A collaborative management system for integrating online and offline store operations

Abstract

This invention provides a collaborative management system for the integrated online and offline operation of stores, including a data processing module for acquiring store management data and a store management twin, and generating a store management knowledge graph; a model training module for localizing and training a standard store management model generated by a cloud server using store management data, and generating a store management model; a perception analysis module for perceiving and analyzing real-time store management data, and generating store status perception results; a perception reasoning module for perceiving and reasoning the store status perception results, and generating perception reasoning results; a strategy generation module for generating store management strategies based on the perception reasoning results; and an optimization and update module for optimizing the store management model and updating the standard store management model based on federated learning. Based on these modules, collaborative management of the integrated online and offline operation of stores is achieved.

Description

Technical Field

[0001] This invention relates to the field of store operation and management technology, and more specifically, to a collaborative management system for the integrated online and offline operation of stores. Background Technology

[0002] With the digital transformation of the retail industry and the upgrading of consumer demand for convenience, unmanned convenience stores have become an important direction in the retail field due to their core advantages of "24-hour operation, low labor costs, and high space utilization". However, the operation and management of unmanned convenience stores have certain limitations.

[0003] On the one hand, stores offer a wide variety of products and users exhibit diverse behaviors, such as taking, putting back, or swapping items, or lingering in front of a particular shelf. Environmental disturbances are also dynamic, including fluctuations in foot traffic density, electromagnetic interference, and changes in lighting. This places extremely high demands on the robustness of store status perception. On the other hand, different stores have different customer profiles, product structures, and operational pain points. Generalized management solutions are difficult to adapt to personalized needs. Furthermore, store data involves private information such as user behavior and consumption preferences, and traditional centralized data collection and homogeneous model training methods often fail to meet the management needs of various offline stores. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention provides a collaborative management system for the integrated online and offline operation of stores. This system addresses the technical problems in existing technologies, such as the high robustness requirements for store status perception and the difficulty in meeting the management needs of various offline stores due to traditional centralized data collection and homogeneous model training methods.

[0005] The purpose and effectiveness of the collaborative management system for integrated online and offline store operations of the present invention are achieved by the following specific technical means: A collaborative management system for integrating online and offline store operations includes: The data processing module is used to acquire store management data and store management twins, and generate a store management knowledge graph. The model training module is used to perform local adaptation training on the standard store management model generated by the cloud server in combination with store management data, so as to generate a store management model. The perception and analysis module is used to perceive and analyze real-time store management data and generate store status perception results. The perception reasoning module is used to perform perception reasoning on the store status perception results and generate perception reasoning results. A strategy generation module, which is used to generate store management strategies based on the results of perception reasoning; The optimization and update module is used to optimize the store management model and update the standard store management model based on federated learning.

[0006] As a further aspect of the present invention, the collaborative management system includes the following control steps: Acquire store management data and store management twins, and generate a store management knowledge graph based on the store management data and store management twins; Obtain a standard store management model and pre-train the standard store management model based on store management data to generate a store management model; Based on the store management model, real-time store management data is perceptually analyzed to obtain store status perception results, and perceptual reasoning is performed on the store status perception results based on GNN. Acquire the perception and reasoning results, generate and execute store management strategies based on the perception and reasoning results, and generate incremental learning data packages based on the store status perception results and perception and reasoning results. The store management model is optimized based on incremental learning data packages, and the incremental learning data packages are uploaded to the cloud. The standard store management model is then updated in conjunction with federated learning.

[0007] As a further aspect of the present invention, a standard store management model is obtained, and the standard store management model is pre-trained based on store management data to generate a store management model, including: Obtain a standard store management model generated by a cloud server, and use a combination of stratified sampling and entropy filtering to extract features from the store management data to generate a model adaptation training set. The model adaptation training set includes at least commodity sales fluctuation data, shelf heat distribution data, customer dwell time data, and standard process data. The standard process data refers to the standardized process data that describes the entire process of customers selecting, paying for, and verifying various types and price points of goods sold in the store. The model adaptation training set is input into the standard store management model, and the model is trained according to the Monte Carlo dropout method. The training effect is evaluated by preset validation rules. When the evaluation result meets the predetermined conditions, the training ends and the store management model is generated.

[0008] As a further aspect of the present invention, the method further includes: The store management model includes a data identification unit, a rationality verification unit, an uncertainty verification unit, and a model fine-tuning unit. The data recognition unit is used to perform behavior recognition and uncertainty calculation on the store management data input into the model, and generate initial behavior recognition results and uncertainty calculation results; The rationality verification unit is used to generate the corresponding ideal data curve based on the standard process data of the goods sold in the store. By performing deviation analysis between the store management data input to the model and the corresponding ideal data curve, and combining the initial behavior recognition results, the store anomaly judgment result is generated. The uncertainty verification unit uses deep learning and Bayesian networks to perform cognitive uncertainty, accidental uncertainty and environmental uncertainty analysis on the store management data that is judged to be abnormal in the store anomaly judgment results, and obtains the cognitive uncertainty analysis results, accidental uncertainty analysis results and environmental uncertainty analysis results. Based on the cognitive uncertainty analysis results, accidental uncertainty analysis results and environmental uncertainty analysis results, uncertainty analysis results are generated. The model fine-tuning unit is used to fine-tune the parameters of the store management model based on the results of uncertainty analysis.

[0009] As a further aspect of the present invention, a perception analysis is performed on real-time store management data based on a store management model to obtain store status perception results, and perception inference is performed on the store status perception results based on a GNN, including: The system acquires real-time store management data and inputs it into the store management model. The data recognition unit within the store management model performs data recognition on the real-time store management data and generates initial behavior recognition results and uncertainty calculation results. The rationality verification unit within the store management model combines real-time store management data and uncertainty calculation results to perform anomaly judgment on the initial behavior identification results and generate anomaly judgment results. If the anomaly judgment result is that there is no anomaly in the initial behavior recognition, then the initial behavior recognition result will be output as the store status perception result. If the anomaly judgment result indicates that the initial behavior recognition is abnormal, a store equipment control strategy is generated based on reinforcement learning. The store equipment in the abnormal area of ​​the store is controlled according to the store equipment control strategy to obtain enhanced store management data. Secondary behavior recognition is performed based on the enhanced store management data to generate behavior recognition results. The behavior recognition results are output as store status perception results. Based on the GNN, the store status perception results are used to perform perception inference and generate perception inference results.

[0010] As a further aspect of the present invention, the method further includes: For real-time store management data where the anomaly judgment result indicates an anomaly in the initial behavior recognition, the uncertainty calculation result corresponding to the real-time store management data is input into the uncertainty verification unit in the store management model for uncertainty verification, generating uncertainty analysis results. The model fine-tuning unit in the store management model then fine-tunes the store management model based on the uncertainty analysis results.

[0011] As a further aspect of the present invention, perceptual reasoning is performed on the store status perception results based on GNN to generate perceptual reasoning results, including: A store management graph neural network is constructed based on a store management knowledge graph, and the store status perception results are clustered to generate abnormal status data and non-abnormal status data. For abnormal state data, obtain the corresponding nodes and causal connection paths of the abnormal state data in the store management graph neural network, combine the corresponding historical data under the no-abnormal situation to perform first-level counterfactual reasoning to generate the minimum intervention hypothesis, and compare the minimum intervention hypothesis with the expected data generated by the store management twin. If the similarity value between the minimum intervention hypothesis and the expected data is not lower than the preset similarity threshold, then the minimum intervention hypothesis will be output as the result of perceptual reasoning. If the similarity between the minimum intervention hypothesis and the expected data is lower than the preset similarity threshold, then second-level counterfactual reasoning based on conditional generative adversarial networks and neuronormal differential equations is performed to obtain the second-level counterfactual reasoning result and output it as the perceptual reasoning result.

[0012] As a further aspect of the present invention, the perceptual reasoning results are obtained, a store management strategy is generated and executed based on the perceptual reasoning results, and an incremental learning data package is generated based on the store status perception results and the perceptual reasoning results, including: Based on the DS evidence theory and the non-dominated sorting genetic algorithm, the Pareto optimal solution set of the perceptual reasoning results is obtained to generate a candidate store management strategy set. The optimal solution is selected from the candidate store management strategy set using a sequential decision-making method to generate a store management strategy. Cluster the store status perception results to generate an abnormal store status perception cluster and a normal store status perception cluster. Obtain the perception inference results corresponding to the abnormal store status perception clusters and generate incremental learning data packages based on the perception inference results.

[0013] As a further aspect of the present invention, the store management model is optimized based on incremental learning data packages, and the incremental learning data packages are uploaded to the cloud. Combined with federated learning, the standard store management model is updated, including: The incremental learning data package is encrypted and uploaded to the cloud federated learning aggregation server. The cloud uses a secure aggregation algorithm to perform a weighted average of the incremental learning data from multiple stores and update the corresponding parameters of the standard store management model. The updated standard store management model is tested and verified, and the testing and verification includes at least performance improvement verification and robustness testing; If the standard store management model passes the test and verification, the standard store management model update is considered complete. If the standard store management model fails the test verification, return to the model update process to readjust the model parameters until it passes the test verification; The standard store management model is used as the management model for the online store platform. The online and offline collaborative management of stores is achieved through the standard store management model and the store management model that integrates the characteristics of offline stores.

[0014] As a further aspect of the present invention, the method involves acquiring store management data and a store management twin, and generating a store management knowledge graph based on the store management data and the store management twin, including: Based on a pre-set store data sensing network, data is collected and pre-processed from offline stores to obtain store management data, which includes at least product management data and shelf management data. The product management data refers to data describing the status of products, and includes at least the product name, product type, product inventory, and product placement location; The shelf management data refers to data describing the merchandise shelves, and includes at least the shelf placement location, the number of shelf layers, and the types of merchandise placed on the shelves; A digital twin is constructed using a probabilistic graphical model to obtain the state probability distribution of store management data within the store, and the store management data is mapped to the digital twin according to the state probability distribution to generate a store management twin. Feature extraction and association recognition are performed on the store management twin to obtain the store management feature vector and the corresponding association network. A store management knowledge graph is constructed based on the store management feature vectors and the association network.

[0015] Based on the above, this application embodiment realizes the acquisition of store management data and store management twins, and generates a store management knowledge graph based on the store management data and store management twins. By constructing digital twins and knowledge graphs, a unified logical framework and data foundation are provided for subsequent perception analysis and reasoning, ensuring the accuracy of interpreting the collected data. A standard store management model is acquired and pre-trained based on store management data to generate a new store management model. Based on the store management model, real-time store management data is perceptually analyzed to obtain store status perception results. Perceptual reasoning is performed on the store status perception results using GNN. Store management data is analyzed and reasoned based on multiple sub-units within the store management model. At the same time, abnormal data collection and abnormal attribution are enhanced by actively scheduling resources within the store. The model is then fine-tuned based on the corresponding data, thereby improving the robustness of the model. The system acquires perception and reasoning results, generates and executes store management strategies based on these results, generates incremental learning data packages based on the store status perception and reasoning results, optimizes the store management model using these incremental learning data packages, and uploads them to the cloud. Combined with federated learning, the standard store management model is updated. By optimizing the model and extracting knowledge locally, the store management model can be deeply adapted to the unique environment of specific stores. The cloud aggregates the wisdom of all stores to discover universal patterns in store management, thereby continuously improving the standard store management model. This allows all stores deploying this standard store management model to benefit from the collective experience of the entire network, achieving collaborative development. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the execution flow of control steps in a collaborative management system for the integrated online and offline operation of stores, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of a store management model in a collaborative management system for integrated online and offline store operations provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a collaborative management system for the integrated online and offline operation of stores, provided by an embodiment of the present invention. Detailed Implementation

[0017] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the technical solutions of the present invention, but should not be used to limit the scope of protection of the present invention.

[0018] Example: As attached Figure 1 , Figure 2 , Figure 3 As shown: This invention provides a collaborative management system for the integrated online and offline operation of stores, applicable to store management, including: The data processing module is used to acquire store management data and store management twins, and generate a store management knowledge graph. The model training module is used to perform local adaptation training on the standard store management model generated by the cloud server in combination with store management data, so as to generate a store management model. The perception and analysis module is used to perceive and analyze real-time store management data and generate store status perception results. The perception reasoning module is used to perform perception reasoning on the store status perception results and generate perception reasoning results. A strategy generation module, which is used to generate store management strategies based on the results of perception reasoning; The optimization and update module is used to optimize the store management model and update the standard store management model based on federated learning.

[0019] The specific usage and function of this embodiment are as follows: Step S10: Obtain store management data and store management twin, and generate a store management knowledge graph based on the store management data and store management twin.

[0020] In this embodiment, step S10 includes: Step S11: Obtain store management data.

[0021] Specifically, data is collected and preprocessed from offline stores based on a pre-defined store data sensing network to obtain store management data, which includes at least merchandise management data and shelf management data.

[0022] Understandably, the product management data refers to data describing the status of products, and includes at least the product name, product type, product inventory, and product placement location; the shelf management data refers to data describing product shelves, and includes at least the shelf placement location, shelf number, and product type.

[0023] In one possible embodiment, a store merchandise status coverage network is constructed based on wireless nodes deployed within the store. Channel status information corresponding to the store's shelf areas is collected based on this network. This channel status information is used to capture real-time environmental changes within the store. Specifically, actions performed by customers and movement of objects within the store will generate subtle fluctuations in the environment, which will affect the channel status information. By analyzing the phase and amplitude changes corresponding to the time-series data of the channel status information, it is possible to identify customer actions of taking and putting away merchandise. Simultaneously, a pressure sensor array pre-deployed on the shelves continuously monitors changes in shelf weight, generating weight change data. This weight change data includes the weight change of the shelf and its corresponding duration. When the detected weight change data exceeds a preset threshold, a merchandise taking and putting away event is triggered. By jointly analyzing the channel status information and the weight change data, a preliminary judgment of the merchandise status is achieved.

[0024] Understandably, the placement of goods in a store is usually fixed in a certain area. By scanning this area through a pre-built store status coverage network, wireless channel status information is collected when goods are present in the area. The collected raw wireless channel status information is then processed by machine learning algorithms, such as using convolutional neural networks, to extract features and generate a corresponding feature vector for each product. This feature vector is used as a status determination benchmark. By comparing the real-time collected wireless channel status information with its corresponding status determination benchmark, it can be determined whether a customer has taken or put back the product.

[0025] For example, by using pre-deployed Wi-Fi nodes, the wireless channel status information of a product is collected in real time. This real-time wireless channel status information is then compared with the corresponding status judgment benchmark to determine whether the product has been picked up or returned by a customer. Simultaneously, by combining this with data changes from the pressure sensor array on the shelf where the product is placed, the system determines whether the product has been picked up or returned. It's important to note that since customers picking up a product does not necessarily mean they will purchase it, simply monitoring short-term weight changes cannot determine the product's status. Therefore, the system first monitors changes in the product's wireless channel status information to determine if the product's status has changed. Then, it combines this with data from the pressure sensors to further determine the product's status, thus achieving preliminary monitoring of the product's status.

[0026] Furthermore, products whose channel status information and weight change data change are defined as products to be managed. The channel status information, weight change data, monitoring video data, and basic product information corresponding to the products to be managed are encapsulated to generate corresponding product management data. The basic product information includes at least the product name, product type, product inventory, and product placement location. At the same time, the shelf management data corresponding to the products to be managed is obtained. All product management data and shelf management data are encapsulated into store management data for output.

[0027] Understandably, the collected multi-source raw data needs to undergo spatiotemporal alignment and noise filtering before the final store management data can be formed. Spatiotemporal alignment of multi-source raw data can be accomplished through the Network Time Protocol (NTP). Noise filtering employs differentiated processing methods. For channel state information data, wavelet denoising is used. For example, a db4 wavelet basis is selected, and after decomposition into three layers, noise is removed using a soft threshold, thus preserving the signal fluctuation characteristics caused by customer actions. Weight change data uses sliding window filtering to reduce instantaneous errors caused by environmental noise. For surveillance video data, irrelevant backgrounds are removed using the region of interest (ROI) cropping method to reduce data volume. For example, only image data of the area where products interact with customers' hands are retained.

[0028] Meanwhile, the store management data is stored using a layered encryption and access control approach. The original data corresponding to the store management data is stored on a local encrypted server. The store management data generated after preprocessing is encrypted using differential privacy technology, such as injecting Laplace noise into the store management data, to ensure that the original scene in the store cannot be inferred from the store management data. At the same time, only model access is granted, and manual viewing is prohibited, providing a data foundation for subsequent model updates and optimizations based on federated learning.

[0029] Step S12: Construct a store management twin.

[0030] Specifically, a probabilistic graphical model is used to construct a digital twin, obtain the state probability distribution of store management data within the store, and map the store management data into the digital twin according to the state probability distribution to generate a store management twin.

[0031] In one possible embodiment, random variable nodes in the probabilistic graphical model are first defined. Specifically, three candidate entities—store shelves, merchandise, and customers—are extracted from store management data, and these candidate entities are used as nodes in the probabilistic graphical model. The node attributes are described by a set of multidimensional random variables, which specifically include the entity's six-degree-of-freedom state vector, existence probability variable, and semantic category variable. The six-degree-of-freedom state vector describes the entity's precise position and orientation in the three-dimensional store space; the existence probability variable is represented as a value ranging from [0,1], used to characterize the entity's position in the current scene. The probability of occurrence is considered. For example, in the scenario of predicting customer traffic in an unmanned convenience store, based on historical sales data and factors such as real-time weather and holidays, it is inferred that the probability of more than 50 customers entering the store between 3 pm and 5 pm is 0.7, meaning that the event has a 70% chance of happening in the current scenario. If the probability of a promotional item being picked up and viewed by a customer is calculated to be 0.2, it means that the item has only a 20% chance of being noticed by the customer. Semantic category variables are used to label the category to which an entity belongs, such as {entity: shelf, category: beverage shelf}, {entity: product, category: low-fat milk}, {entity: customer, category: adult male}.

[0032] Based on the product classification hierarchy, shelf layout topology, and customer-product / shelf interaction relationships stored in the pre-established store association knowledge base, semantic edges between entities are constructed. For example, a "displayed on" relationship edge is established between product entities and shelf entities, and dynamic relationship edges such as "view" and "take" are established between customers and products, thereby initializing a scene diagram that highly replicates the real store scene.

[0033] This approach combines graph attention networks (GNNs) with gated recurrent units (GRNs) to achieve efficient information exchange between nodes within the scene graph through a message passing mechanism. Specifically, each node selectively aggregates neighboring nodes based on the weights of the edges connecting it to its neighbors, and transmits information through these edges. For example, when a "take" relationship edge is detected between a product node and a customer node, the product node aggregates the customer node's behavioral characteristics, such as dwell time and the magnitude of the take-up action. Simultaneously, it combines these characteristics with the product node's inventory features and sales history, updating the product node's state via the GRN and attention network. This enables real-time replenishment demand prediction and customer behavior intention inference and prediction, such as potential purchase probability assessment. For instance, if GNN inference reveals that a product's inventory is below a preset safety threshold and customer take-up frequency has increased significantly recently, it predicts a potential stockout risk for the product.

[0034] Step S13: Construct a knowledge graph for store management.

[0035] Specifically, feature extraction and relationship identification are performed on the store management twin to obtain the store management feature vector and the corresponding relationship network. A store management knowledge graph is then constructed based on the store management feature vector and the relationship network.

[0036] In one possible embodiment, nodes within the store management twin are reused as nodes in the store management knowledge graph. Time-series event streams, such as product retrieval and return, and customer browsing events, are periodically exported from the database corresponding to the store management twin. A frequent pattern mining algorithm is used to analyze these event streams to obtain latent semantic relationships between products, and corresponding nodes are connected based on these latent semantic relationships. For example, a potential association is extracted from a time-series event stream: "Taking bread A is often accompanied by taking cola B." The support and confidence scores for this potential association are calculated to be 0.8 and 0.75, respectively. The support can be expressed as: Support (A→B) = Number of transactions containing both goods A and B / Total number of transactions. Higher support indicates that goods A and B are frequently purchased together, representing a more common, rather than accidental, combination pattern. The confidence score can be expressed as: Confidence (A→B) = Number of transactions containing both goods A and B / Number of transactions containing only goods A. This function represents the measure of rule A→B given that goods A has been purchased; that is, if goods A has been purchased... Assuming the reliability of buying B while simultaneously buying B is determined by predefined support and confidence thresholds of 0.8 and 0.7 respectively, since the support and confidence of this potential association reach the predetermined thresholds, a "complementary" edge connection is established between the nodes of bread A and cola B to represent the potential association that "taking bread A is often accompanied by taking cola B." Similarly, if there is a potential association that "when beverage A is out of stock, customers mostly choose to buy cola B," and the support and confidence of this potential association reach the predetermined thresholds, then a "substitution" edge connection is established between the beverage A node and the cola B node in the knowledge graph to represent the substitution relationship between beverage A and cola B.

[0037] It should be explained that the edge connection weights between knowledge graphs are obtained by weighted summation of their corresponding support and confidence values. The weights used in the weighted summation are preset by domain experts based on historical data. For example, if the support and confidence of a potential association are 0.8 and 0.75 respectively, and assuming the preset weight allocation ratio is 6:4, then the edge weight corresponding to this potential association is 0.8*0.6+0.75*0.4=0.78.

[0038] Understandably, a store management twin is a dynamic instance graph whose core function is to map and synchronize the real-time state of physical entities. The nodes of this graph correspond to specific entity instances, such as a specific product being picked up or a customer lingering in front of a shelf. The edges of this graph are used to represent the dynamic spatial topological relationships and instantaneous interaction relationships between entities. A knowledge graph is a static knowledge base used to describe product categories, business rules, and relationships. The nodes of this graph are conceptual entity categories and abstract business objects, while the edges are used to formally define the inherent semantic relationships and business rules between entity categories, such as cola, bread, and potato chips being complementary products, and cola B belonging to the beverage category.

[0039] Step S20: Obtain the standard store management model and pre-train the standard store management model based on the store management data to generate the store management model.

[0040] Specifically, a standard store management model generated by a cloud server is obtained. Features are extracted from the store management data using a combination of stratified sampling and entropy filtering to generate a model adaptation training set. This training set includes at least product sales fluctuation data, shelf heat distribution data, customer dwell time data, and standard process data. The standard process data describes the standardized process data corresponding to the entire process of customer selection, payment, and verification for various types and price points of goods sold in the store. The model adaptation training set is input into the standard store management model, and the model is trained using the Monte Carlo dropout method. The training effect is evaluated using preset validation rules. Training ends when the evaluation results meet predetermined conditions, thus generating the store management model.

[0041] In one possible embodiment, store management data is stratified sampling according to time, product attribute, and customer behavior dimensions. Stratified sampling of product attributes can be represented by selecting products based on three attributes: value, size, and material. Products are then categorized into high, medium, and low price levels, with 3 to 5 SKUs selected from each level. SKU represents the minimum inventory unit. For example, products with a selling price of at least 50 yuan are classified as high-value products, and products with a selling price between 10 and 50 yuan are classified as medium-value products. For low-value products, those priced below 10 yuan are classified as such. Similarly, size and material dimensions are also stratified according to predetermined thresholds, with 1 to 3 SKUs selected for each stratum. Similarly, stratified sampling for the time dimension can be represented by stratified sampling at three levels: daily, weekly, and monthly. Customer behavior dimensions are stratified based on purchase frequency and single purchase amount. During stratified sampling, an entropy filtering algorithm is used to evaluate the information entropy of the corresponding data, prioritizing features with high entropy and large information content to generate the corresponding model adaptation training set.

[0042] The generated model is adapted to the training set and input into the standard store management model. Based on Monte Carlo dropout technology and a composite loss function combining mean squared error and cross-entropy loss functions, the standard store management model is locally adapted and optimized. The mean squared error loss function is optimized for numerical data such as customer traffic prediction and inventory quantity, minimizing the squared error between the predicted and actual values. The cross-entropy loss function focuses on classification tasks such as customer consumption categories and product sales status, improving the accuracy of classification results by optimizing the category probability distribution. In this way, the joint optimization of numerical and classification tasks is achieved, thereby improving the overall performance of the model and ultimately generating a store management model adapted to the local store situation.

[0043] Understandably, the model adaptation training set includes product sales fluctuation data, shelf heat map data, customer dwell time data, and standard process data. Product sales fluctuation data records the sales changes of stratified sampled products at different times, and includes at least year-on-year and month-on-month growth rates. Shelf heat map data describes the customer flow distribution in each shelf area, including at least the pedestrian density and dwell time in that area. Customer dwell time data represents the time customers spend in different product areas, used to analyze customer shopping paths. Standard process data represents standardized operation records covering the entire process of product selection, pickup, return, and payment verification. The system records data and simultaneously labels each step with timestamps, operation types, and result status. For example, it uses channel status information and weight change data for 30 consecutive seconds under static display conditions to reflect baseline data when the product is not touched; it collects channel status information data for 5 consecutive seconds of two actions: picking up the product horizontally with one hand and picking up the product vertically with one hand, to reflect baseline data of customers performing routine picking-up actions; it collects channel status information data for 5 consecutive seconds of two actions: putting the product back horizontally with one hand and putting the product back vertically with one hand, to reflect baseline data of customers performing routine putting-back actions, and each data point must be labeled with the picking-up and putting-back action tags.

[0044] Understandably, the cloud server generates a standard store management model based on neural networks and Gaussian processes. The neural network is used to extract features from the channel state information and weight change data in the store management data, and combines the store management twin and the store management knowledge graph for feature fusion. The Gaussian process module is used to perform uncertainty analysis on the subsequently acquired store state perception results and output probability distribution data that includes at least the mean and variance.

[0045] Understandably, the store management model includes a data identification unit, a rationality verification unit, an uncertainty verification unit, and a model fine-tuning unit. The data identification unit performs behavior recognition and uncertainty calculation on the store management data input to the model, generating initial behavior recognition results and uncertainty calculation results. The rationality verification unit generates corresponding ideal data curves based on the standard process data of goods sold in the store. By performing deviation analysis between the store management data input to the model and the corresponding ideal data curves, and combining the initial behavior recognition results, it generates store anomaly judgment results. The uncertainty verification unit performs cognitive uncertainty, accidental uncertainty, and environmental uncertainty analysis on the store management data judged as abnormal in the store anomaly judgment results based on deep learning and Bayesian networks, obtaining cognitive uncertainty analysis results, accidental uncertainty analysis results, and environmental uncertainty analysis results. Based on the cognitive uncertainty analysis results, accidental uncertainty analysis results, and environmental uncertainty analysis results, it generates uncertainty analysis results. The model fine-tuning unit is used to fine-tune the parameters corresponding to the store management model according to the uncertainty analysis results.

[0046] Step S30: Based on the store management model, perform perception analysis on real-time store management data to obtain store status perception results, and perform perception inference on the store status perception results based on GNN.

[0047] In this embodiment, step S30 includes: Step S31: Based on the store management model, perform perception analysis on real-time store management data to obtain store status perception results.

[0048] Specifically, real-time store management data is acquired and input into the store management model. The data recognition module within the store management model performs data recognition on the real-time store management data, generating initial behavior recognition results and uncertainty calculation results. The rationality verification module within the store management model combines the real-time store management data and uncertainty calculation results to perform anomaly judgment on the initial behavior recognition results, generating anomaly judgment results. If the anomaly judgment result indicates that the initial behavior recognition is normal, the initial behavior recognition result is output as the store status perception result. If the anomaly judgment result indicates that the initial behavior recognition is abnormal, a store equipment control strategy is generated based on reinforcement learning. The store equipment in the abnormal area is controlled according to the store equipment control strategy to acquire reinforced store management data. Secondary behavior recognition is performed based on the reinforced store management data to generate behavior recognition results, which are then output as the store status perception results.

[0049] In this embodiment, step S31 includes: Step S31-1: Generate initial behavior recognition results and uncertainty calculation results.

[0050] In one possible embodiment, real-time store management data is input into the store management model. The data recognition module locates the approximate area of ​​the product based on the pressure sensor position corresponding to the weight change data of the current interaction event. It then retrieves all candidate product sets, the channel state information benchmark associated with the product, and the corresponding reference spatial coordinates from the store management knowledge graph. For example, if a weight change occurs on the third shelf of shelf B, the corresponding data of all products placed on the third shelf of shelf B are retrieved from the store management knowledge graph. The weighted K-nearest neighbor algorithm is used to match the real-time channel state information with the channel state information benchmark of the candidate product set. Outliers, such as abnormal signals caused by environmental interference or multiple people walking around, are eliminated by calculating the cosine similarity between the two. The coordinates of the remaining candidate reference points are then weighted and averaged for similarity. Finally, the spatial coordinates of the current interaction event and the interaction behavior type are output, such as {coordinates: (x, y, z), interaction behavior: product retrieval}.

[0051] Feature extraction is performed on real-time store management data. The extracted product images, text descriptions, and other information are transformed into high-dimensional semantic vectors. The multi-head attention mechanism in the Transformer architecture is used to process the high-dimensional semantic vectors to capture the correlation between features of different dimensions. The processed semantic vectors are then compared with the product SKUs in the store management knowledge graph using cosine similarity calculation. Combined with the spatial coordinates of the current interaction event, product matching is achieved. For example, if the similarity between the processed semantic vector and the pre-stored semantic vectors corresponding to SKU001 Cola, SKU002 Juice, and SKU003 Tea exceeds a predetermined threshold, and the product corresponding to the spatial coordinates of the current interaction event is SKU001 Cola, then the matched product is determined to be SKU001 Cola.

[0052] The multilayer perceptron processes the action timing information and weight change data in the feature data after product matching, mapping them to the probability interval [0,1]. For example, the multilayer perceptron analyzes the duration of the channel state information change caused by the customer's hand action when picking up the product, the curve of the weight change data recorded by the pressure sensor, and other data to output a probability value representing the probability of picking up the product. Finally, the matched product name, the probability of picking up the product, and the corresponding spatial coordinates are output as the initial behavior recognition result.

[0053] Uncertainty calculation results include cognitive uncertainty data, accidental uncertainty data, and environmental uncertainty data: Cognitive uncertainty data is used to characterize the uncertainty of a model due to insufficient knowledge or lack of training data. By performing multiple forward propagation sampling, a set of prediction results is generated, such as predicting the probability of product classification or predicting spatial coordinates. Then, the variance of the set of prediction results is obtained and the variance is output as cognitive uncertainty data. Random uncertainty data is used to characterize the uncertainty caused by inherent, unavoidable noise or randomness in the data itself. For example, multipath effects, electromagnetic interference, and sensor thermal noise in wireless channels are all random uncertainties. Because random uncertainty data has the above characteristics, the corresponding random uncertainty data will also be calculated simultaneously when training the store management model. Therefore, random uncertainty data is a fixed parameter value that can be directly extracted after real-time store management data is input into the store management model. Environmental uncertainty data is used to characterize the decay of perceived reliability caused by dynamic changes in the macro environment within the store. For example, the low reliability of the data output results of the store management model is caused by fluctuations in the density of people in the store, the intensity of electromagnetic interference, and the performance of equipment. By inputting the current environmental state data into a pre-trained lightweight gating attention network, the weights of each environmental feature in the environmental state data are calculated according to the attention mechanism, and a nonlinear transformation is performed through the activation function to generate an environmental uncertainty score. The environmental uncertainty score is output as environmental uncertainty data, wherein the environmental state data includes at least the real-time density of people in the store and the noise intensity of a predefined frequency band in the store.

[0054] It should be noted that the training data of the lightweight gated attention network uses historical continuous time series environmental state data, and for each training sample in the training data, there is a corresponding quantified true value that is manually labeled. The mean squared error loss is used as the loss function, and the training objective is to minimize the mean squared error loss between the output prediction value of the lightweight gated attention network and the quantified true value. The training ends when the mean squared error loss is lower than a preset threshold and the fluctuation amplitude in multiple consecutive periods is lower than the preset threshold, or when the preset maximum number of training steps is reached.

[0055] Step S31-2: Perform anomaly judgment on the initial behavior recognition result and generate anomaly judgment result.

[0056] Specifically, the rationality verification module calls upon the standard physical parameters corresponding to the products identified in the store management knowledge graph and the initial behavior recognition results, such as unit weight and size specifications. Based on the identified interaction behavior type and the product's standard physical parameters, it generates an ideal weight change curve under the current interaction behavior. Simultaneously, it calculates the DTW distance and morphological similarity between the weight change data in the store management data and the ideal weight change curve, and performs a weighted summation to generate a quantified matching confidence score. If the matching confidence score is higher than a preset threshold and the cognitive uncertainty data, accidental uncertainty data, and environmental uncertainty data in the uncertainty calculation results all reach the predetermined threshold, it is marked as a normal event, indicating that the initial behavior recognition result is not abnormal, and no subsequent uncertainty verification and model fine-tuning are performed. If the matching confidence score is lower than a preset threshold or any of the cognitive uncertainty data, accidental uncertainty data, and environmental uncertainty data in the uncertainty calculation results does not reach the predetermined threshold, it is marked as an abnormal event, indicating that the initial behavior recognition result is abnormal.

[0057] For initial behavior identification results where the anomaly judgment result is an abnormal event, the corresponding uncertainty calculation result and real-time store management data are input into the uncertainty verification module in the store management model for uncertainty verification. The uncertainty verification unit performs deep feature extraction on the real-time store management data based on deep learning to obtain a deep feature vector. The deep feature vector and the uncertainty calculation result are input into a multilayer perceptron for feature fusion to generate a joint diagnostic feature vector. The joint diagnostic feature vector is then input into a predefined Bayesian network for attribution analysis based on cognitive uncertainty, accidental uncertainty, and environmental uncertainty to obtain cognitive uncertainty analysis results, accidental uncertainty analysis results, and environmental uncertainty analysis results. An uncertainty analysis result is generated based on the cognitive uncertainty analysis results, accidental uncertainty analysis results, and environmental uncertainty analysis results. The model fine-tuning module in the store management model fine-tunes the store management model according to the uncertainty analysis results.

[0058] Understandably, the nodes of a Bayesian network can be divided into three categories: top-level nodes, intermediate-level nodes, and bottom-level nodes. Top-level nodes represent potential root causes of faults to be inferred, such as hardware drift, unknown gesture patterns, multipath interference, and environmental occlusion. Intermediate-level nodes represent observable evidence, including joint diagnostic feature vectors and uncertainty calculation results. Bottom-level nodes represent real-time store management data. Directed edges in a Bayesian network define causal relationships between nodes. For example, multipath interference nodes point to nodes with random uncertainty and abnormal channel state information spectrum; unknown gesture pattern nodes point to nodes with cognitive uncertainty and abnormal skeletal joint points. During attribution analysis, the joint diagnostic feature vector is input into the Bayesian network as observed evidence. Variational inference algorithms are then used to infer the causal relationships given the evidence. Under these conditions, the complete posterior probability P(potential fault root cause | observed evidence) of all potential fault root cause nodes is approximately calculated. The posterior probability quantifies the likelihood of each fault root cause occurring. The type of potential fault root cause with the highest posterior probability is selected as the dominant attribution result for this anomaly. If the dominant attribution result is attributed to cognitive uncertainty, such as recognizing a new gesture, incremental learning is triggered, and the corresponding data is added as a sample to the training set of the subsequently generated incremental learning package. If it is attributed to accidental uncertainty, such as diagnosing specific hardware noise, the corresponding algorithm parameters are adjusted or the specific hardware is marked as needing calibration. If it is attributed to environmental uncertainty, such as shelf vibration causing pressure sensor malfunction, the event is ignored, and the physical stability check of the corresponding shelf is triggered instead of adjusting the model parameters.

[0059] Step S32: Perform perception inference based on the GNN's perception results of the store status.

[0060] Specifically, a store management graph neural network is constructed based on a store management knowledge graph, and the store status perception results are clustered to generate abnormal state data and non-abnormal state data. For abnormal state data, the corresponding nodes and causal connection paths in the store management graph neural network are obtained. Combined with the historical data corresponding to the non-abnormal situation, first-level counterfactual reasoning is performed to generate a minimum intervention hypothesis. The minimum intervention hypothesis is compared with the expected data generated by the store management twin. If the similarity value between the minimum intervention hypothesis and the expected data is not lower than a preset similarity threshold, the minimum intervention hypothesis is output as the perception reasoning result. If the similarity value between the minimum intervention hypothesis and the expected data is lower than the preset similarity threshold, second-level counterfactual reasoning based on conditional generative adversarial networks and neural ordinary differential equations is performed, and the second-level counterfactual reasoning result is obtained and output as the perception reasoning result.

[0061] Understandably, clustering the results of store status perception means that the data corresponding to normal events determined in the anomaly judgment process of the rationality verification module will be divided into data with no abnormality status, and the data corresponding to abnormal events will be divided into data with abnormality status.

[0062] In one possible embodiment, first-level counterfactual reasoning involves automatically generating and testing multiple sets of minimal intervention strategies based on a store management knowledge graph and a store management twin. These minimal intervention strategies are used to repair abnormal events corresponding to abnormal state data. Specifically, multiple sets of the lowest-cost and easiest-to-execute minimal intervention strategies are generated based on the store management knowledge graph. Each minimal intervention strategy is simulated and executed in the store management twin. The simulated data is then compared with the baseline data in the store management twin under a normal state to calculate similarity. If the similarity exceeds a predetermined threshold, the strategy is deemed effective and output as the result of perceptual reasoning. If the similarity does not reach the predetermined threshold, it is marked as a complex abnormal event, triggering second-level reasoning. For example, the abnormal event might be "shelf B's ​​weight decreased by 250 grams, but the model did not recognize any picking action." The minimal intervention strategy generated by first-level reasoning is to restart the pressure sensor driver of shelf B. Simulation in the store management twin shows that the data flow has returned to normal, and the similarity with the baseline data under normal conditions reaches 0.97, exceeding the predetermined threshold of 0.85. At this point, the current minimal intervention strategy is deemed effective, and the restart command is immediately executed, resolving the abnormality.

[0063] For complex anomaly events, a conditional generative adversarial network (GAN) generates multiple potential root cause hypotheses, such as equipment failure, unknown user behavior, display errors, and systemic vulnerabilities. A pre-constructed potential strategy space is then used to generate corresponding root cause intervention strategies for each hypothesis. This potential strategy space is a strategy database built by the GAN generator through learning from numerous historical success cases. The long-term effects of the root cause intervention strategies are extrapolated using neural network constant differential equations. Specifically, the long-term dynamic effects of the root cause intervention strategies are extrapolated within a store management twin, including security investment costs, loss reduction rates, and impacts on customer experience. Adjoint sensitivity analysis is used to calculate the global optimization indices for each strategy, generating performance gradient vectors. These global optimization indices include economic benefit indices (maximizing profits or minimizing operating costs), operational efficiency indices (maximizing inventory turnover or minimizing stockout rates), and risk control indices (minimizing loss rates or minimizing customer complaint rates). Finally, the root cause intervention strategies and performance gradient vectors are output as the result of perceptual inference.

[0064] Understandably, the adjoint sensitivity analysis method quantifies the impact of small changes in each root cause intervention strategy on the global optimization index by solving an adjoint differential equation, thereby generating a performance gradient vector for each root cause intervention strategy, such as: [profit gradient: +2.5, cost gradient: -1.8, loss gradient: -0.5]. This performance gradient vector quantifies the optimization direction of the strategy.

[0065] Step S40: Obtain the perception reasoning results, generate and execute store management strategies based on the perception reasoning results, and generate incremental learning data packages based on the store status perception results and perception reasoning results.

[0066] Specifically, based on DS evidence theory and non-dominated sorting genetic algorithm, Pareto optimal solution set is obtained from the perception reasoning results to generate candidate store management strategy set. The optimal solution is selected from the candidate store management strategy set by sequential decision method combined with global optimization index to generate store management strategy. The store status perception results are clustered to generate abnormal clusters and normal clusters of store status perception. The perception reasoning results corresponding to the abnormal clusters of store status perception are obtained. Incremental learning data package is generated based on the perception reasoning results.

[0067] It should be noted that both steps S32 and S40 cluster the store status perception results, but their clustering objectives are different. Step S32 clusters based on the anomaly judgment results output by the rationality verification module. The final clustering result is that all stores with anomalies are clustered as anomaly status data, and all stores without anomalies are clustered as no-anomaly status data. Step S40, on the other hand, clusters the data whose anomaly judgment results are incorrect. Specifically, it uses a density clustering algorithm combined with the store anomaly judgment results, uncertainty analysis results, and corresponding store management data to cluster the store status perception results. After clustering, it uses small-sample manual review feedback to perform semantic interpretation on the identified clusters, such as random sampling and labeling the authenticity of some events. Finally, the data clusters in the store anomaly judgment results, uncertainty analysis results, and store management data that deviate significantly from the benchmark are defined as store status perception anomaly clusters. For example, data in the store anomaly judgment results with a similarity of less than 0.5 to the corresponding benchmark data are defined as store status perception normal clusters.

[0068] Step S50: The store management model is optimized based on the incremental learning data package, and the incremental learning data package is uploaded to the cloud. The standard store management model is then updated in conjunction with federated learning.

[0069] Specifically, incremental learning data packets are encrypted and uploaded to a cloud-based federated learning aggregation server. The cloud uses a secure aggregation algorithm to perform a weighted average of incremental learning data from multiple stores, thereby updating the corresponding parameters of the standard store management model. After testing and verification, the updated standard store management model is compressed and distributed to each store. Upon receiving the model, the store devices perform knowledge distillation and fusion with the local model, thereby achieving collaborative evolution and personalized adaptation of model performance. The standard store management model is used as the management model of the online store platform, and online and offline collaborative management of stores is achieved through the standard store management model and the store management model that integrates offline store characteristics.

[0070] Understandably, after updating the standard store management model, the cloud platform obtains a benchmark test set that did not participate in federated learning training. It then uses the real business scenarios of the stores and the benchmark test set to evaluate the performance of the updated standard store management model. Specifically, it quantifies and compares the performance before and after the model update by calculating accuracy, recall, and F1 score to ensure that key indicators do not significantly degrade. By introducing Gaussian noise of different intensities, simulating sensor failures, and generating adversarial samples, it comprehensively evaluates the performance degradation of the model under extreme conditions and verifies its generalization ability.

[0071] Furthermore, embodiments of the present invention also provide an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the control steps described above.

[0072] The following is a detailed introduction to the various components of the electronic device: In this context, the processor is the control center of the electronic device. It can be a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs) or one or more field-programmable gate arrays (FPGAs).

[0073] The processor can perform various functions of an electronic device by running or executing software programs stored in memory and by calling data stored in memory.

[0074] The memory is used to store the software program that executes the solution of the present invention, and the execution is controlled by the processor. The specific implementation method can be referred to the above embodiments, and will not be repeated here.

[0075] The memory can be a real-only memory (ROM) or other type of static storage device capable of storing static information and instructions, a random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only (CD-ROM), or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through an interface circuit of an electronic device; this embodiment of the invention does not specifically limit this.

[0076] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via limited means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0077] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0078] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0079] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A collaborative management system for the integrated online and offline operation of stores, characterized in that, include: The data processing module is used to acquire store management data and store management twins, and generate a store management knowledge graph. The model training module is used to perform local adaptation training on the standard store management model generated by the cloud server in combination with store management data, so as to generate a store management model. The perception and analysis module is used to perceive and analyze real-time store management data and generate store status perception results. The perception reasoning module is used to perform perception reasoning on the store status perception results and generate perception reasoning results. A strategy generation module, which is used to generate store management strategies based on the results of perception reasoning; The optimization and update module is used to optimize the store management model and update the standard store management model based on federated learning.

2. The collaborative management system for integrated online and offline store operations according to claim 1, characterized in that, The collaborative management system includes the following control steps: Acquire store management data and store management twins, and generate a store management knowledge graph based on the store management data and store management twins; Obtain a standard store management model and pre-train the standard store management model based on store management data to generate a store management model; Based on the store management model, real-time store management data is perceptually analyzed to obtain store status perception results, and perceptual reasoning is performed on the store status perception results based on GNN. Acquire the perception and reasoning results, generate and execute store management strategies based on the perception and reasoning results, and generate incremental learning data packages based on the store status perception results and perception and reasoning results. The store management model is optimized based on incremental learning data packages, and the incremental learning data packages are uploaded to the cloud. The standard store management model is then updated in conjunction with federated learning.

3. The collaborative management system for integrated online and offline store operations according to claim 2, characterized in that, Obtain a standard store management model, and pre-train the standard store management model based on store management data to generate a store management model, including: Obtain a standard store management model generated by a cloud server, and use a combination of stratified sampling and entropy filtering to extract features from the store management data to generate a model adaptation training set. The model adaptation training set includes at least commodity sales fluctuation data, shelf heat distribution data, customer dwell time data, and standard process data. The standard process data refers to the standardized process data that describes the entire process of customers selecting, paying for, and verifying various types and price points of goods sold in the store. The model adaptation training set is input into the standard store management model, and the model is trained according to the Monte Carlo dropout method. The training effect is evaluated by preset validation rules. When the evaluation result meets the predetermined conditions, the training ends and the store management model is generated.

4. The collaborative management system for integrated online and offline store operations according to claim 3, characterized in that, The method further includes: The store management model includes a data identification unit, a rationality verification unit, an uncertainty verification unit, and a model fine-tuning unit. The data recognition unit is used to perform behavior recognition and uncertainty calculation on the store management data input into the model, and generate initial behavior recognition results and uncertainty calculation results; The rationality verification unit is used to generate the corresponding ideal data curve based on the standard process data of the goods sold in the store. By performing deviation analysis between the store management data input to the model and the corresponding ideal data curve, and combining the initial behavior recognition results, the store anomaly judgment result is generated. The uncertainty verification unit uses deep learning and Bayesian networks to perform cognitive uncertainty, accidental uncertainty and environmental uncertainty analysis on the store management data that is judged to be abnormal in the store anomaly judgment results, and obtains the cognitive uncertainty analysis results, accidental uncertainty analysis results and environmental uncertainty analysis results. Based on the cognitive uncertainty analysis results, accidental uncertainty analysis results and environmental uncertainty analysis results, uncertainty analysis results are generated. The model fine-tuning unit is used to fine-tune the parameters of the store management model based on the results of uncertainty analysis.

5. The collaborative management system for integrated online and offline store operations according to claim 2, characterized in that, Based on the store management model, real-time store management data is analyzed to obtain store status perception results. Then, based on the GNN (Generative Neural Network), perception inference is performed on these store status perception results, including: The system acquires real-time store management data and inputs it into the store management model. The data recognition unit within the store management model performs data recognition on the real-time store management data and generates initial behavior recognition results and uncertainty calculation results. The rationality verification unit within the store management model combines real-time store management data and uncertainty calculation results to perform anomaly judgment on the initial behavior identification results and generate anomaly judgment results. If the anomaly judgment result is that there is no anomaly in the initial behavior recognition, then the initial behavior recognition result will be output as the store status perception result. If the anomaly judgment result indicates that the initial behavior recognition is abnormal, a store equipment control strategy is generated based on reinforcement learning. The store equipment in the abnormal area of ​​the store is controlled according to the store equipment control strategy to obtain enhanced store management data. Secondary behavior recognition is performed based on the enhanced store management data to generate behavior recognition results. The behavior recognition results are output as store status perception results. Based on the GNN, the store status perception results are used to perform perception inference and generate perception inference results.

6. The collaborative management system for integrated online and offline store operations according to claim 5, characterized in that, The method further includes: For real-time store management data where the anomaly judgment result indicates an anomaly in the initial behavior recognition, the uncertainty calculation result corresponding to the real-time store management data is input into the uncertainty verification unit in the store management model for uncertainty verification, generating uncertainty analysis results. The model fine-tuning unit in the store management model then fine-tunes the store management model based on the uncertainty analysis results.

7. A collaborative management system for integrated online and offline store operations according to claim 5, characterized in that, Based on the GNN's perception of store status, perceptual reasoning is performed to generate perceptual reasoning results, including: A store management graph neural network is constructed based on a store management knowledge graph, and the store status perception results are clustered to generate abnormal status data and non-abnormal status data. For abnormal state data, obtain the corresponding nodes and causal connection paths of the abnormal state data in the store management graph neural network, combine the corresponding historical data under the condition of no abnormality to perform first-level counterfactual reasoning to generate the minimum intervention hypothesis, and compare the minimum intervention hypothesis with the expected data generated by the store management twin. If the similarity value between the minimum intervention hypothesis and the expected data is not lower than the preset similarity threshold, then the minimum intervention hypothesis will be output as the result of perceptual reasoning. If the similarity between the minimum intervention hypothesis and the expected data is lower than the preset similarity threshold, then second-level counterfactual reasoning based on conditional generative adversarial networks and neuronormal differential equations is performed to obtain the second-level counterfactual reasoning result and output it as the perceptual reasoning result.

8. The collaborative management system for integrated online and offline store operations according to claim 2, characterized in that, Obtain the perceptual reasoning results, generate and execute store management strategies based on the perceptual reasoning results, and generate incremental learning data packages based on the store status perception results and perceptual reasoning results, including: Based on the DS evidence theory and the non-dominated sorting genetic algorithm, the Pareto optimal solution set of the perceptual reasoning results is obtained to generate a candidate store management strategy set. The optimal solution is selected from the candidate store management strategy set using a sequential decision-making method to generate a store management strategy. Cluster the store status perception results to generate an abnormal store status perception cluster and a normal store status perception cluster. Obtain the perception inference results corresponding to the abnormal store status perception clusters, and generate incremental learning data packages based on the perception inference results.

9. A collaborative management system for integrated online and offline store operations according to claim 2, characterized in that, The store management model is optimized based on incremental learning data packages, which are then uploaded to the cloud. Combined with federated learning, the standard store management model is updated, including: The incremental learning data package is encrypted and uploaded to the cloud-based federated learning aggregation server. The cloud uses a secure aggregation algorithm to perform a weighted average of the incremental learning data from multiple stores and update the corresponding parameters of the standard store management model. The updated standard store management model is tested and verified, and the testing and verification includes at least performance improvement verification and robustness testing; If the standard store management model passes the test and verification, the standard store management model update is considered complete. If the standard store management model fails the test verification, return to the model update process to readjust the model parameters until it passes the test verification; The standard store management model is used as the management model for the online store platform. The online and offline collaborative management of stores is achieved through the standard store management model and the store management model that integrates the characteristics of offline stores.

10. A collaborative management system for integrated online and offline store operations according to claim 2, characterized in that, Acquire store management data and a store management twin; generate a store management knowledge graph based on the store management data and the store management twin, including: Based on a pre-set store data sensing network, data is collected and pre-processed from offline stores to obtain store management data, which includes at least product management data and shelf management data. The product management data refers to data describing the status of products, and includes at least the product name, product type, product inventory, and product placement location; The shelf management data refers to data describing the merchandise shelves, and includes at least the shelf placement location, the number of shelf layers, and the types of merchandise placed on the shelves; A digital twin is constructed using a probabilistic graphical model to obtain the state probability distribution of store management data within the store, and the store management data is mapped to the digital twin according to the state probability distribution to generate a store management twin. Feature extraction and association recognition are performed on the store management twin to obtain the store management feature vector and the corresponding association network. A store management knowledge graph is constructed based on the store management feature vectors and the association network.

Images