Method for zero-contact operation management platform of AI intelligent adult sharing cabinet

Abstract

The invention relates to a method, a device, equipment and a medium for a zero-contact operation management platform of an AI intelligent adult shared cabinet. The method comprises the steps of obtaining a commodity reservation purchase request of a consumer terminal, analyzing a target intelligent cabinet and a target commodity, querying real-time states of the target intelligent cabinet and the target commodity through a digital twinborn model, generating a one-time dynamic token, performing homomorphic encryption binding on the one-time dynamic token and a grid physical coordinate to generate an authorization voucher, and sending the authorization voucher to the consumer terminal, and when the consumer terminal meets the near-field condition, analyzing the voucher to obtain a grid opening instruction, checking related data to obtain commodity popularity data, and inputting the commodity popularity data into a pre-trained space-time diagram neural network prediction model to obtain an intelligent operation and maintenance decision suggestion. By adopting the method, the commodity state can be sensed in real time, accurate operation and maintenance are driven by data, zero contact in the whole process is realized, and consumption safety and operation efficiency are guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of Internet e-commerce technology, and in particular relates to a method for a zero-contact operation and management platform for an AI-powered intelligent adult shared locker. Background Technology

[0002] With the integration of IoT and AI technologies, smart retail terminal technology is gradually maturing. Its core feature lies in the remote integration of merchandise management, user interaction, and equipment control through software platforms, enabling basic self-service consumption. Advances in edge computing and image recognition technologies further enable terminal devices to process visual information locally, providing new possibilities for refined operations.

[0003] In the traditional operating model, this type of consumption is mainly realized through modified vending machines or lockers. Consumers pay online and obtain an unlocking voucher, then retrieve their goods on-site. The lockers can only determine the presence of goods through simple sensors, without being able to sense their specific status. All operational aspects, including inventory checks, product cleaning, and replenishment decisions, rely entirely on regular manual inspections and experience-based judgment.

[0004] However, traditional methods suffer from two major drawbacks. First, the system is completely unable to perceive the actual physical state of goods in real time, such as packaging integrity, cleanliness, and whether they have been opened. This leads to significant hygiene risks, as problematic goods cannot be detected and handled promptly, seriously affecting consumer safety and user experience. Second, operational decisions lack data support; the timing, quantity, and product selection for replenishment rely on manual experience, failing to accurately match the dynamic demands of different locations and time periods, resulting in low inventory turnover efficiency and the coexistence of stockouts and slow-moving goods. Furthermore, necessary manual inspections contradict the original intention of zero-contact operations. These two problems together hinder the industry's development towards higher service quality and operational efficiency. Summary of the Invention

[0005] Based on this, it is necessary to provide a method, device, equipment, and medium for an AI-powered intelligent adult shared locker zero-contact operation management platform that can accurately sense the physical status of goods in real time, achieve precise replenishment and intelligent operation and maintenance decisions driven by data, and ensure zero-contact safety and controllability throughout the entire consumption process, in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a method for a zero-contact operation and management platform for AI-powered smart adult shared lockers, including:

[0007] S1. Obtain the product reservation purchase request initiated by the consumer terminal, and parse the product reservation purchase request to obtain the target smart cabinet and the target product;

[0008] S2. Query the real-time status of the target smart cabinet and the target product through the digital twin model to obtain the real-time status of the target.

[0009] S3. Based on the real-time status of the target, a one-time dynamic token is generated through an encryption algorithm;

[0010] S4. Homomorphically bind the one-time dynamic token to the physical coordinates of the grid corresponding to the target product, generate an encrypted authorization certificate, and send the encrypted authorization certificate to the consumer terminal.

[0011] S5. When the positioning information of the consumer terminal meets the near-field condition, the grid opening command is obtained by parsing the encrypted authorization certificate.

[0012] S6. Calculate the product identifier, opening timestamp, and target smart cabinet location corresponding to all compartment opening commands to obtain product popularity data; among which, product popularity data is product demand representation data based on compartment opening behavior quantification, including single product opening frequency, regional demand distribution, and time period demand peak;

[0013] S7. Input the product popularity data into the pre-trained spatiotemporal graph neural network prediction model to obtain intelligent operation and maintenance decision suggestions.

[0014] In one embodiment of the present invention, when the positioning information of the consumer terminal meets the near-field condition, after obtaining the gate opening command by parsing the encrypted authorization certificate, the method further includes:

[0015] S11. Receive synchronous multimodal data of the target product collected by the target smart cabinet. The synchronous multimodal data includes visible light images and depth point cloud data of the product.

[0016] S12. Input the visible light image of the product into the pre-trained product appearance recognition neural network model, perform target detection on the product packaging area in the visible light image of the product, and obtain the bounding box and surface anomaly mask;

[0017] S13. Based on deep point cloud data, calculate the structural deformation fraction between the target product and the standard product 3D model using a point cloud registration algorithm;

[0018] S14. Calculate the ratio of the area of ​​the surface anomaly mask to the area of ​​the bounding box to obtain the surface integrity score;

[0019] S15. Input the surface integrity score and structural deformation score into the evaluation function for fusion calculation to obtain the commodity status evaluation value;

[0020] S16. Compare the product status evaluation value with the preset status threshold to obtain the comparison result; map the corresponding maintenance requirement level based on the comparison result;

[0021] S17. Based on the maintenance requirement level, the target smart cabinet, and the target product, generate the corresponding digital maintenance instruction and send the digital maintenance instruction to the back-end administrator terminal.

[0022] Based on the above embodiments, the expression for the evaluation function is as follows:

[0023]

[0024] in, This is the product condition evaluation value. The surface integrity score. The structural deformation fraction; These are preset weighting coefficients used to adjust the contribution ratio of surface integrity and structural deformation in the final evaluation; This is the deformation influence factor, used to control the decay rate of deformation error in the evaluation function; It is a natural constant.

[0025] In one embodiment, product popularity data is input into a pre-trained spatiotemporal graph neural network prediction model to obtain intelligent operation and maintenance decision suggestions, including:

[0026] S21. Extract time-series features from product popularity data to obtain historical demand time-series feature vectors;

[0027] S22. Extract all successful locker opening records within a preset historical period from the platform database. Each locker opening record includes the smart locker identifier, product identifier, opening timestamp, and geographical location.

[0028] S23. Extract the current virtual status marker and inventory information of the target smart cabinet through a digital twin model;

[0029] S24. Integrate historical demand time-series feature vectors, current virtual state markers, and inventory information to construct a dynamic spatiotemporal feature map;

[0030] S25. Input the dynamic spatiotemporal feature map into the spatiotemporal graph neural network prediction model, and use the spatiotemporal graph neural network prediction model to model the spatiotemporal pattern of the dynamic spatiotemporal feature map to obtain the predicted demand and the strength of the correlation between commodities.

[0031] S26. Based on predicted demand, the strength of relationships between products, and the real-time status of the target, intelligent operation and maintenance decision suggestions are generated through optimization algorithms.

[0032] Based on the above embodiments, the expression for the objective function of the optimization algorithm is:

[0033]

[0034] in, For product indexing, For smart cabinet indexing, The stockout penalty coefficient per unit of goods. Commodity prediction for spatiotemporal graph neural network prediction model At the counter The future demand, For goods At the counter Current virtual inventory; The set of tasks to be processed, obtained from the digital twin model, includes cleaning tasks and recycling tasks; To carry out the mission Fixed cost coefficient, This represents a decision variable with values ​​between 0 and 1, when the task... When scheduled to be executed The value is 1 if it is not 0 otherwise. For the set of all potential delivery routes, For use path The cost coefficient, This represents a decision variable with values ​​between 0 and 1, when the path When included in the delivery plan The value is 1 if it is 1, otherwise it is 0.

[0035] In one embodiment, a one-time dynamic token is generated using an encryption algorithm based on the target's real-time state, including:

[0036] S31. Based on the real-time status of the target, obtain the order information, current timestamp, and device fingerprint of the target smart cabinet for the target product;

[0037] S32. Combine the order information, current timestamp, and device fingerprint to obtain the original signature information;

[0038] S33. Use an encryption algorithm to calculate the original signature information and generate a digital signature value;

[0039] S34. Encode the digital signature value to generate a one-time dynamic token.

[0040] Secondly, this application also provides an AI-powered intelligent adult shared locker zero-contact operation and management platform device for implementing the method described in the first aspect, including:

[0041] The product reservation parsing module is used to obtain product reservation purchase requests initiated by consumer terminals and parse the product reservation purchase requests to obtain the target smart cabinet and the target product;

[0042] The status query module is used to query the real-time status of the target smart cabinet and the target product through the digital twin model, and obtain the real-time status of the target.

[0043] The dynamic token generation module is used to generate a one-time dynamic token based on the real-time status of the target using an encryption algorithm;

[0044] The authorization certificate generation module is used to homomorphically encrypt and bind a one-time dynamic token to the physical coordinates of the grid corresponding to the target product, generate an encrypted authorization certificate, and send the encrypted authorization certificate to the consumer terminal.

[0045] The credential parsing module is used to obtain the gate opening command by parsing the encrypted authorization credential when the location information of the consumer terminal meets the near-field conditions;

[0046] The product popularity data generation module is used to calculate the product identifier, opening timestamp, and target smart cabinet location corresponding to all compartment opening commands to obtain product popularity data. Among them, the product popularity data is product demand characterization data based on compartment opening behavior quantification, including single product opening frequency, regional demand distribution, and time period demand peak.

[0047] The decision generation module is used to input product popularity data into a pre-trained spatiotemporal graph neural network prediction model to obtain intelligent operation and maintenance decision suggestions.

[0048] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect of this application.

[0049] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods in the first aspect of this application.

[0050] This application discloses a method for a zero-contact operation and management platform for AI-powered intelligent adult shared lockers. Through the synergistic effect of proprietary core technologies, it specifically addresses key issues in the background technology: Based on a digital twin model, it combines multimodal data to detect product appearance and structural deformation, accurately identifying packaging integrity and deformation status, promptly identifying problematic products, eliminating hygiene hazards imperceptible in traditional models, and improving consumer safety and experience; it calculates product popularity data such as individual product frequency, regional distribution, and peak time periods generated by locker opening behavior, and relies on a spatiotemporal neural network prediction model to mine demand patterns and product correlations at different locations and times, generating scientific replenishment and delivery plans, replacing manual experience-based judgment, solving problems of unreasonable replenishment and inefficient inventory turnover, reducing stockouts and unsold inventory, and lowering operating costs; the entire process is digitalized to achieve zero-contact operation, eliminating the need for direct human intervention, avoiding the risk of cross-infection from manual inspections and on-site interactions, aligning with the original intention of zero-contact operation, reducing reliance on manual labor, and improving operational convenience and safety. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A schematic diagram of an implementation environment provided for one embodiment of the present invention;

[0053] Figure 2 This is a flowchart of an AI-powered intelligent adult shared locker zero-contact operation management platform method in one embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the AI-powered intelligent adult shared locker zero-contact operation and management platform device in one embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] First, a brief introduction to the terms used in the embodiments of this application will be given.

[0057] C2f (Cross Stage Partial Network with 2 convolutions) is a lightweight cross-stage partial connection module used in YOLOv8. It divides the input feature map into two parts and performs convolution and cross-layer connections on each part respectively. This reduces the amount of computation while promoting gradient flow and feature reuse, effectively improving the efficiency and expressive power of the backbone network.

[0058] SPPF (Spatial Pyramid Pooling-Fast) is a fast implementation of spatial pyramid pooling. It captures and fuses contextual information at different scales in feature maps by concatenating multiple max pooling layers of different sizes, enabling the model to adapt to targets of different sizes and significantly enhancing the multi-scale perception capability of the backbone network.

[0059] FPN (Feature Pyramid Network) is a top-down feature fusion structure that upsamples and transfers rich semantic information from deep features to shallow layers, enhancing the model's ability to recognize multi-scale targets (especially small targets).

[0060] PAN (Path Aggregation Network) adds a bottom-up feature enhancement path to FPN. It feeds back and integrates the fine location and detail information of shallow features into the deep layers, thereby significantly improving the accuracy of target localization and boundary regression of the model.

[0061] The method of the AI-powered intelligent adult shared locker zero-contact operation and management platform provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, the operating terminal 100 communicates with the mobile terminal 101 and the smart cabinet 102 via a network.

[0062] In one exemplary embodiment, such as Figure 2 As shown, a method for a zero-contact operation and management platform for AI-powered smart adult shared lockers is provided, which is then applied to... Figure 1 Taking the running terminal 100 as an example, this embodiment of the method includes the following steps:

[0063] S1. Obtain the product reservation purchase request initiated by the consumer terminal, and parse the product reservation purchase request to obtain the target smart cabinet and the target product.

[0064] For example, mobile terminal 101 initiates a product reservation purchase request to operating terminal 100, and the product reservation purchase request data is transmitted via an encrypted communication network. Operating terminal 100 and mobile terminal 101 complete two-way identity authentication. After successful authentication, the request data is parsed to extract information about the target smart locker and the target product. The target smart locker corresponds to a unique code for smart locker 102, and the target product corresponds to a unique product code. Operating terminal 100 performs a legality check on the parsed results, covering the validity of the target smart locker, the ownership of the target product, and the reasonableness of the reservation time. If the check passes, the relevant information is temporarily stored and the process proceeds to subsequent steps; if the check fails, an error message is returned to mobile terminal 101.

[0065] S2. Query the real-time status of the target smart cabinet and the target product through the digital twin model to obtain the real-time status of the target.

[0066] For example, the operating terminal 100 has a built-in digital twin model, which is a virtual model mapped to the target smart cabinet 102. It has pre-established real-time data interaction links with the pressure sensor, high-definition vision sensor, temperature and humidity sensor, and door magnetic sensor on the smart cabinet 102. The operating terminal 100 sends a query command carrying the target smart cabinet code and the target product code to the digital twin model. The digital twin model immediately locates the corresponding virtual cabinet and virtual compartment based on the code information, and simultaneously calls the real-time data collected by the physical sensors bound to the virtual compartment to perform targeted status analysis. For the target product, the pressure sensor data is used to determine its presence; the image data collected by the high-definition vision sensor identifies its packaging integrity, surface cleanliness, and whether there are signs of tampering; and the temperature and humidity sensor data determines whether its storage environment meets the requirements. For the target smart cabinet 102, its overall operating status is obtained through power supply monitoring data, communication link status data, and door opening / closing data. The digital twin model summarizes and integrates the above analysis results to form target real-time status data that includes the operating status of the target smart cabinet and the physical status of the target product.

[0067] S3. Based on the target's real-time status, generate a one-time dynamic token using an encryption algorithm.

[0068] For example, the terminal 100 receives multi-dimensional target real-time status data, splices and fuses the multi-dimensional target real-time status data to obtain fused multi-dimensional information to be processed, verifies the validity of the fused multi-dimensional information to be processed, and inputs the fused multi-dimensional information that passes the verification into the following formula to obtain a one-time dynamic token:

[0069]

[0070] In the formula, It is a one-time dynamic token. For the pre-stored key, To integrate multi-dimensional information, The standardized key after key padding is... Derived from this, with the aim of adaptation. Hash block size; The inner padding byte is repeated 64 times for 0x36. The outer padding byte is repeated 64 times for 0x5C. express Hash function; This represents a byte string concatenation operation. This is an XOR operation.

[0071] S4. Homomorphically bind the one-time dynamic token to the physical coordinates of the grid corresponding to the target product, generate an encrypted authorization certificate, and send the encrypted authorization certificate to the consumer terminal.

[0072] The encrypted authorization credential can be an encrypted data packet generated by binding a one-time dynamic token with the physical coordinates of the target product compartment using a homomorphic encryption algorithm. It contains an encrypted authorization token and location information, which can be securely parsed when the consumer's terminal arrives within the vicinity of the smart locker, thereby generating a compartment opening command. This enables secure and accurate contactless pickup authorization and effectively prevents credential forgery or interception and reuse.

[0073] For example, the operating terminal 100 queries a preset grid information database based on the target product code, matches and obtains the physical coordinates of the grid corresponding to the product, and completes the association and positioning of the target product with the grid location. The operating terminal 100 uses the Paillier homomorphic encryption algorithm to perform an encrypted binding operation between the one-time dynamic token and the grid physical coordinates: the one-time dynamic token and the grid physical coordinates are converted into integer format data adapted for encryption operations, and the two types of integer data are encrypted separately using the algorithm's public key. The two encrypted ciphertexts are then fused using the homomorphic property of the algorithm's addition to obtain a fused encrypted ciphertext containing dual information. The operating terminal 100 formats the fused encrypted ciphertext, adds auxiliary verification information such as encryption algorithm identifier, ciphertext generation timestamp, and data length, and integrates them to form a complete encrypted authorization certificate. The certificate is sent to the mobile terminal 101 through an encrypted communication channel, and a timeout retransmission mechanism is initiated. If a receipt confirmation is not received within a preset time, it is automatically retransmitted; if multiple retransmissions still fail, the encryption binding process is re-executed to generate a new encrypted authorization certificate and resend it.

[0074] S5. When the positioning information of the consumer terminal meets the near-field conditions, the grid opening command is obtained by parsing the encrypted authorization certificate.

[0075] For example, the operating terminal 100 receives location information, including latitude and longitude coordinates and a location timestamp, uploaded by the mobile terminal 101 in real time, and determines whether the consumer is close to the target smart cabinet 102 based on near-field conditions. The core of the near-field condition determination is to calculate the straight-line distance between the mobile terminal 101 and the target smart cabinet 102 using the Haversine formula, and the location timestamp must meet real-time requirements. The Haversine formula is as follows:

[0076]

[0077]

[0078]

[0079] in, The difference in latitude between the mobile terminal 101 and the target smart cabinet 102; This is the difference in longitude between the two. For mobile terminals, latitude 101; 102 degrees latitude is the target smart cabinet; a is the square of the half-chord length corresponding to the spherical distance between the two points, used to convert latitude and longitude coordinates into actual distance; c is the angle formed by the ray connecting the consumer and the smart cabinet from the center of the Earth. is a two-parameter arctangent function; R is the Earth's radius; d is the straight-line distance between the mobile terminal 101 and the target smart cabinet 102.

[0080] If the calculated straight-line distance matches the near-field range and the positioning timestamp is valid, the near-field condition is deemed met; otherwise, the consumer is prompted to approach the target smart cabinet 102. After the near-field condition is met, the operating terminal 100 receives the encrypted authorization credential uploaded by the mobile terminal 101, decrypts it using the corresponding private key according to the encryption algorithm identifier in the credential, and obtains a one-time dynamic token and the physical coordinates of the compartment. After verifying the token's consistency, validity period, and coordinate matching, a compartment opening command containing the target compartment ID and opening verification code is generated and sent to the controller of the target smart cabinet 102. After the controller verifies the command, it drives the compartment to unlock and simultaneously synchronizes the opening status to the mobile terminal 101.

[0081] S6. Calculate the product identifier, opening timestamp, and target smart cabinet location corresponding to all compartment opening commands to obtain product popularity data; among which, product popularity data is product demand characterization data based on compartment opening behavior quantification, including single product opening frequency, regional demand distribution, and time period demand peak.

[0082] For example, terminal 100 establishes a data acquisition pool to collect core data corresponding to all locker opening commands in real time, covering product identification, opening timestamp, target smart locker location, opening result, and anonymized user identification, ensuring data coverage of all associated smart lockers 102. The collected data needs to be preprocessed to remove abnormal data such as failed opening records and invalid timestamps, supplement a small amount of missing valid data, and simultaneously complete timestamp standardization, mapping of smart locker locations to regional grids, and user identification hash anonymization processing.

[0083] The preprocessed data is used for quantitative calculations based on the core dimension of product popularity. The frequency of each product's launch is calculated using the following formula:

[0084]

[0085] in, Let n be the frequency of opening the grid for the i-th product identifier, and n be the total number of grid opening records for that product identifier within the statistical period. Open a record for the k-th entry of the i-th product identifier. This is an indicator function; the value of the open record is 1 when the opening result is successful and 0 when it fails.

[0086] The formula for calculating the percentage of times the device is activated is:

[0087]

[0088] in, Let m be the percentage of times the i-th product identifier is opened, and m be the total number of all product identifiers within the statistical period. The frequency of opening the i-th product identifier. The frequency of opening the single item for the j-th product identifier;

[0089] Regional demand density is calculated using the following formula:

[0090]

[0091] In the formula, Let g be the required density of the g-th region grid. This represents the total number of times the i-th product identifier was successfully opened within the g-th region grid. Let g be the area of ​​the g-th region grid.

[0092] Peak demand during a given period is calculated using the following formula:

[0093]

[0094] In the formula, is the peak demand for the i-th product identifier during a given time period; h is the hourly time period divided by a natural day; The total number of grid opening records for the i-th product identifier during the h-th time period; Open the k-th record for the i-th product identifier within the h-th time period; The function is an indicator function. The value of 1 indicates that the opening result is successful and the value of 0 indicates that the opening result is unsuccessful. This formula can be used to directly locate the period when the demand for a single product is the strongest and the corresponding number of peak times.

[0095] During the calculation process, the running terminal 100 uses a distributed computing framework to improve the efficiency of processing massive amounts of data. After completing the calculation of the frequency of single product activation, regional demand distribution, and peak demand during a time period, it summarizes and forms complete product popularity data and stores it in a time-series database.

[0096] S7. Input the product popularity data into the pre-trained spatiotemporal graph neural network prediction model to obtain intelligent operation and maintenance decision suggestions.

[0097] For example, the terminal 100 preprocesses the product popularity data using the following formula:

[0098]

[0099] In the formula, The original data, The mean, Standard deviation The data is preprocessed. The preprocessed product popularity data is used with the physical location of the smart cabinet as nodes, and the node features are the preprocessed product-smart cabinet association feature vectors. The Pearson correlation coefficient of the historical popularity of products between smart cabinets is used as the edge weight to construct a static graph structure, and then the data is connected by time slices to form a dynamic spatiotemporal graph sequence.

[0100] The pre-trained spatiotemporal graph neural network prediction model adopts a GraphSAGE-Bi-LSTM fusion architecture. The GraphSAGE layer extracts spatial correlation features through neighborhood sampling and mean aggregation, while the Bi-LSTM layer captures temporal evolution patterns and outputs spatiotemporal fusion features. Fully connected layers, combined with Dropout regularization, output prediction results. A training dataset is constructed using historical data from the past six months, divided into training and validation sets in an 8:2 ratio. The pre-trained spatiotemporal graph neural network prediction model has two GraphSAGE layers and 128 hidden units in the Bi-LSTM, using RMSE as the loss function. The expression for the loss function is:

[0101]

[0102] In the formula, For error, The total number of training samples, The total number of time steps for each sample. Indicates that the calculation is in progress. The sample is denoted by t, where t represents the time point at which the calculation of the i-th sample is performed. The model's predicted value for the i-th sample at time step t. The true value of the i-th sample at time step t; iteratively trained using the RMSprop optimizer, training stops when the error reaches the target, resulting in a pre-trained spatiotemporal graph neural network prediction model;

[0103] The extracted dynamic spatiotemporal graph sequence is input into a pre-trained spatiotemporal graph neural network prediction model. Spatial correlation features are extracted through the GraphSAGE layer in the spatiotemporal graph neural network prediction model by neighborhood sampling and mean aggregation. The following is the calculation formula for spatial correlation feature extraction:

[0104]

[0105] In the formula, Let be the feature vector of the k-th layer of node v. Let v be the set of sampling neighborhoods of node v. Let be the weight matrix of the k-th layer. This indicates the mean aggregation operation. It is the ReLU activation function. Let be the feature vector of node u at layer k-1; input the obtained spatial correlation features into the Bi-LSTM layer, and capture the temporal evolution pattern through forward LSTM and backward LSTM layers. The calculation formula is:

[0106]

[0107]

[0108] In the formula, Let t be the hidden state of the forward LSTM. Let be the hidden state of the LSTM after time t. The spatial correlation features output by the GraphSAGE layer are used. After computation, the bidirectional LSTM output features are concatenated and fused to obtain spatiotemporal fusion features. Multi-objective decision optimization is carried out through non-dominated sorting genetic algorithm. First, the spatiotemporal fusion features are input to clarify replenishment costs and stockout rates, product selection profits and turnover rates, inspection efficiency and fault risks, and set constraints such as inventory capacity, budget limit, and inspection resource quota. Then, multiple sets of decision scheme candidate sets are generated through population initialization. The Pareto front level is divided by non-dominated sorting, and the diversity of schemes is maintained by computational congestion. The selection, crossover and mutation operations are performed iteratively to optimize the solution set. Finally, the target weights are dynamically adjusted in combination with equipment health status and market demand fluctuations to select the global optimal solution and generate intelligent operation and maintenance decision suggestions covering replenishment strategies, product selection adjustment schemes and inspection plans.

[0109] This application provides a method for a zero-contact operation and management platform for AI-powered intelligent adult shared lockers. Through the synergistic effect of proprietary core technologies, it specifically addresses key issues in the background technology: Based on a digital twin model, it combines multimodal data to detect product appearance and structural deformation, accurately identifying packaging integrity and deformation status, promptly identifying problematic products, eliminating hygiene hazards imperceptible in traditional models, and improving consumer safety and experience; it calculates product popularity data such as individual product frequency, regional distribution, and peak time periods generated by locker opening behavior, and relies on a spatiotemporal neural network prediction model to mine demand patterns and product correlations at different locations and times, generating scientific replenishment and delivery plans, replacing manual experience-based judgment, solving problems of unreasonable replenishment and inefficient inventory turnover, reducing stockouts and unsold inventory, and lowering operating costs; the entire process is digitalized to achieve zero-contact operation, eliminating the need for direct human intervention, avoiding the risk of cross-infection from manual inspections and on-site interactions, aligning with the original intention of zero-contact operation, reducing reliance on manual labor, and improving operational convenience and safety.

[0110] In one embodiment of the present invention, when the positioning information of the consumer terminal meets the near-field condition, after obtaining the gate opening command by parsing the encrypted authorization certificate, the method further includes:

[0111] S11. Receive synchronous multimodal data of the target product collected by the target smart cabinet. The synchronous multimodal data includes visible light images and depth point cloud data of the product.

[0112] For example, the operating terminal 100 receives synchronous multimodal data of the target product collected by the target smart cabinet: after the smart cabinet 102 is unlocked, its high-definition visible light camera and depth camera are started simultaneously to collect synchronous multimodal data of the target product and upload it to the operating terminal 100. The data specifically includes visible light images and depth point cloud data of the product.

[0113] S12. Input the visible light image of the product into the pre-trained neural network model for product appearance recognition, perform target detection on the product packaging area in the visible light image of the product, and obtain the bounding box and surface anomaly mask.

[0114] For example, the visible light image of the product is preprocessed by size standardization and pixel normalization to ensure it meets the input requirements of the model. A pre-trained neural network model for product appearance recognition is used, which adopts a "C2f+SPPF" combination as the main module. The deep and shallow layer features are complemented by the FPN-PAN cross-layer fusion structure, and the target prediction is completed by the detection head. After the preprocessed image is input, the main module first extracts features layer by layer, and then the FPN-PAN enhances the feature expression. The detection head simultaneously outputs the bounding box of the product packaging area, the surface anomaly mask and the anomaly type. Redundant boxes are removed by non-maximum suppression and the mask is optimized by morphological operations to obtain accurate detection results.

[0115] The pre-training process of the neural network model for product appearance recognition is as follows: using images covering various product appearances as the dataset, the training set and the test set are divided; hyperparameters such as convolution kernel and learning rate are initialized; the Adam optimization algorithm is used, with the weighted sum of prediction box coordinate error, mask segmentation error and classification error as the loss function, to iteratively optimize the model parameters; after the detection accuracy is verified to meet the standard on the test set, the pre-training is completed and the model parameters are fixed.

[0116] S13. Based on deep point cloud data, calculate the structural deformation score between the target product and the standard product 3D model using a point cloud registration algorithm.

[0117] For example, the iterative nearest point registration algorithm is used to register the target product point cloud with a preset standard product 3D model. The structural deformation fraction is obtained by calculating the distance deviation between corresponding point pairs after registration using the following formula:

[0118]

[0119] in, This represents the number of valid point pairs after registration. Point cloud for target product The three-dimensional coordinates of the points These are the three-dimensional coordinates of the points corresponding to the standard model. This represents the structural deformation score; a smaller value indicates a more complete product structure. express and The Euclidean distance between two points.

[0120] S14. Calculate the ratio of the area of ​​the surface anomaly mask to the area of ​​the bounding box to obtain the surface integrity score.

[0121] For example, the pixel area of ​​the surface anomaly mask is calculated using image pixel statistics methods.

[0122] Pixel area corresponding to the bounding box The ratio of the two is the surface integrity score; the pixel area of ​​the surface anomaly mask is calculated using the following formula:

[0123]

[0124] In the formula, The width in pixels of the visible light image of the product. The height in pixels of the visible light image of the product. For surface anomaly mask in coordinates The pixel value at that location; the pixel area corresponding to the bounding box is calculated using the following formula:

[0125]

[0126] In the formula, The coordinates of the top-left corner of the bounding box are in pixels. The coordinates of the bottom right corner of the bounding box are given; the surface integrity score is calculated using the following formula:

[0127]

[0128] in, This is the surface integrity score; the closer the value is to 1, the more intact the product surface is.

[0129] S15. Input the surface integrity score and structural deformation score into the evaluation function for fusion calculation to obtain the product status evaluation value.

[0130] For example, the product condition rating is calculated using the following formula:

[0131]

[0132] in, To integrate weights, + =1, if the condition of the product depends more on surface integrity, then increase. If structural deformation has a greater impact on the product's function or preservation, then adjust the setting. , This is the product condition rating; a higher value indicates a better product condition.

[0133] S16. Compare the product status evaluation value with the preset status threshold to obtain the comparison result; map the corresponding maintenance requirement level based on the comparison result.

[0134] For example, the product status evaluation value Compared with preset state thresholds (including first-level thresholds) ,and Comparison: If If it is determined that there is no maintenance required; This is determined to be a routine maintenance requirement; if It was determined to be an urgent maintenance requirement.

[0135] S17. Based on the maintenance requirement level, the target smart cabinet, and the target product, generate the corresponding digital maintenance instruction and send the digital maintenance instruction to the back-end administrator terminal.

[0136] For example, based on the determined maintenance requirement level, combined with the unique code of the target smart cabinet, the code of the target product, and the collected multimodal data summary, a standardized digital maintenance instruction is generated. The instruction includes core information such as maintenance priority, task type, target location, and associated products, and is sent to the back-end administrator terminal through an encrypted communication channel.

[0137] The method of the AI-powered smart adult shared locker zero-contact operation management platform provided in this application embodiment synchronously collects multimodal data of goods, and uses an improved YOLOv8 model and point cloud registration algorithm to achieve accurate detection of the appearance and structural status of goods. After multi-dimensional score fusion, the status assessment and maintenance level determination are completed, and finally, accurate digital maintenance instructions are generated. This enables automated and comprehensive monitoring of the status of goods and rapid response to operation and maintenance needs, ensuring the quality of goods and the stability of operation of the smart locker.

[0138] Based on the above embodiments, the expression for the evaluation function is as follows:

[0139]

[0140] in, This is the product condition evaluation value. The surface integrity score. The structural deformation fraction; These are preset weighting coefficients used to adjust the contribution ratio of surface integrity and structural deformation in the final evaluation; This is the deformation influence factor, used to control the decay rate of deformation error in the evaluation function; It is a natural constant.

[0141] For example, the condition of a product is affected by both surface damage and structural deformation, but their mechanisms of influence differ: surface defects typically have a linear effect, while structural deformation often exhibits a non-linear deterioration trend. Therefore, this scheme employs a fusion strategy combining linear weighting and exponential decay to model the surface integrity score and structural deformation score separately, obtaining a weighted product condition evaluation value.

[0142] The evaluation function is adjusted and The value can be flexibly adapted to the evaluation standards of different product types, improving the accuracy and applicability of status judgment and providing a reliable basis for intelligent operation and maintenance.

[0143] In one embodiment of the present invention, product popularity data is input into a pre-trained spatiotemporal graph neural network prediction model to obtain intelligent operation and maintenance decision suggestions, including:

[0144] S21. Extract time-series features from product popularity data to obtain historical demand time-series feature vectors.

[0145] For example, a time-series analysis algorithm is used to perform feature mining on product popularity data. The sliding window method is used to extract demand features from different historical periods, and the historical demand time-series feature vector is generated using the following time-series feature quantification formula:

[0146]

[0147] In the formula, Let t be the historical demand time-series feature vector. Frequency of item activation during TK moments Let be the regional demand density at time tk. This represents the peak demand at time tk. , , is the feature weight, and n is the number of historical periods.

[0148] S22. Extract all successful locker opening records within a preset historical period from the platform database. Each locker opening record includes the smart locker identifier, product identifier, opening timestamp, and geographical location.

[0149] For example, the operating terminal 100 extracts all successful locker opening records within a preset historical period from the platform database: it filters valid successful locker opening records within a preset historical period from the operating platform database. Each record contains at least the core fields of smart locker identifier, product identifier, opening timestamp, and geographical location. The extracted records are deduplicated and formatted to form a structured dataset.

[0150] S23. Extract the current virtual status marker and inventory information of the target smart cabinet through a digital twin model.

[0151] For example, the digital twin model is a full-dimensional virtual mapping of the target smart cabinet. The digital twin model structure adopts a four-layer architecture: perception layer, fusion layer, mapping layer, and decision support layer. The perception layer is responsible for receiving data from physical sensors, the fusion layer cleans and fuses multi-source sensor data, the mapping layer achieves accurate mapping from the physical state to the virtual model, and the decision support layer outputs the state evaluation results. The input of the digital twin model can be real-time raw data collected by various sensing devices on the smart cabinet, such as pressure sensors, high-definition vision sensors, temperature and humidity sensors, and door magnetic sensors, covering commodity weight signals, cabinet image data, environmental temperature and humidity data, door opening and closing status data, and power supply and communication data. The output of the digital twin model includes two types of information: one is the current virtual state marker reflecting the operating status of the smart cabinet, and the other is the real-time inventory information of each commodity.

[0152] The digital twin model is trained through the following steps: collecting historical sensor data and corresponding entity status labels throughout the entire lifecycle of the smart cabinet, and dividing it into training and testing sets; using historical sensor data as input and entity status labels as output, training the multi-source data fusion model of the fusion layer and the state mapping model of the mapping layer; using the gradient descent optimization algorithm to minimize the error between the predicted state and the actual state, and iteratively optimizing the model parameters; after verifying that the model mapping accuracy meets the standards on the test set, completing the model training and deployment, ensuring that the model output data is synchronized with the real-time status of the physical smart cabinet.

[0153] S24. Integrate historical demand time-series feature vectors, current virtual state markers, and inventory information to construct a dynamic spatiotemporal feature map.

[0154] For example, a multi-source data fusion algorithm is used to achieve feature alignment and fusion through the following weight fusion formula:

[0155]

[0156] in, As a feature of fusion, , , To integrate the weights, S is the current virtual state marker, and I is the standardized value of inventory information. Using the geographical location of the smart cabinet as the node and the degree of demand correlation between nodes as the edge weight, combined with the time dimension demand change characteristics, a dynamic spatiotemporal feature map is constructed to accurately depict the spatiotemporal correlation between demand and equipment status.

[0157] S25. Input the dynamic spatiotemporal feature map into the spatiotemporal graph neural network prediction model, and use the spatiotemporal graph neural network prediction model to model the spatiotemporal pattern of the dynamic spatiotemporal feature map to obtain the predicted demand and the strength of the correlation between commodities.

[0158] Among them, the spatiotemporal graph neural network prediction model adopts a progressive architecture of spatial feature extraction, temporal feature mining, and cross-dimensional feature fusion. The spatiotemporal graph neural network prediction model has built-in graph convolutional network module, long short-term memory module, and feature fusion module.

[0159] For example, the dynamic spatiotemporal feature map is input into the graph convolutional network module of the spatiotemporal graph neural network prediction model for neighborhood aggregation. The features of the neighboring nodes of each node in the dynamic spatiotemporal feature map are weighted and aggregated using a defined graph convolutional kernel. The calculation formula is as follows:

[0160]

[0161] in, This is the original node feature matrix. The normalized adjacency matrix, , For convolution kernel weights, The bias term is ReLU, and the activation function is ReLU. The extracted spatial correlation features are output. The spatial correlation features are input into the Long Short-Term Memory (LSTM) network module. The gating mechanism of input gate, forget gate, and output gate alleviates the problem of long sequence dependence and mines the temporal evolution pattern in the historical demand time-series feature vector. Long-term time-series information is stored through cell states, and the input and forgetting of information are dynamically adjusted through gating units. The demand trend features in the time-series dimension are output. The demand trend features are then input into the feature fusion module. The spatial features output by the graph convolutional network module and the time-series features output by the LSM module are fused dimension by dimension to obtain the spatiotemporal fusion feature vector.

[0162] The spatiotemporal fusion feature vector is input into a pre-defined fully connected mapping layer. A weight matrix maps the fusion features into feature sub-vectors corresponding to the smart cabinet and product combinations. A sigmoid activation function then converts these feature sub-vectors into quantified demand prediction values. Combining node indices and product identification information from the dynamic spatiotemporal feature map, a three-dimensional prediction matrix is ​​constructed: smart cabinet number - product identification - predicted demand quantity. The predicted demand quantity of the products can be obtained from this three-dimensional prediction matrix. Feature sub-vectors corresponding to each product are extracted from the spatiotemporal fusion feature vector. A cosine similarity algorithm is used to calculate the similarity score between any two product feature sub-vectors. The similarity score calculation formula is as follows:

[0163]

[0164] In the formula, This represents the similarity score between product i and product j. Let i be the feature vector of product i. Let j be the feature vector of product j. Let L2 norm be the vector corresponding to the feature subvector of product i. Let L2 be the L2 norm of the vector corresponding to the feature subvector of product j; the strength of the association between products is obtained through similarity scores.

[0165] S26. Based on predicted demand, the strength of relationships between products, and the real-time status of the target, intelligent operation and maintenance decision suggestions are generated through optimization algorithms.

[0166] For example, the output includes the predicted demand for each smart locker's corresponding product, the product correlation strength matrix, and the target real-time status feedback from the digital twin model, such as current inventory, smart locker compartment capacity, pending tasks, and equipment status. Simultaneously, operational configuration parameters such as stockout penalty coefficient and task execution cost coefficient, as well as two types of binary decision variables—task execution and route selection—are determined. The generated intelligent operation and maintenance decision suggestions aim to reduce costs. Constraints are set such as no overflow of smart locker compartment inventory, upper limit of operation and maintenance resource capacity, full coverage of delivery routes, and reasonable values ​​for decision variables. A multi-objective optimization algorithm is used to initialize a population of feasible solutions, iteratively optimize and generate offspring, and determine convergence to select the globally optimal decision scheme with the lowest overall cost. Combined with product correlation strength optimization decisions, the precise replenishment quantity is calculated based on the gap between predicted demand and current inventory, and priorities are ranked. Task execution plans and optimal delivery routes are organized, and highly correlated products are arranged adjacently, integrating them into standardized suggestions containing decision type, execution object, parameters, and time limits.

[0167] The method of the AI-powered smart adult shared locker zero-contact operation management platform provided in this application embodiment constructs a dynamic spatiotemporal feature map by extracting temporal features and fusing multi-source data. It accurately mines the spatiotemporal evolution law of commodity demand by relying on spatiotemporal graph neural network, and achieves global optimization of operation and maintenance decision-making by combining multi-objective optimization algorithm. This effectively improves the accuracy of demand forecasting and the scientific nature of operation and maintenance decision-making, and provides core technical support for the efficient and accurate operation of smart adult shared lockers with zero contact.

[0168] Based on the above embodiments, the expression for the objective function of the optimization algorithm is:

[0169]

[0170] in, For product indexing, For smart cabinet indexing, The stockout penalty coefficient per unit of goods. Commodity prediction for spatiotemporal graph neural network prediction model At the counter The future demand, For goods At the counter Current virtual inventory; The set of tasks to be processed, obtained from the digital twin model, includes cleaning tasks and recycling tasks; To carry out the mission Fixed cost coefficient, This represents a decision variable with values ​​between 0 and 1, when the task... When scheduled to be executed The value is 1 if it is not 0 otherwise. For the set of all potential delivery routes, For use path The cost coefficient, This represents a decision variable with values ​​between 0 and 1, when the path When included in the delivery plan The value is 1 if it is 1, otherwise it is 0.

[0171] By optimizing the objective function of the algorithm, a precise balance between overall operational costs and demand satisfaction is achieved. The multi-objective optimization algorithm seeks global optimization, and the feasibility of decision implementation is ensured by combining actual constraints such as smart cabinet capacity and maintenance resources. The replenishment and display solutions are optimized by incorporating the product association characteristics, and finally a standardized closed-loop decision output is formed, providing core support for the efficient and precise implementation of zero-contact operations.

[0172] For example, based on the real-time state of the target, a one-time dynamic token is generated using an encryption algorithm, including:

[0173] S31. Based on the real-time status of the target, obtain the order information, current timestamp, and device fingerprint of the target smart cabinet.

[0174] For example, after obtaining the real-time status of the target, the operating terminal 100 synchronously retrieves the order information corresponding to the target product. The order information may include core fields such as order number, product code, and consumer anonymity identifier. At the same time, it obtains the system's millisecond-level current timestamp to ensure the validity of the token. It also extracts the device fingerprint through the hardware information collection interface of the target smart cabinet.

[0175] S32. Combine the order information, current timestamp, and device fingerprint to obtain the original signature information.

[0176] For example, the three types of data—order information, current timestamp, and device fingerprint—are standardized in format. The order information is converted into a JSON string format, the timestamp is converted into a hexadecimal string, and the device fingerprint is preserved in a hash string format. The strings are then concatenated in a fixed order of order information-timestamp-device fingerprint to form the complete original signature information.

[0177] S33. Use an encryption algorithm to calculate the original signature information and generate a digital signature value.

[0178] For example, the SHA-256 hash algorithm combined with the RSA asymmetric encryption algorithm is selected for signature calculation. First, the original signature information is hashed using the SHA-256 algorithm to obtain a 256-bit fixed-length hash value. Then, the RSA private key built into the running terminal is called to encrypt the hash value. The encryption process uses PKCS#1v1.5 padding mode to finally generate a fixed-length digital signature value to ensure that the original information cannot be tampered with.

[0179] S34. Encode the digital signature value to generate a one-time dynamic token.

[0180] For example, the digital signature value is Base64 encoded, and redundant characters such as newline characters are removed during the encoding process. At the same time, an expiration field is added to the token. The encoded signature value is concatenated with the expiration field to obtain a one-time dynamic token. This token has uniqueness, timeliness and non-forgeability, and can be used for identity authentication when opening the grid.

[0181] The method of the AI-powered smart adult shared locker zero-contact operation management platform provided in this application integrates multi-dimensional core data such as order information, timestamps, and device fingerprints. Through standardized splicing and a combination algorithm of hash + asymmetric encryption, a digital signature value with unique and tamper-proof characteristics is generated. This lays the core foundation for the secure generation of subsequent one-time dynamic tokens, while ensuring the identity legitimacy and data integrity of the token generation source.

[0182] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0183] Based on the same inventive concept, this application also provides an AI intelligent adult shared locker zero-contact operation management platform device for implementing the aforementioned AI intelligent adult shared locker zero-contact operation management platform method. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more AI intelligent adult shared locker zero-contact operation management platform device embodiments provided below can be found in the limitations of the AI ​​intelligent adult shared locker zero-contact operation management platform method described above, and will not be repeated here.

[0184] In an exemplary embodiment, as shown in FIG03, an AI-powered intelligent adult shared locker zero-contact operation management platform device is provided to implement the methods in the above-described method embodiments, including:

[0185] The product reservation parsing module 401 is used to obtain the product reservation purchase request initiated by the consumer terminal and parse the product reservation purchase request to obtain the target smart cabinet and the target product;

[0186] The status query module 402 is used to query the real-time status of the target smart cabinet and the target product through the digital twin model to obtain the real-time status of the target.

[0187] The dynamic token generation module 403 is used to generate a one-time dynamic token based on the real-time status of the target using an encryption algorithm;

[0188] The authorization certificate generation module 404 is used to homomorphically encrypt and bind a one-time dynamic token to the physical coordinates of the grid corresponding to the target product, generate an encrypted authorization certificate, and send the encrypted authorization certificate to the consumer terminal.

[0189] The credential parsing module 405 is used to obtain the grid opening command by parsing the encrypted authorization credential when the positioning information of the consumer terminal meets the near-field conditions.

[0190] The product popularity data generation module 406 is used to calculate the product identifier, opening timestamp, and target smart cabinet location of all compartment opening instructions to obtain product popularity data. The product popularity data is a product demand characterization data based on the quantification of compartment opening behavior, including the opening frequency of individual products, regional demand distribution, and peak demand during time periods.

[0191] The decision generation module 407 is used to input product popularity data into a pre-trained spatiotemporal graph neural network prediction model to obtain intelligent operation and maintenance decision suggestions.

[0192] In one embodiment of the present invention, the credential parsing module 405 can also be used for:

[0193] The system receives synchronous multimodal data of the target product collected by the target smart cabinet. The synchronous multimodal data includes visible light images and depth point cloud data of the product.

[0194] The visible light image of the product is input into a pre-trained neural network model for product appearance recognition. The model performs target detection on the product packaging area in the visible light image to obtain bounding boxes and surface anomaly masks.

[0195] Based on deep point cloud data, the structural deformation fraction between the target product and the standard product's 3D model is calculated using a point cloud registration algorithm.

[0196] The surface integrity score is obtained by calculating the ratio of the area of ​​the surface anomaly mask to the area of ​​the bounding box.

[0197] The surface integrity score and structural deformation score are input into the evaluation function for fusion calculation to obtain the product status evaluation value;

[0198] The product status evaluation value is compared with the preset status threshold to obtain the comparison result; the corresponding maintenance requirement level is obtained based on the comparison result.

[0199] Based on the maintenance requirement level, the target smart cabinet, and the target product, a corresponding digital maintenance instruction is generated and sent to the back-end administrator terminal.

[0200] Based on the above embodiments, the expression for the evaluation function is as follows:

[0201]

[0202] in, This is the product condition evaluation value. The surface integrity score. The structural deformation fraction; These are preset weighting coefficients used to adjust the contribution ratio of surface integrity and structural deformation in the final evaluation; This is the deformation influence factor, used to control the decay rate of deformation error in the evaluation function; It is a natural constant.

[0203] In one embodiment of the present invention, the decision generation module 407 can also be used for:

[0204] Extract time-series features from product popularity data to obtain historical demand time-series feature vectors;

[0205] Extract all successful locker opening records within a preset historical period from the platform database. Each locker opening record includes the smart locker identifier, product identifier, opening timestamp, and geographical location.

[0206] The current virtual status marker and inventory information of the target smart cabinet are extracted using a digital twin model.

[0207] By integrating historical demand time-series feature vectors, current virtual state markers, and inventory information, a dynamic spatiotemporal feature map is constructed.

[0208] The dynamic spatiotemporal feature map is input into the spatiotemporal graph neural network prediction model. The spatiotemporal pattern model of the dynamic spatiotemporal feature map is then used to obtain the predicted demand and the strength of the correlation between commodities.

[0209] Based on predicted demand, the strength of relationships between products, and the real-time status of targets, intelligent operation and maintenance decision suggestions are generated through optimization algorithms.

[0210] Based on the above embodiments, the expression for the objective function of the optimization algorithm is:

[0211]

[0212] in, For product indexing, For smart cabinet indexing, The stockout penalty coefficient per unit of goods. Commodity prediction for spatiotemporal graph neural network prediction model At the counter The future demand, For goods At the counter Current virtual inventory; The set of tasks to be processed, obtained from the digital twin model, includes cleaning tasks and recycling tasks; To carry out the mission Fixed cost coefficient, This represents a decision variable with values ​​between 0 and 1, when the task... When scheduled to be executed The value is 1 if it is not 0 otherwise. For the set of all potential delivery routes, For use path The cost coefficient, This represents a decision variable with values ​​between 0 and 1, when the path When included in the delivery plan The value is 1 if it is 1, otherwise it is 0.

[0213] In one embodiment of the present invention, the dynamic token generation module 403 can also be used for:

[0214] Based on the real-time status of the target, obtain the order information, current timestamp, and device fingerprint of the target smart cabinet for the target product;

[0215] The order information, current timestamp, and device fingerprint are concatenated to obtain the original signature information;

[0216] The original signature information is calculated using an encryption algorithm to generate a digital signature value;

[0217] The digital signature value is encoded to generate a one-time dynamic token.

[0218] In one embodiment, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0219] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0220] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0221] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for a zero-contact operation and management platform for AI-powered intelligent adult shared lockers, characterized in that, The method includes: S1. Obtain the product reservation purchase request initiated by the consumer terminal, and parse the product reservation purchase request to obtain the target smart cabinet and the target product; S2. Query the real-time status of the target smart cabinet and the target product through the digital twin model to obtain the real-time status of the target; S3. Based on the real-time state of the target, generate a one-time dynamic token using an encryption algorithm; S4. Homomorphically bind the one-time dynamic token to the physical coordinates of the grid corresponding to the target product to generate an encrypted authorization certificate, and send the encrypted authorization certificate to the consumer terminal; S5. When the positioning information of the consumer terminal meets the near-field condition, the grid opening command is obtained by parsing the encrypted authorization certificate; S6. Calculate the product identifier, opening timestamp, and target smart cabinet location corresponding to all the compartment opening instructions to obtain product popularity data; wherein, the product popularity data is product demand characterization data based on compartment opening behavior quantification, including single product opening frequency, regional demand distribution, and time period demand peak; S7. Input the product popularity data into a pre-trained spatiotemporal graph neural network prediction model to obtain intelligent operation and maintenance decision suggestions.

2. The method according to claim 1, characterized in that, When the location information of the consumer terminal meets the near-field condition, after obtaining the gate opening command by parsing the encrypted authorization credential, the method further includes: S11. Receive synchronous multimodal data of the target product collected by the target smart cabinet, wherein the synchronous multimodal data includes visible light images and depth point cloud data of the product; S12. Input the visible light image of the commodity into a pre-trained commodity appearance recognition neural network model, perform target detection on the commodity packaging area in the visible light image of the commodity, and obtain bounding boxes and surface anomaly masks; S13. Based on the deep point cloud data, calculate the structural deformation fraction between the target product and the standard product 3D model using a point cloud registration algorithm; S14. Calculate the ratio of the area of ​​the surface anomaly mask to the area of ​​the bounding box to obtain the surface integrity score; S15. Input the surface integrity score and the structural deformation score into the evaluation function for fusion calculation to obtain the commodity status evaluation value; S16. Compare the product status evaluation value with a preset status threshold to obtain a comparison result; map the corresponding maintenance requirement level based on the comparison result. S17. Based on the maintenance requirement level, the target smart cabinet, and the target product, generate a corresponding digital maintenance instruction and send the digital maintenance instruction to the backend administrator terminal.

3. The method according to claim 2, characterized in that, The expression for the evaluation function is: in, This is the product status evaluation value. The surface integrity score is... The structural deformation fraction; These are preset weighting coefficients used to adjust the contribution ratio of surface integrity and structural deformation in the final evaluation; This is the deformation influence factor, used to control the decay rate of deformation error in the evaluation function; It is a natural constant.

4. The method according to claim 1, characterized in that, The step of inputting the product popularity data into a pre-trained spatiotemporal graph neural network prediction model to obtain intelligent operation and maintenance decision suggestions includes: S21. Extract time-series features from the product popularity data to obtain a historical demand time-series feature vector; S22. Extract all successful lock opening records within a preset historical period from the platform database. Each lock opening record includes a smart cabinet identifier, a product identifier, an opening timestamp, and a geographical location. S23. Using the digital twin model, extract the current virtual status marker and inventory information of the target smart cabinet; S24. Integrate the historical demand time-series feature vector, the current virtual state marker, and the inventory information to construct a dynamic spatiotemporal feature map; S25. Input the dynamic spatiotemporal feature map into the spatiotemporal graph neural network prediction model, and perform spatiotemporal pattern modeling on the dynamic spatiotemporal feature map through the spatiotemporal graph neural network prediction model to obtain the predicted demand and the strength of the correlation between commodities. S26. Based on the predicted demand, the strength of the correlation between the products, and the real-time status of the target, generate intelligent operation and maintenance decision suggestions through an optimization algorithm.

5. The method according to claim 4, characterized in that, The objective function of the optimization algorithm is expressed as follows: in, For product indexing, For smart cabinet indexing, The stockout penalty coefficient per unit of goods. The goods predicted by the spatiotemporal graph neural network prediction model At the counter The future demand, For goods At the counter Current virtual inventory; The set of tasks to be processed obtained from the digital twin model includes cleaning tasks and recycling tasks; To carry out the mission Fixed cost coefficient, This represents a decision variable with values ​​between 0 and 1, when the task... When scheduled to be executed The value is 1 if it is not 0 otherwise. For the set of all potential delivery routes, For use path The cost coefficient, This represents a decision variable with values ​​between 0 and 1, when the path When included in the delivery plan The value is 1 if it is 1, otherwise it is 0.

6. The method according to claim 1, characterized in that, The step of generating a one-time dynamic token based on the real-time state of the target using an encryption algorithm includes: S31. Based on the real-time status of the target, obtain the order information, current timestamp, and device fingerprint of the target smart cabinet for the target product; S32. Concatenate the order information, the current timestamp, and the device fingerprint to obtain the original signature information; S33. Calculate the original signature information using an encryption algorithm to generate a digital signature value; S34. Encode the digital signature value to generate the one-time dynamic token.

7. An AI-powered intelligent adult shared locker zero-contact operation and management platform device, used to implement the method according to any one of claims 1 to 6, characterized in that, The device includes: The product reservation parsing module is used to obtain product reservation purchase requests initiated by consumer terminals and parse the product reservation purchase requests to obtain the target smart cabinet and the target product; The status query module is used to query the real-time status of the target smart cabinet and the target product through a digital twin model to obtain the real-time status of the target. The dynamic token generation module is used to generate a one-time dynamic token based on the real-time state of the target using an encryption algorithm; The authorization certificate generation module is used to homomorphically encrypt and bind the one-time dynamic token with the physical coordinates of the grid corresponding to the target product, generate an encrypted authorization certificate, and send the encrypted authorization certificate to the consumer terminal; The credential parsing module is used to obtain the grid opening command by parsing the encrypted authorization credential when the positioning information of the consumer terminal meets the near-field conditions; The product popularity data generation module is used to calculate the product identifier, opening timestamp, and target smart cabinet location corresponding to all the compartment opening instructions to obtain product popularity data; wherein, the product popularity data is product demand characterization data based on compartment opening behavior quantification, including single product opening frequency, regional demand distribution, and time period demand peak; The decision generation module is used to input the product popularity data into a pre-trained spatiotemporal graph neural network prediction model to obtain intelligent operation and maintenance decision suggestions.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images