and a trusted geographic location code to the multi-source behavioral data, wherein the trusted timestamp is generated based on the standard time of the National Time Service Center, and the trusted geographic location code is generated based on the positioning data of the BeiDou Navigation Satellite System or the Global Positioning System; The trusted timestamp and trusted geographic location encoding are used as spatiotemporal attributes of event nodes in the commodity circulation hypergraph to construct a hypergraph structure with a spatiotemporal trusted foundation.

3. The method according to claim 1, characterized in that, Step S3, which involves learning the audit rules through differentiable inductive logic programming, includes: Deep features of the multi-source behavioral data are extracted using a deep neural network. By combining the deep features with domain prior knowledge through differentiable inductive logic programming, first-order logic rules with confidence weights are learned. The rule base is incrementally updated in response to new audit cases or newly detected fraud patterns.

4. The method according to claim 1, characterized in that, Step S4 involves finding the game strategy, which includes: Based on the commodity circulation supermap, feature vectors of the current market status are extracted. These feature vectors include circulation distribution characteristics of different commodity categories, statistical characteristics of historical audit results, and emotional characteristics of consumer feedback. By using generative adversarial imitation learning, the strategy space and behavioral preferences of fraudsters are estimated based on historical fraud case data; A graph theory-based counterfactual regret minimization algorithm is used to solve the approximate Nash equilibrium strategy of the partially observable random game. Based on the approximate Nash equilibrium strategy, proactive audit instructions are generated, which include key audit targets, audit channels, and audit timing.

5. The method according to claim 1, characterized in that, Step S2, which involves constructing a spatiotemporally coupled commodity circulation hypergraph, also includes: Perceptual hash fingerprints are extracted from different modal data in the multi-source behavioral data, including product images, radio frequency identification signals, logistics codes, and consumer scanning behavior sequences. The perceptual hash fingerprint is used as the original feature vector of the product node, and then mapped to a low-dimensional hyperbolic representation through the hyperbolic space embedding.

6. The method according to claim 5, characterized in that, The multi-source behavioral data includes: The scanning behavior data generated when consumers scan a code to check the authenticity of a product using a mobile terminal includes scanning time, scanning location, scanning frequency, and scanning result feedback. The construction of the spatiotemporally coupled commodity circulation hypergraph in step S2 also includes: The consumer scanning behavior data is used as consumer-end interaction event nodes, and hyper-edge connections are established with corresponding product nodes and sales channel nodes to form a holographic map of commodity circulation that includes consumer behavior dimensions.

7. The method according to claim 1, characterized in that, The multi-source behavioral data obtained in step S1 also includes: Process parameter data and quality inspection data of the product during the production process, as well as environmental monitoring data of the product during the distribution process; In step S2, the process parameter data and environmental monitoring data are used as attribute features of event nodes to participate in the hypergraph construction.

8. The method according to claim 1, characterized in that, The rules in the evolvable rule base constructed in step S3 are stored in an interpretable first-order logic form and are associated with rule application conditions, rule confidence and rule lifecycle information; The rule base supports rule transfer learning across brands and product categories.

9. The method according to claim 1, characterized in that, After the audit results are fed back to the rule base in step S5, steps S3 and S4 are re-executed to achieve online updates of the rule base and game strategy. The online update uses an incremental learning approach, updating model parameters only based on newly added data.

10. A big data model system for anti-counterfeiting inspection of consumer products that integrates artificial intelligence and behavioral graphs, characterized in that, Configured to perform the method of any one of claims 1 to 9.

Images