Intelligent selling system and method based on Internet platform

Abstract

The invention discloses an intelligent selling system and method based on an internet platform, and the method comprises the steps: constructing a high-dimensional state embedding vector which can capture the spatial dependence relation between global devices through the deep mining of the static attributes, dynamic flow and external environment data of a selling machine; on the basis, a demand curve and a commodity decay risk value at a future moment are predicted in parallel; and further, the benefit and the loss are comprehensively balanced by using a collaborative decision-making mechanism, and a suggested replenishment time window considering a logistics response period is output. Through the mode, the problem of replenishment lag or blind stockpiling caused by static threshold management is effectively solved, and the expiration rejection rate and stockout opportunity cost of fresh food commodities are remarkably reduced, so that the optimal cooperation of fine management, profit maximization and operation efficiency of an unmanned retail network in a complex dynamic environment is realized.

Description

Technical Field

[0001] This application relates to the field of smart vending, and more specifically, to a smart vending system and method based on an internet platform. Background Technology

[0002] With the deep integration of IoT technology and new retail models, smart vending systems based on internet platforms have become the mainstream form of modern unmanned retail. These systems aim to break the time and space limitations of traditional retail through digitalization, improving operational efficiency and user experience. Typically, these systems utilize built-in sensors to collect real-time data on inventory levels, equipment operating status, and sales records for each aisle. This data is then uploaded to a cloud-based backend via a communication module, allowing operators to comprehensively monitor and remotely manage vending machines through various terminal devices without needing to be physically present. Furthermore, they can trigger replenishment notifications when inventory falls below a preset threshold.

[0003] However, while existing cloud platform solutions have achieved some degree of visualization of operational data and basic remote control, their technological limitations are becoming increasingly apparent when dealing with perishable goods with short shelf lives (such as boxed meals and salads). Current technologies largely rely on static inventory thresholds or simple human experience for replenishment decisions, lacking in-depth mining and correlation analysis of multimodal data (such as weather, pedestrian traffic, and product lifecycle). In actual operation, this passive management model struggles to address the coordination issue between sales forecasting and spoilage risk assessment: if the replenishment window lags behind the demand curve, it can lead to the instant depletion of inventory for best-selling items before replenishment personnel can arrive in time, resulting in significant lost sales opportunities; conversely, if spoilage risks are not accurately predicted and replenishment is done blindly, goods can easily expire and become unusable in the cabinets.

[0004] Therefore, we look forward to an optimized smart vending system based on an internet platform. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides an intelligent vending system and method based on an internet platform.

[0006] According to one aspect of this application, an intelligent vending system based on an internet platform is provided, comprising: The data acquisition module is used to acquire static data, dynamic data, and external environment data; The multimodal feature preprocessing module is used to perform multimodal feature preprocessing on static data, dynamic data, and external environment data to obtain a sales status feature vector sequence; The global state high-dimensional embedding module is used to generate a global vending machine state high-dimensional embedding vector sequence based on the vending machine state feature vector sequence and vending machine location coordinates; The sales forecasting and spoilage risk assessment module is used to perform sales forecasting and spoilage risk assessment on the high-dimensional embedded vector sequence of global vending machine status to obtain the demand forecast curve and the estimated loss risk value. The collaborative decision-making module is used to make collaborative decisions based on the demand forecast curve and the estimated loss risk value to obtain a suggested replenishment time window.

[0007] According to another aspect of this application, a smart sales method based on an internet platform is provided, comprising: Acquire static data, dynamic data, and external environment data; Multimodal feature preprocessing is performed on static data, dynamic data, and external environment data to obtain a sequence of sales status feature vectors; Based on the vending status feature vector sequence and vending machine location coordinates, a global vending machine status high-dimensional embedding vector sequence is generated; Sales forecasting and spoilage risk assessment are performed on the high-dimensional embedded vector sequence of global vending machine status to obtain demand forecast curves and estimated loss risk values. A collaborative decision is made based on the demand forecast curve and the estimated loss risk value to obtain a suggested replenishment window.

[0008] Compared with existing technologies, this application provides an intelligent vending system and method based on an internet platform. It constructs a high-dimensional state embedding vector by deeply mining the static attributes, dynamic flow, and external environmental data of vending machines. This vector captures the spatial dependencies between global devices, and uses it to predict future demand curves and product spoilage risk values ​​in parallel. Furthermore, a collaborative decision-making mechanism comprehensively weighs revenue and losses, outputting suggested replenishment time windows that take into account logistics response cycles. This approach effectively solves the problems of delayed replenishment or blind stockpiling caused by static threshold management, significantly reducing the expiration and spoilage rate of fresh food products and the opportunity cost of stockouts. Thus, it achieves optimal synergy between refined management, profit maximization, and operational efficiency in unmanned retail networks under complex and dynamic environments. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a block diagram of an intelligent vending system based on an internet platform according to an embodiment of this application; Figure 2This is a schematic diagram of data flow in an intelligent vending system based on an internet platform according to an embodiment of this application; Figure 3 This is a flowchart of an intelligent vending method based on an internet platform according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] The technical solution of this application proposes an intelligent vending system based on an Internet platform. Figure 1 This is a block diagram of an intelligent vending system based on an internet platform according to an embodiment of this application. Figure 2 This is a system architecture diagram of an intelligent vending system based on an internet platform according to an embodiment of this application. Figure 1 and Figure 2As shown, the intelligent vending system 300 based on an internet platform according to an embodiment of this application includes: a data acquisition module 310, used to acquire static data, dynamic data, and external environment data; a multimodal feature preprocessing module 320, used to perform multimodal feature preprocessing on the static data, dynamic data, and external environment data to obtain a sales status feature vector sequence; a global state high-dimensional embedding module 330, used to generate a global vending machine state high-dimensional embedding vector sequence based on the sales status feature vector sequence and the vending machine location coordinates; a sales forecasting and spoilage risk assessment module 340, used to perform sales forecasting and spoilage risk assessment on the global vending machine state high-dimensional embedding vector sequence to obtain a demand forecast curve and an estimated loss risk value; and a collaborative decision-making module 350, used to perform collaborative decision-making on the demand forecast curve and the estimated loss risk value to obtain a suggested replenishment time window.

[0016] Specifically, the data acquisition module 310 is used to acquire static data, dynamic data, and external environment data. It should be understood that the core value of the intelligent vending system of this application lies in achieving refined operation through data-driven approaches, namely, using deep learning algorithms to balance profit maximization and loss minimization. To achieve this goal, the technical solution of this application first acquires static data, dynamic data, and external environment data to transform the physical state of the vending machine, transaction behavior, and its environmental context into digital signals that can be processed by a computer. This provides a reliable input basis for subsequent feature preprocessing, high-dimensional embedding vector generation, and complex sales forecasting and spoilage risk assessment, thereby avoiding the problems of delayed replenishment or blind stockpiling caused by information silos in traditional operating models. Static data is the basic information describing the inherent attributes of the vending machine, including the vending machine ID, geographical coordinates, and a basic information table of product SKUs. It provides anchor points for identity and location; the geographical coordinates are not only used for map display but also serve as index keys for subsequently acquiring weather and pedestrian flow data for specific areas. The product SKU information defines the basic capabilities for selling. Dynamic data refers to time-series data reflecting the real-time operating status of vending machines. This includes real-time sales logs, inventory sensor logs, and heartbeat data from the vending machines. Real-time sales logs directly reflect market demand, inventory sensor logs provide physical feedback on current remaining stock, and heartbeat data monitors the online status and health of the equipment, ensuring data continuity. External environmental data refers to external contextual variables that influence consumer behavior. This includes minute-level weather data for the machine's location from third-party APIs, hourly pedestrian traffic heatmaps, and a pre-built library of holiday / event tags. It provides contextual features; for example, weather and pedestrian traffic directly affect the willingness to purchase fresh food. By introducing these variables, the system can capture demand fluctuations that cannot be explained by historical sales data alone.

[0017] In practice, the collection of both static and dynamic data primarily relies on the hardware of the vending machine terminal. The vending machine uses built-in sensors and systems to collect its own operational data in real time, such as inventory levels in each compartment, equipment operating status (e.g., temperature, fault information), and sales records. This data is then encrypted and uploaded to a back-end management system hosted on a cloud server (e.g., Alibaba Cloud). This process ensures data real-time performance and security, allowing operators to monitor the system without being physically present, using any internet-connected device. External environmental data is primarily collected by the cloud-based back-end management system. Based on the vending machine's geographical coordinates, the system periodically sends application programming interface (API) requests to third-party data service providers to retrieve real-time weather and pedestrian flow heatmaps for the area, and combines this with a pre-built database to retrieve holiday information. This data from various sources is uniformly aggregated and timestamped in the cloud, providing a data foundation for generating a time-aligned raw dataset.

[0018] Specifically, the multimodal feature preprocessing module 320 is used to perform multimodal feature preprocessing on static data, dynamic data, and external environment data to obtain a sales status feature vector sequence. It should be understood that although the system can collect massive amounts of raw data from the cloud and terminals, this data is inherently multi-source, heterogeneous, and unstructured. Specifically, raw sensor logs, JSON data returned by weather APIs, and static SKU tables differ significantly in time granularity, units of measurement, and data distribution; directly inputting them into a deep learning model can lead to gradient explosion or failure to converge. Furthermore, for fresh food products, simple inventory quantity cannot reflect their value decay; the concept of a lifecycle must be introduced. Therefore, in the technical solution of this application, multimodal feature preprocessing is performed on static data, dynamic data, and external environment data to transform the chaotic raw signals into model-understandable, aligned, and normalized numerical tensors through standardized mathematical transformations, laying the foundation for subsequently generating a high-dimensional embedding vector sequence of global vending machine status.

[0019] In practice, the first step is to perform time-series alignment on static data, dynamic data, and external environment data to obtain a time-series aligned original dataset. Since timestamps from different data sources may not be perfectly synchronized—for example, sales records are event-triggered, while sensor logs and external weather data may be collected periodically—the technical solution in this application resamples and aligns static data, dynamic data, and external environment data along a unified time axis to obtain a time-series aligned original dataset. Specifically, by setting a basic time interval (e.g., 15 minutes), all data is aggregated or interpolated to this unified time point, ensuring that all features have a consistent correspondence in the time dimension.

[0020] Next, the time-aligned original dataset is subjected to inventory unit lifecycle quantification to obtain an inventory dataset with quantified lifecycles. That is, for the special handling of perishable goods management, the system not only records the inventory quantity but also calculates the remaining lifespan or spoilage risk of each inventory unit based on the shelf-life information in the static data and the current time. For example, freshness can be quantified by calculating the ratio of the difference between the current time and the production date to the total shelf life, a process expressed by the formula:

[0021] in, Indicates at time Product freshness factor This refers to the expiration date of the product. This represents the total shelf life of the product. Using this formula, the system converts absolute time into a relative value reflecting the product's condition, thus obtaining a quantified inventory dataset representing its lifecycle.

[0022] Furthermore, multimodal feature fusion and normalization are performed on the quantified lifecycle inventory dataset to obtain a sales status feature vector sequence. That is, all features in the previously processed dataset (including aligned original features and quantified lifecycle features) are merged and transformed into a single numerical vector, which is then normalized to obtain the final sales status feature vector sequence. Specifically, for each vending machine at each aligned time point, all its static attributes (such as geolocation-encoded features), dynamic indicators (such as sales revenue, inventory level, and percentage of remaining shelf life), and external environmental indicators (such as temperature, foot traffic index, and whether it is a holiday) are concatenated into a high-dimensional original feature vector. Subsequently, to eliminate the impact of differences in feature dimensions on model training, this original feature vector needs to be normalized, for example, using Z-score normalization to make its mean 0 and standard deviation 1.

[0023] Specifically, the global state high-dimensional embedding module 330 is used to generate a global vending machine state high-dimensional embedding vector sequence based on the sales state feature vector sequence and the vending machine location coordinates. It should be understood that smart vending machines are not isolated entities; they have complex interdependencies in physical space and consumer demand. Relying solely on the characteristics of a single machine (such as individual machine inventory or sales volume) often fails to capture systemic risks or opportunities at the network level, such as the spillover effect (siphon effect) caused by stockouts at neighboring machines or regional surges in foot traffic. Therefore, in the technical solution of this application, the sales state feature vector sequence obtained in the previous step, which only describes the state of a single device, is combined with the vending machine location coordinates describing the relative positions between devices to construct a dynamic vending machine relationship graph. Then, using spatial feature aggregation techniques such as graph convolutional networks, a global vending machine state high-dimensional embedding vector is generated for each vending machine, which not only contains its own state information but also incorporates its influence from neighboring nodes throughout the network. These vectors, arranged in chronological order, constitute a high-dimensional embedding vector sequence of the global vending machine status. This sequence can more profoundly and comprehensively reflect the true status of each device in the global network, providing high-quality input for subsequent accurate prediction and risk assessment.

[0024] In practice, the first step is to extract historical sales time-series data from the sales status feature vector sequence. That is, from the pre-processed sales status feature vector sequence, specific feature dimensions directly related to sales performance are extracted, such as sales revenue and sales volume within a past time window, forming independent historical sales time-series data for each vending machine. This step is for subsequent analysis of sales correlations between devices.

[0025] Next, a weighted adjacency matrix is ​​constructed based on historical sales time-series data and vending machine location coordinates. It should be understood that the effectiveness of a graph convolutional network highly depends on whether its input adjacency matrix can truly and accurately reflect the inherent, meaningful connections between nodes. In the context of smart vending, the association between two vending machines is not a simple physical connection, but is determined by a variety of complex factors, the two most important of which are spatial proximity and behavioral similarity. Considering only geographical distance ignores the complementary or competitive relationships of sales patterns; while considering only sales relevance may forcibly associate geographically unrelated devices with similar sales curves merely by chance, leading the model to learn false patterns. Therefore, in the technical solution of this application, a weighted adjacency matrix is ​​constructed to systematically integrate the geographical proximity information contained in the vending machine location coordinates and the sales relevance information contained in the historical sales time-series data, thereby constructing a weighted adjacency matrix that can more comprehensively and reasonably represent the true topology of the vending machine network, laying the foundation for subsequent accurate global state embedding.

[0026] In this process, firstly, a sales correlation weight matrix is ​​generated based on historical sales time-series data. Specifically, this is achieved by calculating the statistical correlation measure between the historical sales time series of every two vending machines in the network. For example, the correlation can be calculated for each pair of vending machines. in the past Sales sequence at a specific point in time and The Pearson correlation coefficient is used to obtain a sales correlation weight matrix, where each element of the matrix represents the vending machine. and The degree of similarity in sales models. Specifically, the formula for calculating the Pearson correlation coefficient is:

[0027] in, and These are sequences and The mean.

[0028] Next, a geographic proximity weight matrix is ​​generated based on the vending machine location coordinates. Specifically, the geographic distance between vending machines is calculated based on their latitude and longitude coordinates, and then the distance is converted into weights representing the degree of proximity. For example, Euclidean distance is first calculated, and then a decay function (such as a Gaussian kernel function) is used to map the distance to a weight value between 0 and 1, with closer distances receiving higher weights. An element of the geographic proximity weight matrix can be calculated as follows:

[0029] in, Indicates vending machine and Geographical distance between them The bandwidth parameter used to control the decay rate; Furthermore, a dynamic relationship gating matrix is ​​determined based on the sales status feature vector sequence. It should be understood that the core flaw in traditional graph construction mechanisms, when performing weight fusion and adjacency matrix generation, lies in the static nature and global homogeneity of the fusion strategy. Specifically, it linearly combines geographical proximity weights and historical sales relevance weights using a globally shared learnable scalar parameter. This approach fails to fully capture the complex dynamics of relationships between nodes in intelligent sales scenarios. Firstly, this mechanism lacks contextual dynamism, incorrectly assuming that the relative importance of geographical location and sales model is constant. However, in real-world operations, this importance fluctuates dramatically with time, events, and other external factors. For example, two geographically close devices may exhibit strong competition for commuter traffic during weekday morning rush hour, where geographical factors should dominate. However, on weekends, their correlation may be more determined by sales relevance due to shared promotional activities. A fixed, globally shared learnable scalar parameter cannot adapt to such changes. Secondly, this mechanism ignores the heterogeneity and special relationships between nodes, applying the same fusion standard to all node pairs in the network. However, vending machines with different functions have different interaction patterns. For example, there is a temporal complementary relationship between lunch box machines and afternoon tea beverage machines in an office building. This is fundamentally different from the competitive relationship between two similar beverage machines. The globally unified fusion strategy obviously cannot accurately depict the unique association pattern of each pair of nodes at a specific moment, thus limiting the true expressive power of the graph model.

[0030] To address the aforementioned issues, this application proposes an optimization mechanism that replaces the original static fusion coefficients with an adaptive relational attention gating mechanism, thereby enabling dynamic and differentiated construction of the association weights between nodes.

[0031] Specifically, firstly, the node-pair joint state representation is performed on the vending state feature vector sequence to obtain the node-pair joint state vector sequence. That is, in order to dynamically determine the relationship between any two vending machine nodes, comprehensive information that fully describes their current interaction state is first obtained. Specifically, at any time t, the feature vectors of node i and node j are extracted from the input state feature vector sequence. (Eigenvector of node i at time t) and (The feature vector of node j at time t), and through vector concatenation, construct a representation vector that reflects the node's state in relation to the joint state. This process can be expressed by the following formula:

[0032] in, For the joint state vector of node pairs: a vector of dimension 2F, where F is the dimension of the feature vector of a single node, which captures the instantaneous combination of the states of nodes i and j at time t; Let be the feature vectors of nodes i and j at time t: represent the F-dimensional feature vectors of vending machines i and j at time t, respectively; Vector concatenation operation: This means joining two vectors end-to-end to form a longer vector. This step provides rich contextual information for subsequent intelligent decision-making; for example, if... The characteristics indicate that machine i's inventory is about to run out, while The characteristics indicate that machine j has sufficient inventory. This allows for the explicit encoding of key information regarding this inventory imbalance.

[0033] Furthermore, gating modeling is applied to the joint state vector sequence of node pairs to obtain a dynamic relation gating matrix. It should be understood that static fusion weights cannot adapt to dynamic changes in relations. Therefore, in the technical solution of this application, gating modeling is applied to the joint state vector sequence of node pairs to intelligently generate a dynamic fusion weight for a specific node pair. Specifically, the joint state vector of node pairs generated in the previous step... The input is fed into a gating unit consisting of two fully connected layers. This network performs a nonlinear transformation, ultimately outputting a gating factor between 0 and 1. That is, dynamic fusion weights, which can be expressed by the formula:

[0034] in, For dynamic relationship gating factor: a scalar value between 0 and 1, representing the degree of importance of geographical proximity to the relationship between node pair (i,j) at time t; , , and : Learnable parameters of the gated neural network, automatically optimized through model training; ReLU: Modified activation function of linear units, providing nonlinear modeling capabilities; The Sigmoid activation function smoothly maps the output value to the (0, 1) interval. This step constructs a relational expert capable of understanding complex scenarios. Its learnable parameters allow the model to learn to recognize specific state patterns during training. For example, when the model recognizes the combination of the features that machine i's product has a very short remaining shelf life and machine j's similar product has just been restocked, it might output a value close to 1. The value indicates that at this moment, they are primarily engaged in geographical competition due to vying for customers seeking novel products. This allows for the generation of a dynamic, heterogeneous, and context-aware gating factor. This forms the core of the subsequent construction of the adaptive graph structure.

[0035] Subsequently, based on the dynamic relational gating matrix, the geographic proximity weight matrix and the sales relevance weight matrix are fused to obtain a weighted adjacency matrix. That is, after obtaining the dynamic gating factor customized for each pair of nodes at each time step, this factor is applied to the static geographic and sales relevance weights to generate the final graph structure that reflects the true relationships within the current network. Specifically, the gating factor for all node pairs at time t is... Combined into a dynamic relational gating matrix And use this as the weight to weight the static geographic proximity weight matrix. Relevance weight matrix to sales Perform element-wise weighted summation to obtain the time-varying weighted adjacency matrix that takes effect at time t. This process can be expressed by the following formula:

[0036] in, Weighted adjacency matrix: An N x N matrix whose elements This represents the combined correlation strength between vending machines i and j at time t; For dynamic relational gating matrix: by The resulting N x N matrix; and These represent the statically calculated weight matrices for geographic proximity and sales relevance, respectively; they are all-one matrices: an N x N matrix where all elements are 1. Element-wise multiplication / Hadamard product: This represents the element-wise multiplication of corresponding positions in two matrices. This step allows the graph's topology to evolve adaptively with changes in the vending machine network's state. For example, during large events, the association weights of all devices within a relevant area may be dynamically adjusted to reflect shared impacts. Ultimately, this completely replaces the original static matrices, providing downstream graph convolutional networks with inputs that more accurately reflect the complex dynamic interactions of real-world vending machine networks, thus achieving a fundamental optimization of the graph structure.

[0037] In summary, this optimization mechanism overcomes the limitations of the original mechanism's static, globally homogeneous fusion strategy. Specifically, by introducing an adaptive relational attention gating mechanism, the adjacency matrix of the vending machine network graph can be dynamically adjusted at each time step based on the real-time joint state of each pair of nodes. This allows the model to distinguish and quantify the differences and time-varying nature of association patterns in different scenarios (such as weekdays vs. weekends, promotional periods vs. non-promotional periods) and between different node pairs (such as competitive and complementary relationships). Ultimately, the model can more accurately and profoundly capture the complex and dynamic interdependencies within the smart vending machine network, providing higher-quality and more information-rich input for subsequent downstream tasks such as sales forecasting and risk assessment. This significantly improves the overall solution's predictive accuracy and decision-making effectiveness, achieving the ultimate business goals of refined operation and profit maximization.

[0038] Furthermore, graph convolutional spatial feature aggregation is performed on the sales status feature vector sequence and the weighted adjacency matrix to obtain a high-dimensional embedding vector sequence of the global vending machine status. It should be understood that smart vending machines are not isolated in physical space and business logic; the operational status of each machine is significantly affected by surrounding machines, such as the spillover effect (siphon effect) caused by stockouts in neighboring machines or the flow of people in a specific area. The cloud platform aims to perform in-depth analysis of accumulated sales data to help operators understand consumption trends and achieve data-driven refined operations. If prediction is based solely on the feature vectors of a single machine, the model will fail to capture these implicit relationships across devices, leading to decision bias. Therefore, in the technical solution of this application, graph convolutional spatial feature aggregation is further performed on the sales status feature vector sequence and the weighted adjacency matrix. Utilizing the powerful message passing mechanism of Graph Convolutional Networks (GCNs), the network topology defined by the weighted adjacency matrix is ​​deeply fused with the local features of the nodes to generate a high-dimensional embedding vector that can represent the global context, allowing the digital twin model of each machine to perceive its surrounding state.

[0039] Among them, graph convolutional spatial feature aggregation refers to the operation of feature extraction on graph data with non-Euclidean structure. It enables each node to aggregate information of its first-order or multi-order neighbors through a message passing mechanism.

[0040] In this process, firstly, the input weighted adjacency matrix is ​​preprocessed by adding self-loops and normalizing it to ensure that nodes retain their own feature information when aggregating information and to prevent feature values ​​from exploding or disappearing during multi-layer propagation. Next, the system performs the core graph convolution operation, continuously multiplying the preprocessed adjacency matrix, the current selling state feature vector sequence, and the learnable weight matrix. This process essentially allows each node to collect feature information from its neighbors and map it to a new feature space through a linear transformation. Finally, a non-linear activation function is used to enhance the model's expressive power.

[0041] Specifically, the sales forecasting and spoilage risk assessment module 340 is used to perform sales forecasting and spoilage risk assessment on the high-dimensional embedded vector sequence of the global vending machine status to obtain a demand forecast curve and an estimated loss risk value. It should be understood that although the high-dimensional embedded vector sequence of the global vending machine status contains rich information, it is a highly abstract feature representation and cannot be directly understood by operators or used to formulate specific replenishment and promotion plans. The two core bases for operational decisions are future product demand (avoiding stockouts or slow sales) and the risk of product spoilage (reducing losses). Therefore, in the technical solution of this application, sales forecasting and spoilage risk assessment are performed on the high-dimensional embedded vector sequence of the global vending machine status. Through a specialized forecasting model, the high-dimensional embedded vector is decoded, and two key business indicators are output in parallel: first, a continuous demand forecast curve for a future period to accurately guide replenishment volume; and second, an estimated loss risk value to quantitatively assess potential losses caused by expired or slow-selling products. These two together constitute the cornerstone of a refined and risk-controllable operational strategy.

[0042] In specific implementation, firstly, the high-dimensional embedding vector sequence of the global vending machine state is temporally dependently encoded to obtain the vending machine state temporal context vector. It should be understood that the operational state of the vending machine (such as sales trends and inventory depletion rates) is a continuous and highly time-dependent dynamic process. Simple static snapshots or aggregation methods that ignore time order cannot effectively model key temporal patterns such as daily / weekly periodicity, trend changes, and holiday effects. Although the high-dimensional embedding vector sequence of the global vending machine state output by the preceding steps contains rich state information at each point in time, its inherent temporal correlation has not been explicitly modeled. Therefore, in the technical solution of this application, the high-dimensional embedding vector sequence of the global vending machine state is temporally dependently encoded to capture the evolution patterns over long time spans (such as daily and weekly periodicity). Specifically, the vending machine state temporal context vector can be obtained by inputting the high-dimensional embedding vector sequence of the global vending machine state into a recurrent neural network (such as LSTM or GRU) for temporal encoding. This vector serves as the input for subsequent decoding steps, ensuring that the prediction is not only based on the current state but also takes into account the evolution trajectory of historical states, thereby significantly improving the accuracy of the prediction.

[0043] Furthermore, the vending machine's temporal context vector is decoded using a multi-task prediction head to obtain the demand forecast curve and the estimated loss risk value. To simultaneously complete two different forecasting tasks, the system constructs two independent decoding networks (i.e., prediction heads) on top of the shared context vector, respectively decoding the demand forecast curve and the estimated loss risk value from the aforementioned temporal context vector. Specifically, for demand forecasting, a fully connected layer (or a small neural network) is typically used to regress the demand values ​​at multiple future time points (such as the predicted sales volume per hour for the next 24 hours), thus forming a demand forecast curve. For spoilage risk assessment, another fully connected layer (or neural network) is used to regress a scalar value, namely the estimated loss risk value, which can be understood as the probability or expected amount of loss that the goods will not be sold due to expiration or slow sales within a certain period of time.

[0044] Specifically, the collaborative decision-making module 350 is used to collaboratively decide on the demand forecast curve and the estimated loss risk value to obtain a suggested replenishment time window. It should be understood that isolated forecast results cannot directly generate an optimal replenishment plan. For example, the demand forecast curve indicates future sales volume, and the estimated loss risk value indicates the possibility of product expiration. However, determining when to replenish requires comprehensive consideration of various complex and potentially conflicting factors, such as logistics costs (e.g., vehicle dispatch fees, route costs), inventory holding costs, stockout losses, and operational constraints (e.g., vehicle capacity, working time windows). Replenishing too early may increase logistics frequency and costs, and increase the risk of loss due to inventory backlog; replenishing too late may lead to stockouts and lost sales. Therefore, in the technical solution of this application, by constructing a decision model that can weigh these multi-objective factors, a suggested replenishment time window considering the global optimization objective is output, providing data-driven decision support for operators, and ultimately achieving a balance between minimizing operating costs and optimizing service quality.

[0045] In practical implementation, firstly, a reinforcement learning environment instance is constructed based on the demand forecast curve, estimated loss risk value, logistics and distribution cost parameters, and a set of constraints. That is, the demand forecast curve and estimated loss risk value output from the previous steps are used as the state input of the environment, and logistics and distribution cost parameters (such as single delivery cost, vehicle travel distance cost) and a set of constraints (such as maximum vehicle load, driver working time window, and replenishment personnel's current location) are introduced as the boundary conditions of the environment, thereby constructing a reinforcement learning environment instance that simulates a real-world operational scenario. In this environment, the system defines a reward function to maximize overall profit. Its calculation logic typically includes positive feedback from sales profit, negative feedback from spoilage and loss, and a penalty term for logistics costs. This process is expressed by the formula:

[0046] in, for Reward value at any moment For goods unit price, To predict demand, For inventory level, For the cost of goods, This is the expected loss calculated based on the risk value. To account for the logistics costs of replenishing stock at that moment, , and The hyperparameters used to adjust the weights.

[0047] Next, PPO agent policy iteration is performed on the reinforcement learning environment instance to obtain a converged network policy. PPO agent policy iteration refers to an advanced policy gradient algorithm training process. Specifically, the agent is trained using the Proximal Policy Optimization (PPO) algorithm. The agent continuously tries different replenishment opportunities in the simulated environment and receives feedback through the aforementioned reward function, thereby updating its policy network parameters. The PPO algorithm ensures training stability by limiting the magnitude of policy updates, ultimately obtaining a converged network policy. The core optimization objective function of this process is expressed by the formula:

[0048] in, For policy network parameters, The probability ratio between the old and new strategies. This is an estimate of the dominance function. This is the truncation range hyperparameter.

[0049] Then, the converged network policy is decoded to obtain the suggested replenishment time window. Specifically, after the network training converges, the system inputs the current real-time state into the policy network, and the network outputs an action probability distribution or deterministic action value. By decoding this output, the system maps the abstract numerical value into specific physical time operation instructions, thereby obtaining the suggested replenishment time window (e.g., "today 14:00-15:00"). Here, the suggested replenishment time window refers to the optimal delivery period calculated by the system, within which replenishment can achieve the best balance between profit and cost.

[0050] As described above, the internet-based smart vending system 300 according to the embodiments of this application can be implemented in various wireless terminals, such as servers with internet-based smart vending algorithms. In one possible implementation, the internet-based smart vending system 300 according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the internet-based smart vending system 300 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the internet-based smart vending system 300 can also be one of many hardware modules of the wireless terminal.

[0051] Alternatively, in another example, the internet-based smart vending system 300 and the wireless terminal can also be separate devices, and the internet-based smart vending system 300 can connect to the wireless terminal via wired and / or wireless networks and transmit interactive information in accordance with an agreed data format.

[0052] Furthermore, an intelligent sales method based on an internet platform is also provided.

[0053] Figure 3 This is a block diagram of an intelligent vending method based on an internet platform according to an embodiment of this application. Figure 3 As shown, the intelligent vending method based on an internet platform according to an embodiment of this application includes the following steps: S1, acquiring static data, dynamic data, and external environment data; S2, performing multimodal feature preprocessing on the static data, dynamic data, and external environment data to obtain a sales status feature vector sequence; S3, generating a global vending machine status high-dimensional embedding vector sequence based on the sales status feature vector sequence and the vending machine location coordinates; S4, performing sales forecasting and spoilage risk assessment on the global vending machine status high-dimensional embedding vector sequence to obtain a demand forecast curve and an estimated loss risk value; S5, performing collaborative decision-making on the demand forecast curve and the estimated loss risk value to obtain a suggested replenishment time window.

[0054] In summary, the intelligent vending method based on an internet platform according to the embodiments of this application is explained. It constructs a high-dimensional state embedding vector that captures the spatial dependencies between global devices by deeply mining the static attributes, dynamic flow, and external environmental data of vending machines. Based on this vector, it predicts the demand curve and the risk value of product spoilage in parallel at future moments. Furthermore, it uses a collaborative decision-making mechanism to comprehensively weigh revenue and losses, outputting a suggested replenishment time window that takes into account the logistics response cycle. In this way, it effectively solves the problems of delayed replenishment or blind stockpiling caused by static threshold management, significantly reduces the expiration and spoilage rate of fresh food products and the opportunity cost of stockouts, thereby achieving optimal synergy between refined management, profit maximization, and operational efficiency of unmanned retail networks in complex and dynamic environments.

[0055] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A smart vending system based on an internet platform, characterized in that, include: The data acquisition module is used to acquire static data, dynamic data, and external environment data; The multimodal feature preprocessing module is used to perform multimodal feature preprocessing on static data, dynamic data, and external environment data to obtain a sales status feature vector sequence; The global state high-dimensional embedding module is used to generate a global vending machine state high-dimensional embedding vector sequence based on the vending machine state feature vector sequence and vending machine location coordinates; The sales forecasting and spoilage risk assessment module is used to perform sales forecasting and spoilage risk assessment on the high-dimensional embedded vector sequence of global vending machine status to obtain the demand forecast curve and the estimated loss risk value. The collaborative decision-making module is used to make collaborative decisions based on the demand forecast curve and the estimated loss risk value to obtain a suggested replenishment time window.

2. The intelligent vending system based on an internet platform according to claim 1, characterized in that, Static data includes vending machine ID, geographic coordinates, and basic product SKU information table; dynamic data includes real-time sales logs from vending machines, inventory sensor logs, and heartbeat data from vending machines; external environment data includes minute-level weather data and hourly pedestrian traffic heatmap data for the machine's location from third-party APIs, as well as a pre-built holiday / event tag library.

3. The intelligent vending system based on an internet platform according to claim 1, characterized in that, The multimodal feature preprocessing module is used for: Temporal alignment is performed on static data, dynamic data, and external environment data to obtain a temporally aligned original dataset; Inventory cell lifecycle quantization is performed on the time-aligned original dataset to obtain an inventory dataset with quantified lifecycle; Multimodal feature fusion and normalization are performed on the inventory dataset with quantified lifecycle to obtain a sequence of sales status feature vectors.

4. The intelligent vending system based on an internet platform according to claim 1, characterized in that, The global state high-dimensional embedding module includes: The historical sales time series data extraction unit is used to extract historical sales time series data from the sales status feature vector sequence; The weighted adjacency matrix construction unit is used to construct a weighted adjacency matrix based on historical sales time series data and vending machine location coordinates; The graph convolutional spatial feature aggregation unit is used to perform graph convolutional spatial feature aggregation on the vending state feature vector sequence and the weighted adjacency matrix to obtain a high-dimensional embedding vector sequence of the global vending machine state.

5. The intelligent vending system based on an internet platform according to claim 4, characterized in that, The weighted adjacency matrix construction unit includes: The sales relevance weight calculation subunit is used to generate a sales relevance weight matrix based on historical sales time series data. The geographic proximity weight calculation subunit is used to generate a geographic proximity weight matrix based on the vending machine's location coordinates; The dynamic relationship gating subunit is used to determine the dynamic relationship gating matrix based on the sales status feature vector sequence; The fusion subunit is used to perform matrix fusion of the geographic proximity weight matrix and the sales relevance weight matrix based on the dynamic relationship gating matrix to obtain a weighted adjacency matrix.

6. The intelligent vending system based on an internet platform according to claim 5, characterized in that, Dynamic relationship gating subunit, used for: The node-pair joint state representation is performed on the sales state feature vector sequence to obtain the node-pair joint state vector sequence. Gating modeling is performed on the joint state vector sequence of nodes to obtain the dynamic relation gating matrix.

7. The intelligent vending system based on an internet platform according to claim 1, characterized in that, The sales forecasting and spoilage risk assessment module is used for: Temporal dependency encoding is performed on the high-dimensional embedding vector sequence of global vending machine state to obtain the vending machine state temporal context vector; Multi-task prediction head decoding is performed on the vending machine state time-series context vector to obtain the demand forecast curve and the estimated loss risk value.

8. The intelligent vending system based on an internet platform according to claim 1, characterized in that, The collaborative decision-making module is used for: Based on demand forecasting curves, estimated loss risk values, logistics and distribution cost parameters, and a set of constraints, a reinforcement learning environment instance is constructed. PPO agent policy iteration is performed on reinforcement learning environment instances to obtain a converged network policy; The converged network strategy is action-decoded to obtain the suggested replenishment time window.

9. A smart sales method based on an internet platform, characterized in that, include: Acquire static data, dynamic data, and external environment data; Multimodal feature preprocessing is performed on static data, dynamic data, and external environment data to obtain a sequence of sales status feature vectors; Based on the vending status feature vector sequence and vending machine location coordinates, a global vending machine status high-dimensional embedding vector sequence is generated; Sales forecasting and spoilage risk assessment are performed on the high-dimensional embedded vector sequence of global vending machine status to obtain demand forecast curves and estimated loss risk values. A collaborative decision is made based on the demand forecast curve and the estimated loss risk value to obtain a suggested replenishment window.

Images