Fruit fresh cash register integrated intelligent control method, device, equipment and storage medium

Abstract

The application discloses a fruit fresh food cash register integrated intelligent control method, device, equipment and storage medium, relates to the technical field of intelligent retail, and the method comprises the following steps: acquiring a fresh food transaction log containing visual identification information of a cash register terminal and a member interaction log of a member terminal, performing multi-source data cleaning and structural alignment on the two types of logs, and extracting a fresh food transaction and member interaction feature set; then, through cross-domain feature fusion and portrait mining, a multi-dimensional member portrait and real-time transaction context feature are generated; subsequently, the features are input into a multi-task cooperative network for bidirectional instruction deduction, real-time cash register control instructions and asynchronous marketing push instructions are generated; finally, the instructions are respectively sent to corresponding terminals to drive the terminals to perform cross-module cooperative actions. The application realizes full-link automatic cooperation of cash register and member management links in a fresh food retail scene, and reduces the dependence on manual intervention in the operation process.

Description

Technical Field

[0001] This application relates to the field of smart retail technology, and in particular to a method, device, equipment and storage medium for integrated smart control of fruit and fresh produce checkout. Background Technology

[0002] With the acceleration of digital transformation, merchants have placed higher demands on operational efficiency. They not only need to achieve rapid AI recognition, accurate pricing, and multi-channel payment for bulk fresh produce, but also hope to improve customer repurchase rates and average order value by deeply analyzing member consumption behavior.

[0003] Current technological solutions in the fresh food retail sector often exhibit a focus on single-point breakthroughs. At the checkout end, some systems have introduced computer vision technology for fruit and vegetable identification and integrated various payment methods such as QR code payment and facial recognition payment. For membership management, the common practice is to establish a separate mini-program or membership system, using basic points, tiered management, or periodic group coupon distributions for secondary outreach. At the data level, merchants typically utilize general business intelligence tools to offline synchronize store revenue to generate sales statistics and simple report analysis to support daily operational decisions.

[0004] In current practices, POS terminals and membership systems are often isolated, lacking real-time data exchange and collaborative command mechanisms. POS terminals are only responsible for physical weighing and settlement, unable to dynamically adjust pricing or trigger specific strategies based on real-time member profiles; while membership systems are mostly passively recording data, unable to capture precise modal data at the moment of checkout. This "data silo" between modules means that operational decisions still heavily rely on managerial experience and manual intervention, making it difficult to achieve automated and precise cross-module linkage. Especially when facing high-concurrency transactions and complex member intent deductions, the system's response efficiency and decision-making accuracy are limited. Therefore, how to achieve end-to-end automated collaboration between POS and membership management in fresh food retail scenarios and reduce reliance on manual intervention in operations has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide an integrated intelligent control method, device, equipment and storage medium for fruit and fresh produce checkout, aiming to solve the technical problem of how to achieve full-link automated collaboration between checkout and membership management in fresh produce retail scenarios and reduce the reliance on manual intervention in the operation process.

[0006] To achieve the above objectives, this application proposes an integrated intelligent control method for fruit and fresh produce checkout, the method comprising: Obtain fresh food transaction logs containing visual recognition information uploaded by the POS terminal, and member interaction logs uploaded by the member terminal; The fresh food transaction logs and the member interaction logs are subjected to multi-source data cleaning and structure alignment to obtain the fresh food transaction feature set and the member interaction feature set. Cross-domain feature fusion and profile mining are performed on the fresh food transaction feature set and the member interaction feature set to obtain multi-dimensional member profile features and real-time transaction context features. The multi-dimensional member profile features and the real-time transaction context features are synchronously input into a multi-task collaborative network for bidirectional instruction deduction, generating real-time cashier control instructions and asynchronous marketing push instructions. The real-time cashier control command is sent to the cashier terminal, and the asynchronous marketing push command is sent to the member terminal, so that the cashier terminal and the member terminal can perform corresponding cross-module collaborative actions.

[0007] Furthermore, to achieve the above objectives, this application also proposes an integrated intelligent control device for fruit and fresh produce checkout, the device comprising: The data acquisition module is used to acquire fresh food transaction logs containing visual recognition information uploaded by the POS terminal, and member interaction logs uploaded by the member terminal; The data preprocessing module is used to perform multi-source data cleaning and structure alignment on the fresh food transaction logs and the member interaction logs to obtain fresh food transaction feature sets and member interaction feature sets. The profile fusion module is used to perform cross-domain feature fusion and profile mining on the fresh food transaction feature set and the member interaction feature set to obtain multi-dimensional member profile features and real-time transaction context features. The intelligent deduction module is used to synchronously input the multi-dimensional member profile features and the real-time transaction context features into the multi-task collaborative network for bidirectional instruction deduction, generating real-time cashier control instructions and asynchronous marketing push instructions; The collaborative delivery module is used to send the real-time cashier control command to the cashier terminal and the asynchronous marketing push command to the member terminal, so that the cashier terminal and the member terminal can perform corresponding cross-module collaborative actions.

[0008] In addition, to achieve the above objectives, this application also proposes an integrated intelligent control device for fruit and fresh produce cashiering, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the integrated intelligent control method for fruit and fresh produce cashiering as described above.

[0009] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the integrated intelligent control method for fruit and fresh produce cashier described above.

[0010] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the integrated intelligent control method for fruit and fresh produce cashier described above.

[0011] One or more technical solutions proposed in this application have at least the following technical effects: First, by acquiring fresh food transaction logs containing visual recognition information uploaded from POS terminals and member interaction logs uploaded from membership terminals, a comprehensive data foundation covering the transaction process and member behavior is constructed. Next, multi-source data cleaning and structured alignment are performed on the fresh food transaction logs and member interaction logs to obtain fresh food transaction feature sets and member interaction feature sets, thereby eliminating data noise and unifying analysis dimensions to ensure the accuracy of subsequent processing. Then, cross-domain feature fusion and profile mining are performed on the fresh food transaction feature sets and member interaction feature sets to obtain multi-dimensional member profile features and real-time transaction context features, thereby deeply reconstructing the consumption scenario. The system accurately identifies user intent; then, it inputs multi-dimensional member profile features and real-time transaction context features into a multi-task collaborative network for bidirectional instruction deduction, generating real-time checkout control instructions and asynchronous marketing push instructions. This ensures a high degree of logical consistency and bidirectional linkage between settlement intervention and marketing triggering. Finally, the real-time checkout control instructions are sent to the checkout terminal, and the asynchronous marketing push instructions are sent to the member terminal to drive the checkout terminal and member terminal to execute corresponding cross-module collaborative actions. This achieves full-link automated collaboration between checkout and member management in the fresh food retail scenario and reduces the reliance on manual intervention in the operation process. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating an embodiment of the integrated intelligent control method for fruit and fresh produce payment in this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the intelligent control method for integrated cashier system for fresh fruit in this application. Figure 3This is a schematic diagram of the structure of the multi-task collaborative network provided in Embodiment 2 of the integrated intelligent control method for fruit and fresh produce payment in this application; Figure 4 This is a schematic diagram of the system architecture of the integrated intelligent control method for fruit and fresh produce payment provided in Embodiment 2 of this application; Figure 5 This is a schematic diagram of the modular structure of the integrated intelligent control device for fruit and fresh produce payment in an embodiment of this application; Figure 6 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the integrated intelligent control method for fruit and fresh produce payment in the embodiments of this application.

[0015] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0017] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0018] It should be noted that the executing entity in this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or cloud server capable of performing the above functions. The following description uses a cloud server as an example to illustrate this embodiment and the subsequent embodiments.

[0019] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0020] Based on this, the embodiments of this application provide an integrated intelligent control method for fruit and fresh produce checkout, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the integrated intelligent control method for fruit and fresh produce payment in this application.

[0021] In this embodiment, the integrated intelligent control method for fruit and fresh produce checkout includes steps S10 to S50: Step S10: Obtain the fresh food transaction log containing visual recognition information uploaded by the POS terminal, and the member interaction log uploaded by the member terminal; Step S20: Perform multi-source data cleaning and structure alignment on the fresh food transaction logs and the member interaction logs to obtain the fresh food transaction feature set and the member interaction feature set. Step S30: Perform cross-domain feature fusion and profile mining on the fresh food transaction feature set and the member interaction feature set to obtain multi-dimensional member profile features and real-time transaction context features. Step S40: The multi-dimensional member profile features and the real-time transaction context features are synchronously input into the multi-task collaborative network for bidirectional instruction deduction to generate real-time cashier control instructions and asynchronous marketing push instructions; Step S50: Send the real-time cashier control command to the cashier terminal and the asynchronous marketing push command to the member terminal, so that the cashier terminal and the member terminal can perform corresponding cross-module collaborative actions.

[0022] It should be noted that the POS terminal refers to the intelligent POS equipment deployed in fresh food stores, integrating functions such as AI fruit and vegetable recognition, weighing and pricing, and multiple payment methods. It is the core hardware terminal used to complete product identification, transaction settlement, and data collection. Visual recognition information refers to the feature data extracted after image analysis of fresh produce using computer vision technology (such as the YOLOv8 lightweight model), including identification results such as product category, weight, and appearance quality. Fresh produce transaction logs refer to the raw data collection recording the transaction process of fresh produce, including structured transaction records such as product ID, sales volume, weight, payment method, member ID, and transaction time. Membership terminals refer to the interactive devices used by members, typically mini-programs or apps on smartphones, used for member identification, receiving promotional information, viewing consumption records, and other membership service functions.

[0023] Member interaction logs refer to the raw data recording member interactions with the system, including purchase time, purchased categories, purchase amount, visit frequency, and discount clicks, among other member behavior trajectory information. Fresh produce transaction feature sets refer to the set of fresh produce transaction data features after data cleaning and structuring, containing standardized product characteristics, transaction characteristics, payment characteristics, and other core data dimensions that can be used for analysis. Member interaction feature sets refer to the set of member behavior data features after cleaning and standardization, containing structured feature data reflecting member behavior patterns, such as consumption habits, price sensitivity, and category preferences. Multi-dimensional member profile features refer to a set of structured data tags extracted by the system based on a customer's long-term purchase records and online interaction habits, comprehensively describing the customer. For example, it depicts what types of fresh produce the customer tends to buy, their price sensitivity, and how often they purchase; it portrays the customer's inherent "long-term consumption habits." Real-time transaction context features refer to a set of data formed by combining and aligning the product information the customer is currently checking out (such as type and weight) with the online actions the customer just performed before payment (such as claiming coupons or browsing); it reflects the customer's immediate purchase scenario.

[0024] Multi-task collaborative networks refer to a deep learning model structure capable of handling multiple prediction tasks simultaneously. In this scenario, it shares underlying data, allowing two specialized branches to work concurrently. One branch infers the customer's current checkout intention (e.g., which discount to use), while the other predicts the customer's long-term preferences (e.g., what they might buy in the future), thus achieving shared computing power and synchronized decision-making. Real-time checkout control commands refer to checkout terminal control commands generated instantly based on the current transaction scenario, such as member discount matching, product price verification, and payment method recommendations—operations that take effect immediately. Asynchronous marketing push commands refer to non-real-time marketing commands generated based on member profile analysis, such as marketing actions triggered at appropriate times, like pushing discounts on preferred products, repeat purchase reminders, and personalized recommendations. Cross-module collaborative actions refer to the linked operations executed by the checkout terminal and member terminals under system command, such as triggering member discount pushes upon checkout completion and simultaneously adjusting marketing strategies when inventory warnings are issued—collaborative behaviors across functional modules.

[0025] Understandably, firstly, the cloud server receives real-time fresh food transaction data uploaded by the underlying POS terminal, along with visual recognition results from the camera, and user behavior logs uploaded by members' mobile applications (e.g., simultaneously receiving the weight and amount records of customers weighing apples offline and the click records of browsing fruit coupons online). This is done to comprehensively collect raw business materials covering both physical settlement and virtual online browsing. Secondly, the cloud server directly performs anomaly noise removal and underlying format standardization conversion on these two types of log data from different sources, forcibly aligning and splicing them in their corresponding time series and user identifiers to generate a fresh food transaction feature set and a member interaction feature set with standardized format and consistent dimensions (e.g., forcibly converting the POS terminal's unique wide table fields and the mini-program's tracking fields into the same standard machine-readable array). This is done to completely eliminate data barriers between heterogeneous hardware systems and prevent dimensional conflicts and errors during subsequent multiplication and addition operations in the algorithm.

[0026] Then, the cloud server transcends a single business boundary and deeply reorganizes, filters, and processes the data from the two feature sets mentioned above. From this, it extracts the underlying profile data representing the user's long-term inherent attributes. At the same time, it assembles and aggregates this data with the features of the current instant order that is being checked out, thereby obtaining multi-dimensional member profile features and real-time transaction context features (for example, extracting a customer's fixed attribute of "long-term preference for high-priced imported fresh fruit" and packaging it with the instant action of "currently queuing to pay for near-expiry discounted vegetables" into a whole data payload). This is done in order to completely and three-dimensionally capture and restore the full picture of "what kind of customer group is making a settlement in what specific environment and background" at the data level. Next, the cloud server feeds the prepared profile features and context features as dual parameters to the pre-deployed multi-task collaborative network, driving the parallel computing channels within the network to perform forward inference simultaneously. This generates a real-time cash register control instruction and an asynchronous marketing push instruction at the same time (for example, using the same AI model to simultaneously calculate whether to directly instruct the cash register to offer a 10% discount and to push a cherry delivery notification to the user's mobile phone tomorrow). This is done to use a single central algorithm brain to efficiently solve the two completely different but closely related business needs of "how to collect money immediately" and "how to sell goods in the future".

[0027] Finally, the cloud server precisely sends real-time checkout control commands to the physical POS machine currently processing the order via a communication link, and delivers asynchronous marketing push commands to the customer's corresponding mobile application. This directly drives the underlying hardware to complete automatic pricing and deduction operations and display promotional content on the mobile interface (for example, automatically deducting 3 yuan from the physical electronic scale and displaying a special repurchase coupon in the user's WeChat mini-program). This is done to ensure that the abstract strategy data derived from the cloud is ultimately and reliably implemented as a fully automated, cross-system linkage between the checkout hardware and the user's mobile phone, completely eliminating the need for human store manager intervention.

[0028] As an example, the steps of obtaining the fresh food transaction logs containing visual recognition information uploaded by the POS terminal and the member interaction logs uploaded by the member terminal include: sending an image acquisition trigger signal to the POS terminal so that the POS terminal can obtain the category characteristics and weight characteristics of bulk fresh food through a preset lightweight target detection model; receiving the fresh food transaction logs uploaded by the POS terminal, wherein the fresh food transaction logs are obtained by structurally binding the category characteristics, the weight characteristics, and multiple payment methods; receiving the identity credential to be verified generated by the member terminal based on the user's biometric recognition operation; verifying the legality of the identity credential to be verified in a preset member authentication database, and obtaining the unique identifier of the target user corresponding to the identity credential to be verified after the verification is passed; and indexing the data in the preset member authentication database according to the unique identifier of the target user and retrieving the member interaction logs that match the unique identifier of the target user.

[0029] It should be noted that the image acquisition trigger signal refers to the electronic command signal sent from the cloud server to the POS terminal to initiate the image capture operation. This signal is used to trigger the camera to capture images when fresh produce is placed on the weighing platform. The preset lightweight target detection model refers to a computer vision algorithm (such as YOLOv8-Nano) deployed on the edge computing device of the POS terminal, which has undergone model compression optimization and can quickly and accurately identify the visual features of bulk fresh produce in low-computing-power environments. Category features refer to the feature data extracted from the fresh produce image through visual recognition algorithms that can uniquely identify the product category, including multi-dimensional identification parameters such as appearance color distribution, shape contour features, and texture detail features. Weight features refer to the precise weight value of the bulk fresh produce measured by the POS terminal's weighing sensor, usually in grams and including a weight confidence score.

[0030] Multiple payment methods refer to a set of settlement methods that support multiple payment channels, including WeChat Pay, Alipay, UnionPay QuickPass, membership stored value payment, cash payment, and other payment methods. The identity credential to be verified refers to the user's original biometric data collected by the member terminal through biometric technology (such as fingerprint recognition and facial recognition), which is then encrypted to generate a data packet of identity identifier to be verified. The preset member authentication database refers to a centralized data management system stored on a cloud server that contains member identity information, permission rules, and biometric templates, used to perform member identity authentication and permission verification. The unique identifier of the target user refers to a globally unique identity code assigned to a legitimate member by the system after identity verification, usually in UUID format, used to uniquely identify and track all behavioral data of a specific member in the system.

[0031] Understandably, firstly, when the cloud server detects a business trigger event, it sends an image acquisition trigger signal to the POS terminal, driving the terminal's optical sensor to capture an image of the object on the weighing pan. This image is then processed using a pre-set lightweight object detection model to extract features, thereby obtaining the category and weight characteristics of the bulk fresh produce. This real-time triggering mechanism, achieved through edge-sensing, ensures the objectivity and temporal consistency of the transaction initiation data. Secondly, the cloud server receives fresh produce transaction logs uploaded by the POS terminal in real time. These logs are generated by structurally binding the extracted category and weight characteristics, along with the user's chosen payment methods, on the terminal side. Through this logical coupling of multi-dimensional data, the cloud server can obtain a complete digital description of a single transaction event, providing reliable voucher records for subsequent closed-loop analysis.

[0032] Then, the cloud server synchronously receives the identity credentials to be verified from the member terminal after the user performs biometric identification. It then uses the access control logic within the pre-defined member authentication database to perform security and legality verification on these credentials (e.g., verifying the integrity and timeliness of the token signature). Upon successful verification, a globally unique identifier for the target user is mapped, thus establishing a precise mapping between transaction behavior and the identity of a specific natural person in the complex distributed system. Finally, based on the obtained unique identifier of the target user, the cloud server performs efficient data indexing in the pre-defined member authentication database, retrieving matching member interaction logs that record historical behavioral patterns (e.g., the user's recent promotional activity clicks or coupon claim status). This completes the comprehensive aggregation of the user's static attributes and dynamic behaviors, building a solid data foundation for the accurate deduction of subsequent cross-module collaborative instructions.

[0033] As an example, the step of performing cross-domain feature fusion and profile mining on the fresh food transaction feature set and the member interaction feature set to obtain multi-dimensional member profile features and real-time transaction context features includes: updating the member consumption profile in a preset database based on the fresh food category data, weight data, and consumption frequency data in the fresh food transaction feature set; performing behavioral association analysis based on the member consumption profile and the member interaction feature set to obtain a multi-tag set; clustering and integrating the multi-tag set according to the preset number of tag categories to obtain multi-dimensional member profile features; and concatenating the real-time interaction sequence in the member interaction feature set with the current order transaction features in the fresh food transaction feature set to obtain real-time transaction context features.

[0034] It should be noted that fresh produce category data refers to the product category identification results recorded in the fresh produce transaction feature set, including standardized coding information for specific product categories and their subcategories such as fruits, vegetables, and meats. Weight data refers to the precise weight values ​​of bulk goods recorded in the fresh produce transaction feature set, including quantitative indicators in grams such as single purchase weight and cumulative purchase weight. Purchase frequency data refers to the information on the number of times a member purchases fresh produce within a specific time period, including time-dimensional statistical indicators such as average daily purchase frequency, weekly purchase frequency, and monthly repurchase rate. The preset database refers to the structured data storage system pre-configured by the system for storing and managing member consumption history, including basic configuration parameters such as data table structure, indexing rules, and storage engine. The member consumption profile refers to a personalized data profile created for each member, containing long-term consumption behavior data such as historical purchase records, consumption amount, and category preferences, used for continuous tracking and analysis of member consumption patterns.

[0035] A multi-tag set refers to a combination of tags extracted through behavioral association analysis that can describe user characteristics from multiple perspectives. Each tag represents a user's attribute in a specific dimension, such as descriptive identifiers like "high-end fruit preference," "price-sensitive user," and "high-frequency consumer." The preset tag category quantity refers to the total number of fixed categories predefined by the system for organizing and classifying user tags. For example, user tags can be divided into standardized categories such as spending power, category preference, price sensitivity, and active time periods. Real-time interaction sequence refers to a series of continuous operation trajectory data triggered by customers on online membership terminals such as mobile mini-programs within a time window before or during the checkout of a physical order (e.g., actions such as [opening the membership code], [briefly lingering on the promotion page], and [clicking to claim a cherry discount coupon] occurring sequentially in time). It can extremely accurately reflect the customer's most direct psychological motivation and attention focus at the moment of purchase. Current order transaction characteristics refer to the specific characteristic information of the currently ongoing transaction order in the fresh food transaction characteristic set, including real-time transaction scenario data such as product combination type, transaction amount, purchase time, payment method, and store location.

[0036] Understandably, firstly, the cloud server extracts fresh produce category data, weight data, and consumption frequency data recorded in the fresh produce transaction feature set. Using these quantified entity indicators, it performs data reconstruction operations such as numerical accumulation or timestamp overwriting in a pre-set database to dynamically update the member consumption profile of the corresponding target user (e.g., simultaneously adding the one kilogram of cherries purchased this time to the customer's historical total purchase field and refreshing their last active time). This is to ensure that the underlying factual basis of the system can accurately synchronize the user's latest offline physical consumption strength and actual conversion trajectory at all times. Secondly, the cloud server retrieves the updated member consumption profile and performs cross-modal behavioral correlation analysis with the member interaction feature set that records the user's online operational behavior. By calculating the joint probability mapping between long-term physical purchase patterns and fragmented virtual interaction actions, it extracts a multi-label set that directly addresses the user's deep motivations (e.g., cross-calculating the behavior logic of frequently purchasing high-priced offline goods with rarely clicking on promotional pages online to extract discrete business labels representing their purchasing tendencies). This is to break through the visual blind spots of single online or offline data, accurately extracting feature anchors reflecting true psychological expectations from massive and messy logs.

[0037] Then, the cloud server performs spatial distance calculation and clustering integration operations on the extracted discrete and complex multi-label set in the multi-dimensional vector space according to the preset number of label categories. This compresses and converges the multi-dimensional member profile features into standard dimensions (for example, merging and refining dozens of messy sub-labels into a structured feature vector pointing to the fixed dimension of "high-frequency high-net-worth customer group" through an algorithm). This is done to standardize and reduce the dimensionality of redundant fragmented labels, providing a set of long-term static attribute loads with minimal computational overhead and absolute dimensional alignment for subsequent multi-task networks. Finally, the cloud server precisely extracts the real-time interaction sequence from the member interaction feature set on the timeline, and performs a hard data splicing and assembly operation at the underlying tensor level with the current order transaction feature representing the offline settlement environment in the fresh food transaction feature set. This directly generates real-time transaction context features with both temporal extension and spatial fixation attributes (for example, directly splicing the code of the fresh food product being priced on the customer's current scale with the duration of the customer's stay on the discount page that they just browsed in the mobile app half a minute ago into a complete data packet). This is done to seamlessly weld the customer's fleeting online operation motivation with the objectively existing offline physical checkout fact in the algorithm base, thereby providing the most timely and environmentally constrained decision background matrix for subsequent inference of instantaneous checkout intentions.

[0038] As an example, the step of performing behavioral association analysis based on the member consumption profile and the member interaction feature set to obtain a multi-label set includes: extracting historical purchase frequency features and average order value features from the member consumption profile, and extracting page dwell time features and activity click-through conversion features from the member interaction feature set; aligning the historical purchase frequency features, average order value features, page dwell time features, and activity click-through conversion features to construct a multimodal user feature sequence; inputting the multimodal user feature sequence into a hidden Markov model for behavioral sequence decoding to obtain a behavioral association feature vector; and mapping the behavioral association feature vector to a preset preference label space for threshold filtering to obtain a multi-label set.

[0039] It should be noted that historical purchase frequency characteristics refer to the statistical characteristics of the number of times a user purchases fresh produce within a specific time period, extracted from the member's consumption profile. This includes quantitative indicators such as the purchase frequency of different product categories and the periodic purchase interval. Average order value characteristics refer to the characteristics of the average amount spent per transaction by a user, extracted from the member's consumption profile. This includes data on the overall average order value level, differences in average order value across different product categories, and trends in average order value fluctuations. Page dwell time characteristics refer to the distribution characteristics of the time users spend on various functional pages of mobile applications or websites (such as product detail pages, promotional activity pages, and shopping cart pages), extracted from the member interaction characteristic set. This reflects the degree of user attention to different content. Activity click-through conversion characteristics refer to the characteristics of user click behavior and subsequent conversion effects of marketing activities (such as coupon redemption, limited-time discounts, and new product recommendations), extracted from the member interaction characteristic set. This includes indicators such as click-through rate, conversion rate, and conversion delay time.

[0040] Multimodal user feature sequences refer to time-series feature sequences constructed by aligning historical purchase frequency features, average order value features, page dwell time features, and activity click-through conversion features according to the time dimension, containing heterogeneous behavioral data from multiple sources both online and offline. Hidden Markov Models (HMMs) are statistical models used to model multimodal user feature sequences, inferring and decoding users' hidden behavioral intentions through state transition probabilities and observation probability parameters. Behavioral association feature vectors are high-dimensional vectors output after decoding multimodal user feature sequences using HMMs; these vectors encode the potential association patterns and temporal dependencies between users' online and offline behaviors. Pre-defined preference label space refers to a pre-constructed vector mapping space containing user preference labels (such as price-sensitive, quality-oriented, and high-frequency repurchase), with each label corresponding to a specific vector region and threshold boundary.

[0041] Understandably, the cloud server first uses a unified time dimension as an alignment benchmark to synchronously extract historical purchase frequency and average order value features from member consumption profiles (e.g., extracting the number of times a user has purchased cherries in the past two weeks and the average amount paid each time). It also extracts page dwell time and activity click conversion features from the member interaction feature set (e.g., extracting the number of seconds a user browsed on the cherry promotion page and whether they ultimately clicked to claim a coupon). Subsequently, these four types of heterogeneous data reflecting different dimensions are spliced ​​and aligned into feature matrices according to the chronological order of the behavior, thereby constructing a multimodal user feature sequence that also has time series characteristics (e.g., linking a user's high average order value purchase record at an offline store on Monday with their long browsing and coupon claiming behavior on an online mini-program on Tuesday). This is done to spatiotemporally link the user's online decision-making process with the offline physical checkout result, forming a complete and coherent user behavior trajectory base.

[0042] Secondly, the cloud server inputs the constructed multimodal user feature sequence as the observation state into the Hidden Markov Model. Using the state transition probability matrix and observation probability matrix configured within the model, the system decodes the behavioral sequence of the true shopping intentions (i.e., hidden states) hidden behind surface interactions and transactions, calculates the probability distribution of the current user under various potential intentions, and then obtains an abstract behavioral association feature vector. This approach can overcome the limitations of single features and accurately infer the potential temporal dependencies and deep motivations between seemingly random user actions (for example, inferring that the subsequent high-frequency purchases of users with "long stay and low conversion" are driven by specific promotional events). Finally, the cloud server projects the abstract behavioral association feature vector output by the model into a pre-established preference label space coordinate system. By calculating the Euclidean distance or similarity score between the feature vector and each standard preference label anchor point in the coordinate system, and directly eliminating redundant feature items with scores below a preset benchmark threshold, a set of multi-labels that can directly reflect the user's unique characteristics is selected and combined (e.g., matching and outputting concrete labels such as "imported fruit preference", "high-frequency order value", and "promotion sensitivity"). This can transform high-dimensional vectors that are difficult to understand intuitively into standardized structured data that can be directly identified and called by downstream cashier pricing instructions and marketing push actions.

[0043] As an example, the training steps of the Hidden Markov Model include: defining a set of hidden states and a set of observed states, wherein the set of hidden states corresponds to the user's potential shopping intention and the set of observed states corresponds to the multimodal user feature sequence; initializing the initial state probability distribution, the state transition probability matrix, and the observation probability matrix; using the collected historical multimodal user feature sequence as input samples, iteratively calculating the input samples using the Baum-Welch algorithm to obtain the calculation results; continuously updating the parameter values ​​of the initial state probability distribution, the state transition probability matrix, and the observation probability matrix based on the calculation results; and stopping the iteration and outputting the trained Hidden Markov Model when the change in the parameter values ​​is less than a preset convergence threshold.

[0044] It should be noted that the latent state set refers to the set of a finite number of discrete states in the model that cannot be directly observed but drive behavior at the underlying level. In this embodiment, it consists of various potential user shopping intentions and is used to describe the user's psychological motivation when making decisions. The observed state set refers to the set of various feature states that can be directly recorded and extracted by the cashier or membership terminal in real-world scenarios. In this embodiment, it corresponds to the multimodal user feature sequence, i.e., various quantitative indicators reflecting user behavior. Potential user shopping intentions refer to the true motivations or psychological states hidden behind consumption behavior. For example, a user may be in a state of latent intention such as exploring high-quality fresh food, repurchasing daily necessities, or comparing prices with price sensitivity. These states determine the user's subsequent operational logic. The initial state probability distribution represents the probability vector of the user being in each latent state (i.e., various shopping intentions) at the beginning of the observation sequence. It reflects the user's initial psychological tendency when entering the system.

[0045] The state transition probability matrix describes the probability of a user transitioning from one hidden state to another at different time steps. For example, it records the likelihood of a user transitioning from the "browsing a page" state to the "generating a purchase impulse" state. The observation probability matrix, also known as the emission probability matrix, represents the probability of generating a specific observation characteristic (such as page dwell time exceeding 2 minutes) in a particular hidden state. It establishes a mapping relationship between intent and behavior. Historical multimodal user feature sequences refer to multi-source behavioral feature data extracted from historical databases and processed with time alignment. It integrates offline data such as historical purchase frequency and average order value, as well as online interaction data such as page dwell time and activity clicks, serving as the original samples for training the model.

[0046] The Baum-Welch algorithm is an iterative algorithm based on the expectation-maximization (EM) principle. Since shopping intent (the hidden state) is invisible, the algorithm searches for model parameters that best fit the data distribution by repeatedly executing "expectation steps" and "maximization steps" with only observed data. The calculation results refer to a series of intermediate variables generated by the Baum-Welch algorithm in each iteration, such as forward probabilities, backward probabilities, and specific expectation counts. Parameter values ​​refer to the specific numerical values ​​in the initial state probability distribution, state transition probability matrix, and observation probability matrix. The training goal is to learn from historical data and adjust these parameters to an optimal state to accurately predict user intent. The preset convergence threshold is a set of extremely small positive numbers (0.001 in this example) pre-set by the system to measure the magnitude of change in model parameters between adjacent iterations.

[0047] Understandably, firstly, the cloud server defines a set of hidden states and a set of observed states in the system backend. Through logical mapping, it transforms the user's potential shopping intentions into discrete latent variables and simultaneously transforms the preprocessed multimodal user feature sequences into observable symbolic sequences. This is done to mathematically construct a mapping architecture between the user's inner intentions and external behavioral manifestations. Secondly, the cloud server numerically initializes the initial state probability distribution, state transition probability matrix, and observation probability matrix. By setting initial weights for each transition and emission probability, it provides the necessary logical starting point for subsequent parameter optimization. Then, the cloud server uses the collected historical multimodal user feature sequences as the input sample set to guide the Baum-Welch algorithm to perform multiple rounds of iterative calculations. In each iteration, it uses the expectation-maximization principle to derive the expected values ​​of the forward and backward probabilities. This is done to discover the parameter path that best explains the historical behavioral trajectory through statistical methods, even when the hidden intentions cannot be directly observed.

[0048] Next, based on the calculation results generated in each iteration, the cloud server continuously corrects and updates the specific parameter values ​​of the initial state probability distribution, state transition probability matrix, and observation probability matrix in real time (for example, if historical data shows that users frequently place orders after browsing high-end fruits, the algorithm will increase the probability weight of the transition from "browsing intent" to "purchase intent"). This allows the model to dynamically adjust the accuracy of its understanding of user behavior patterns. Finally, after each parameter update, the cloud server calculates the change in parameter values ​​and quantizes and aligns it with a preset convergence threshold. When the difference in change shrinks to within the threshold range, the iteration stops and the trained Hidden Markov Model is output. This is done to ensure that the model locks the parameters in time after reaching statistical steady state, avoiding resource waste due to over-computation, while ensuring that the output model has robust interpretability for real-world scenarios.

[0049] As an example, the steps of sending the real-time checkout control command to the checkout terminal and the asynchronous marketing push command to the member terminal, so that the checkout terminal and the member terminal can perform corresponding cross-module collaborative actions, include: obtaining the hardware driver layer protocol parameters of the real-time checkout control command and converting the asynchronous marketing push command into an application layer network message; sending the real-time checkout control command containing the hardware driver layer protocol parameters to the checkout terminal through a preset transport layer security encryption protocol channel, so that the checkout terminal can perform pricing and discount deduction operations; and sending the application layer network message to the member terminal through a preset program push interface, so that the member terminal can display the preferred product discount information corresponding to the asynchronous marketing push command.

[0050] It should be noted that hardware driver layer protocol parameters refer to configuration data used to instruct the underlying hardware devices of the POS terminal (such as scanners, scales, or printers) to perform specific physical actions. These include, but are not limited to, device communication addresses, specific control command codes, and timing parameters such as baud rates. Application layer network packets refer to data units encapsulated according to specific application protocols (such as HTTP, MQTT, etc.) and capable of transmission over the network. They typically contain structured business data in JSON or XML format, used to transmit complex marketing logic instructions between the cloud and the terminal. The preset transport layer secure encryption protocol channel refers to an end-to-end logical link established based on security protocols such as TLS or SSL. By applying symmetric or asymmetric encryption to the data transmission process, it aims to prevent unauthorized interception or tampering of POS control instructions by third parties during the issuance process.

[0051] Pricing and discount deduction operations refer to the process by which the POS terminal, upon receiving a control instruction, calculates the amount based on the original price and weight data of the goods in the current transaction, combined with membership benefits and real-time marketing strategies, and automatically deducts the corresponding discount amount to generate the final payable amount. Preset program push interface refers to a standardized application programming interface provided by a mobile application (App) or mini-program platform (such as WeChat mini-program), used to deliver and present server-generated notifications or business data to users' mobile devices in real-time and asynchronously. Preferred product discount information refers to personalized promotional content generated based on big data mining of members' historical consumption behavior and profile characteristics, including but not limited to exclusive discounts, cash vouchers, or points rebate information for members' frequently purchased fresh produce.

[0052] Understandably, firstly, the cloud server parses and extracts fields from the generated instructions to obtain hardware driver layer protocol parameters (such as extracting specific baud rates and device communication codes for scanning docks or electronic scales) that can be recognized by the underlying hardware in the real-time POS control instructions. Simultaneously, it calls a data encapsulation algorithm to convert asynchronous marketing push instructions into application layer network packets conforming to network transmission formats (such as encapsulating marketing content into JSON format data packets). This ensures that the control signals are directly compatible with the physical hardware of the POS terminal and that the marketing content can be transmitted in a mobile network environment. Secondly, the cloud server establishes a pre-defined transport layer security encryption protocol channel (such as establishing a secure transmission link based on the TLS protocol) through a handshake connection. It then injects the real-time POS control instructions containing hardware driver layer protocol parameters into this channel and sends them to the POS terminal. This instructs the POS terminal to call the underlying hardware to perform pricing and discount deduction operations according to the protocol parameters, thereby ensuring the security of transaction data transmission while enabling remote automated intervention of the POS settlement process by the cloud. Finally, the cloud server activates the preset program push interface (such as calling the WeChat mini program subscription message interface) by sending a call request to the corresponding message server, and delivers the packaged application layer network message to the member terminal. This drives the member terminal to render and display the preferred product discount information corresponding to the asynchronous marketing push instruction on the front-end interface (such as pushing a limited-time discount coupon for the fruit category that the user frequently purchases). This achieves precise marketing outreach to specific members without interfering with the front-end cashier process.

[0053] This embodiment provides an integrated intelligent control method for fruit and fresh produce checkout. First, it acquires fresh produce transaction logs containing visual recognition information uploaded by the checkout terminal, and member interaction logs uploaded by the member terminal, constructing a comprehensive data foundation covering the transaction process and member behavior. Next, it performs multi-source data cleaning and structured alignment on the fresh produce transaction logs and member interaction logs to obtain fresh produce transaction feature sets and member interaction feature sets, thereby eliminating data noise and unifying analysis dimensions to ensure the accuracy of subsequent processing. Then, it performs cross-domain feature fusion and profile mining on the fresh produce transaction feature sets and member interaction feature sets to obtain multi-dimensional member profile features and real-time transaction context features. The system uses data collection to deeply reconstruct consumption scenarios and accurately identify user intent. Then, multi-dimensional member profile features and real-time transaction context features are input into a multi-task collaborative network for bidirectional instruction deduction, generating real-time checkout control instructions and asynchronous marketing push instructions. This ensures a high degree of logical consistency and bidirectional linkage between settlement intervention and marketing triggers. Finally, real-time checkout control instructions are sent to the checkout terminal, and asynchronous marketing push instructions are sent to the member terminal to drive the checkout terminal and member terminal to execute corresponding cross-module collaborative actions. This achieves full-link automated collaboration between checkout and member management in the fresh food retail scenario and reduces the reliance on manual intervention in the operation process.

[0054] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the integrated intelligent control method for fruit and fresh produce checkout in this application. Step S40 of the integrated intelligent control method for fruit and fresh produce checkout includes steps S41 to S45: Step S41: Input the multi-dimensional member profile features and the real-time transaction context features into the shared feature extraction layer for global feature mapping to obtain global shared semantic features; Step S42: The intent recognition branch is used to perform intent mapping on the globally shared semantic features to obtain a real-time cashier preference vector representing the probability distribution of the strategy. Step S43: Perform temporal inference on the globally shared semantic features through the behavior prediction branch to obtain a long-term marketing preference matrix representing multi-dimensional marketing attributes; Step S44: Based on the real-time cashier preference vector, dynamically sort the discount rules in the preset instruction rule base to generate real-time cashier control instructions; Step S45: Extract recommended product identifiers from the preset instruction rule base according to the long-term marketing preference matrix, and encapsulate and generate asynchronous marketing push instructions.

[0055] It should be noted that globally shared semantic features refer to a high-level data representation obtained by uniformly fusing and compressing long-term customer profile tags with current checkout environment data through an underlying neural network. The strategy probability distribution refers to the system's quantitative assessment of the likelihood of customers accepting various available checkout discounts or promotional rules in the backend. It is typically represented as a set of values ​​summing to 100%, intuitively reflecting the customer's preference or priority for different discount options at the current checkout moment. The real-time checkout preference vector is a one-dimensional mathematical array calculated and output by the model's intent recognition branch. It directly carries the aforementioned "strategy probability distribution" values, expressing which specific checkout deduction rule is most suitable for triggering the current order. Multi-dimensional marketing attributes refer to multiple different aspects of business indicators that need to be considered when predicting customers' future long-term consumption behavior, such as the customer's willingness to purchase a certain type of fresh produce, their sensitivity to price changes, and their expected repurchase cycle. These together constitute a three-dimensional long-term marketing reference dimension.

[0056] The long-term marketing preference matrix is ​​a two-dimensional data array derived from the model's behavior prediction branch. Its row dimension typically represents different fresh produce categories, while the column dimension corresponds to the aforementioned "multi-dimensional marketing attributes." This structured approach provides a panoramic view of customers' comprehensive consumption intentions for various products over a future period. The preset instruction rule base is a pre-configured business strategy control center in the system backend. It stores all promotional activity conditions, rigid discount calculation logic, and available marketing script templates, categorized into different types. It represents the "rigid execution standard" that translates the "soft probabilities" calculated by the algorithm into specific business actions. The recommended product identifier is a unique identification code (such as SKU code) assigned by the system to each specific fresh produce (e.g., cherries from a specific origin and specification) within the vast product master data. Extracting this identifier means that, after calculation, the system has accurately identified the unique marketing target to be pushed to the customer from a massive amount of products.

[0057] Understandably, the cloud server first inputs multi-dimensional member profile features and real-time transaction context features into the shared feature extraction layer of the multi-task collaborative network front end. This layer performs underlying global feature mapping calculations on the input data, deeply integrating these two types of heterogeneous data from different sources and converting them into unified global shared semantic features (for example, compressing and reconstructing the customer's long-term preference tags and the current settlement environment matrix into a set of general underlying tensors). This is done to extract a general data base with strong generalization ability under different task splitting conditions to avoid redundant calculations. Secondly, the cloud server synchronously distributes the generated globally shared semantic features to two independent computing branches of the network. On the one hand, the intent recognition branch performs an instant intent mapping inference on the feature to accurately derive a real-time checkout preference vector that internally represents the probability distribution of the applicability of various checkout strategies (e.g., outputting a one-dimensional array containing the trigger probabilities of multiple discount rules). On the other hand, the behavior prediction branch performs long-term time-series inference operations on the same feature in parallel to derive a long-term marketing preference matrix that can simultaneously represent various multi-dimensional marketing attributes such as product category preference (e.g., outputting a multi-row, multi-column two-dimensional array indicating future consumption trends). This is done to utilize the same data source to simultaneously address current decisions and future predictions.

[0058] Then, the cloud server carries the aforementioned real-time POS preference vector into the system's backend preset instruction rule library. It directly uses the probability values ​​of each strategy in the vector to dynamically score and reorder all available discount rules in the library, selecting the optimal rule with the highest priority to directly generate real-time POS control instructions for controlling hardware actions (e.g., generating a machine pricing command for a full discount based on the highest probability item). This transforms the abstract instantaneous probability assessment directly into physical billing logic. Finally, the cloud server uses the synchronously generated long-term marketing preference matrix to perform feature comparison and targeted retrieval within the same preset instruction rule library. It accurately extracts the recommended product identifier with the highest consistency with the long-term attributes represented by the matrix. This identifier and its accompanying content are then formatted and encapsulated at the network layer, ultimately generating asynchronous marketing push instructions for remotely reaching users (e.g., locking a specific strawberry code and packaging it into a mobile app pop-up message). This ensures that physical-level pricing intervention and virtual-level marketing deployment are completed simultaneously within a single algorithm processing flow.

[0059] As an example, the construction steps of the multi-task collaborative network include: constructing a shared feature extraction layer based on linear layers and convolutional layers; constructing an intent recognition branch based on a multi-head attention mechanism layer, a fully connected network layer, and a normalized activation function; constructing a behavior prediction branch based on a residual network layer, a gated recurrent unit network layer, and a tensor reconstruction layer; and combining the shared feature extraction layer, the intent recognition branch, and the behavior prediction branch to obtain the multi-task collaborative network.

[0060] It's important to note that a linear layer is the fundamental structure in a neural network that performs linear spatial transformations. It uses a weight matrix and bias vector to perform a weighted summation of the input data features, primarily used to achieve linear mapping between data dimensions and preliminary feature dimensionality reduction or enhancement. A convolutional layer is the core component in a deep learning network used to extract local spatial features. It uses a pre-sized convolutional kernel to perform a local sliding window dot product operation on the input data matrix, effectively reducing the overall model's parameter computation while preserving the correlation of the feature space structure. A shared feature extraction layer is a common network structure deployed at the front end of a multi-task network model. It is mainly responsible for performing unified high-dimensional semantic mining on the original joint features of the input, providing a consistent and basic underlying common data representation for subsequent parallel subtasks to avoid redundant consumption of computational resources.

[0061] A multi-head attention mechanism layer refers to a network architecture capable of simultaneously extracting information weights from multiple different representation subspaces. It calculates the internal correlations of the input feature sequences by running multiple independent attention modules in parallel, thereby comprehensively and accurately focusing on the key feature information most relevant to the current prediction target. A fully connected network layer is a classic network component where each neuron in the current layer is physically connected to all neurons in the previous layer. It is typically located at the end of the feature extraction or branching module, responsible for combining and integrating the previously extracted distributed local features into a global dimension and a deep mapping. A normalized activation function is a mathematical transformation module that performs probability distribution mapping on the computation results at the neural network output. It can compress and convert unstructured real-number feature vectors into a standardized probability distribution form with a sum of one (e.g., Softmax), thus intuitively expressing the confidence level of various decision results or intention classifications.

[0062] Residual network layers are deep network components that incorporate skip connections (or shortcut connections). This structure allows input signals to be directly added across one or more nonlinear layers and passed to subsequent layers, effectively preserving original static properties and mitigating feature degradation and gradient vanishing problems during deep model training. Gated recurrent unit (ROU) network layers are lightweight recurrent neural network variants specifically designed to process and capture dependencies in time-series data. They incorporate reset and update gate mechanisms, adaptively preserving long-term historical key time-series states and discarding redundant interference data when extracting sequence features. Tensor reconstruction layers are network computation modules specifically designed to change the shape and dimension of data. Their main function is to forcibly decompose and rearrange the originally tightly compressed high-dimensional complex feature data according to the format required by business needs (e.g., rows representing categories, columns representing various marketing indicators), thereby outputting a clearly structured two-dimensional attribute matrix that can directly guide business actions.

[0063] Understandably, firstly, the cloud server instantiates linear and convolutional layers sequentially within the underlying code framework and connects these two network components in series to build a shared feature extraction layer at the very front of the entire model (for example, connecting the linear layer performing spatial transformation before the convolutional layer performing local feature extraction). This is done to pre-process the complex multidimensional input data to extract underlying fundamental patterns before the network performs task splitting, thereby reducing redundant computation in the overall architecture. Secondly, the cloud server sequentially calls and cascades the multi-head attention mechanism layer, the fully connected network layer, and the normalized activation function according to the logical order of data forward propagation, completing the independent structural construction of the intent recognition branch (for example, directly connecting the attention computation graph used to focus on key information to the neural network used for dimensionality reduction and ending with a normalization function). This is done to customize a dedicated channel that best suits the one-dimensional dimensionality reduction inference characteristics for the single decision objective of transient intent.

[0064] Then, the cloud server synchronously stacks and splices residual network layers, gated recurrent unit network layers, and tensor reconstruction layers sequentially at the code level to construct a behavior prediction branch that is spatially parallel to the aforementioned intent recognition branch (for example, a recurrent unit with memory and forgetting mechanisms is specially configured to handle long-term behavior profiling, and a code layer responsible for spatial dimensionality expansion and reorganization is connected to the end of the computation graph). This is done to customize a dedicated computation path that can deeply process temporal dependencies and ultimately output a two-dimensional array for the complex task of long-term preference prediction. Finally, the cloud server performs the integration and assembly operation of the overall network topology architecture. By configuring the internal routing distribution mechanism, the output tensor nodes of the shared feature extraction layer are simultaneously connected in parallel to the bottom input ends of the intent recognition branch and the behavior prediction branch (for example, so that the same general feature matrix extracted at the bottom layer can be losslessly copied and simultaneously fed into the two parallel branches at the upper layer). This is done to seamlessly stitch the three model components that were previously constructed independently into a unified end-to-end computation graph in physical and logical terms, and finally combine them to obtain a multi-task collaborative network that can concurrently process dual inference objectives using the same underlying data payload.

[0065] Please refer to Figure 3 , Figure 3This is a schematic diagram of the multi-task collaborative network provided in Embodiment 2 of the integrated intelligent control method for fruit and fresh produce checkout in this application. The structure starts from the left input end, where multi-dimensional member profile features and real-time transaction context features are input into a shared feature extraction layer composed of linear and convolutional layers. This layer performs global feature mapping calculations, fusing heterogeneous data and outputting a globally shared semantic feature. This globally shared semantic feature is then synchronously distributed to two parallel computation branches: the upper intent recognition branch sequentially cascades a multi-head attention mechanism layer, a fully connected network layer, and a normalized activation function, aiming to perform intent mapping inference on the features and ultimately output a real-time checkout preference vector representing the probability distribution of the strategy; the lower behavior prediction branch first splits the profile features into a static attribute feature subset and a dynamic preference feature subset through a profile feature segmentation module, and feeds them to a residual network layer and a gated recurrent unit network layer for parallel computation. After element-wise addition and fusion, the computation results are spatially expanded and rearranged through a tensor reconstruction layer at the end, ultimately outputting a long-term marketing preference matrix where the row dimension represents candidate fresh produce categories and the column dimension represents diverse marketing attributes.

[0066] As an example, the training steps of the multi-task collaborative network include: aligning historical fresh food transaction records with historical member interaction records to construct historical real-time transaction context features; extracting historical multi-dimensional member profile features corresponding to the historical real-time transaction context features from the historical database, and obtaining the corresponding real checkout strategy execution records and real repurchase category distribution matrices; encoding the real checkout strategy execution records and the real repurchase category distribution matrices as checkout preference hot tags and marketing preference soft tags, respectively; synchronously inputting the historical real-time transaction context features and the historical multi-dimensional member profile features into the multi-task collaborative network to be trained to obtain a predicted checkout preference vector and a predicted marketing preference matrix; calculating a first loss value between the predicted checkout preference vector and the checkout preference hot tags based on the cross-entropy loss function; calculating a second loss value between the predicted marketing preference matrix and the marketing preference soft tags based on the mean squared error loss function; weighted summing of the first loss value and the second loss value to obtain a total loss value; and updating the parameters of the multi-task collaborative network to be trained using the backpropagation algorithm based on the total loss value until a preset convergence condition is met, thus obtaining the trained multi-task collaborative network.

[0067] It's important to clarify that historical fresh produce transaction records refer to the basic transaction data of customers who purchased fresh produce at offline stores, as recorded by the system. This includes objective physical information such as the specific types of fruits and vegetables purchased, the actual weight, and the payment method used. Historical member interaction records refer to the operational trajectory of customers on online platforms (such as mobile mini-programs) within a certain period of time, including behavioral logs such as which pages were viewed, how long they stayed, and which coupons were claimed. Historical real-time transaction context features refer to the data package obtained by splicing the physical product information of a past checkout order with the customer's online operational trajectory just before checkout. It's like accurately capturing and archiving the real-time context of a past transaction. The historical database refers to the underlying cloud or physical storage warehouse used by the system backend to centrally and persistently store massive amounts of past business data. It contains a large amount of raw transaction records, behavioral logs, and structured member history files.

[0068] Historical multi-dimensional member profile features refer to a set of structured tags that the system has drawn up for a customer's long-term consumption habits at the same point in time when a past transaction occurred. This reflects the customer's perceived category preferences and spending power at that time. Actual checkout strategy execution records refer to the specific discount rules or pricing schemes that the cash register ultimately adopted and deducted fees in a past actual checkout scenario. This serves as the absolute standard answer for verifying the accuracy of the model's immediate decisions. The actual repeat purchase category distribution matrix refers to the array of actual proportions and frequencies of various fresh produce items purchased again by the customer over a subsequent long period, obtained by retrospectively tracking and statistically analyzing data from a past point in time. This serves as the standard answer for verifying the accuracy of the model's long-term predictions.

[0069] The "Hot Checkout Preference Tag" refers to converting the unique, actual checkout strategy into a computer-recognizable exclusive coding format (setting the position corresponding to the actual rule to 1 and all others to 0), explicitly indicating the only correct option to the model at that moment. The "Soft Marketing Preference Tag" converts the actual repurchase distribution into a set of decimal or probability values ​​containing proportional relationships. It preserves the diversity of long-term customer interests and the distinction between different levels of preference, rather than a simple black-and-white division. The predicted checkout preference vector is a one-dimensional mathematical array derived by the intent recognition branch of the multi-task collaborative network during training after receiving input samples. It represents the probability distribution of the model's guess regarding which checkout discount strategy should be used. The predicted marketing preference matrix is ​​a multi-row, multi-column two-dimensional mathematical array derived by the behavior prediction branch of the model during training based on the same input samples. It represents the model's comprehensive prediction of the customer's long-term fresh food purchase intentions and attributes.

[0070] The first loss value refers to the deviation between the model's predicted payment preference and the standard answer (unique hot label), precisely calculated using a specific mathematical formula (cross-entropy loss function). It is specifically used to quantify the degree of prediction error of the model in the task of instantaneous decision-making. The second loss value refers to the distance difference between the model's predicted marketing preference and the standard answer (soft label), calculated using another formula (mean squared error loss function). It is specifically used to quantify the degree of deviation of the model in the task of long-term behavior prediction. The total loss value is a global error index obtained by adding and fusing the first loss value (measuring the instantaneous task) and the second loss value (measuring the long-term task) according to a set proportional weight. It is the core basis for guiding the entire network to perform self-correction and parameter adjustment. The preset convergence condition refers to a stop warning line set in advance by the system to determine whether the model has completed learning (e.g., the total error no longer decreases significantly after several consecutive attempts). It aims to ensure that the model ends training in time after reaching the optimal statistical stability state, preventing overlearning or rote memorization.

[0071] Understandably, firstly, the cloud server uses timestamps as an alignment benchmark to matrix-concatenate historical fresh food transaction records and historical member interaction records at the underlying data dimension, thereby constructing historical real-time transaction context features. This is done to accurately reproduce the complete business environment at the moment a past transaction occurred in mathematical space. Secondly, the cloud server uses associated user identifiers as indexes to selectively extract historical multi-dimensional member profile features from the historical database that occurred at the same time point as the historical transaction. Simultaneously, it retrieves the actual cashier strategy execution records that occurred in the business system at that time, as well as the actual repeat purchase category distribution matrix obtained through long-term tracking and statistics. Next, the cloud server transforms the extracted actual cashier strategy execution records into unique hot tags for cashier preferences, where only the hit items are 1 and the rest are all 0. It also transforms the actual repeat purchase category distribution matrix into marketing preference soft tags that express numerical proportions. This is done to completely convert the original business facts into a mathematical reference that the neural network can directly use as a reference standard when performing backpropagation.

[0072] Then, the cloud server simultaneously feeds the previously constructed historical real-time transaction context features and historical multi-dimensional member profile features as dual input data to the underlying layer of the multi-task collaborative network to be trained. This drives the model to perform a forward inference calculation using the current initial parameters, thereby deriving the predicted checkout preference vector and the predicted marketing preference matrix at two parallel outputs. Next, the cloud server calls the cross-entropy loss function formula to calculate the deviation between the predicted checkout preference vector and the absolutely correct checkout preference one-hot label to obtain the first loss value; in parallel, it calls the mean squared error loss function formula to measure the geometric distance between the predicted marketing preference matrix and the marketing preference soft label to obtain the second loss value.

[0073] Subsequently, the cloud server performs a weighted summation of the first and second loss values ​​according to the pre-allocated task weights, comprehensively integrating them to obtain a global total loss value. This is done to unify the errors of instantaneous and long-term tasks into a global objective, preventing the model from being biased in training. Finally, guided by the total loss value, the cloud server activates the backpropagation algorithm to calculate gradients layer by layer along the network topology from the output to the input, and fine-tunes and updates the weight parameters of all nodes within the multi-task collaborative network to be trained. Then, it iterates by continuously feeding new batches of samples until the calculated overall loss change level reaches the minimum stopping boundary specified by the preset convergence condition. At this point, the calculation is terminated and the parameters are locked, finally solidifying the trained multi-task collaborative network.

[0074] As an example, the intent recognition branch includes a multi-head attention mechanism layer, a fully connected network layer, and a normalized activation function. The step of mapping the globally shared semantic features through the intent recognition branch to obtain a real-time cashier preference vector representing the policy probability distribution includes: inputting the globally shared semantic features into the multi-head attention mechanism layer for weight allocation to obtain intent-focused features; inputting the intent-focused features into the fully connected network layer for dimensionality reduction to obtain a low-dimensional intent feature representation; and performing probability normalization processing on the low-dimensional intent feature representation through the normalized activation function to obtain a real-time cashier preference vector representing the policy probability distribution.

[0075] It's important to clarify that intent-focused features refer to the system's ability to identify and amplify the most important and crucial data related to the current checkout (such as the cherries just picked up or the VIP gold card in the account) through an "attention mechanism" when faced with a large amount of disorganized information, while automatically ignoring irrelevant details (such as the onions bought last week). This extracts a highly condensed set of features that directly address the core motivation for the current payment. Intent-low dimensional feature representation refers to the process where, to make subsequent calculations faster and more accurate, the system compresses and packages the potentially large and complex high-dimensional focused data extracted in the previous step using a neural network, removing unnecessary information and redundant dimensions, ultimately generating a concise, efficient, and highly representative set of values.

[0076] Understandably, firstly, the cloud server directly inputs the globally shared semantic features from the underlying layer into the multi-head attention mechanism layer. This mechanism utilizes parallel computational graph logic to perform deep correlation scoring and adaptive weight allocation on each dimension of the input features (e.g., automatically assigning higher weights to data highly relevant to the current checkout action while simultaneously suppressing irrelevant historical background noise). This accurately filters out redundant interference and extracts intent-focused features, ensuring the model's computational power is absolutely concentrated on the most critical payment decision factors. Secondly, the cloud server feeds the extracted intent-focused features into the fully connected network layer. Through matrix multiplication and addition operations between neurons within the layer, it performs forced dimensionality reduction on the high-dimensional focus features (e.g., mathematically compressing the large and sparse attention output array into a compact numerical structure with fewer nodes). This eliminates redundant dimensions in the data representation and generates a simplified low-dimensional intent feature representation. This significantly reduces the computational load at the network's decision-making end and prevents feature overfitting. Finally, the cloud server directly calls the normalization activation function to perform probability normalization processing on each discrete real number inside the aforementioned intentional low-dimensional feature representation (for example, scaling the originally unbounded feature values ​​proportionally and forcibly mapping them to a set of probability distributions with a total of strictly 100%), thereby smoothly converting the abstract deep network tensor into a real-time cashier preference vector representing the probability distribution of the strategy. This is done in order to finally fix the black-box network output into a standardized business probability array that can be directly read, compared, and sorted by the downstream business rule base.

[0077] As an example, the behavior prediction branch includes a residual network layer, a gated recurrent unit network layer, and a tensor reconstruction layer. The step of inputting the multi-dimensional member profile features into the behavior prediction branch of the multi-task collaborative network for temporal inference to obtain a long-term marketing preference matrix includes: dividing the multi-dimensional member profile features into a subset of static attribute features and a subset of dynamic preference features; inputting the subset of static attribute features into the residual network layer for nonlinear semantic mapping to obtain static attribute features; inputting the subset of dynamic preference features into the gated recurrent unit network layer for temporal dependency calculation to obtain dynamic temporal features; adding the static attribute features and the dynamic temporal features element-wise to obtain a hybrid prediction feature; and performing spatial dimensionality expansion and tensor reconstruction on the hybrid prediction feature through the tensor reconstruction layer to obtain a long-term marketing preference matrix in which the row dimension represents candidate fresh food categories and the column dimension represents multiple marketing attributes.

[0078] It's important to clarify that the static attribute feature subset refers to the set of inherent user tags in a multi-dimensional member profile that remain stable over a long period and are not prone to frequent changes. Examples include a customer's registered store, basic membership level, or long-term fixed spending power stratification. These constitute the underlying static background parameters of the behavior prediction model. The dynamic preference feature subset refers to the set of time-sensitive data slices in the multi-dimensional member profile that are constantly updated based on the customer's recent browsing, interaction, or actual transaction behavior. For example, the fluctuation in a customer's click frequency on a particular seasonal fruit within the past week. These are specifically designed to capture the rapidly changing short-term interest shifts of users. Static attribute features, after being sent to the residual network layer for deep nonlinear spatial transformation, represent a high-order and dense data representation. This not only filters out noise from the original shallow data but also further strengthens the anchoring effect of the customer's inherent characteristics on long-term consumption decisions. Dynamic temporal features refer to the hidden state vectors with clear temporal evolution patterns extracted from a subset of dynamic preference features after being processed by the internal memory and forgetting gating mechanism of the gated recurrent unit (GRU). It accurately quantifies the evolution trajectory and trend of customer consumption interests over time.

[0079] Hybrid predictive features refer to a comprehensive data carrier generated by forcibly superimposing and matrix-merging static attribute features representing the inherent characteristics of customers and dynamic time-series features representing recent behavioral trends element-by-element in the underlying mathematical tensor space. At the algorithmic level, it perfectly balances the dual dimensions of long-term background and short-term trends. Candidate fresh produce categories refer to various specific fresh produce categories available for marketing pushes from the full product master data in the system backend (e.g., imported cherries of specific specifications or local organic cabbage). These are mapped to the "row dimension" of the final output matrix, representing the specific physical targets that customers may purchase in the future, as predicted by the model. Multiple marketing attributes refer to the comprehensive output of various business prediction metrics (e.g., repurchase probability score, price sensitivity elasticity coefficient, expected repurchase time span, etc.) when comprehensively evaluating customers' purchase intentions for specific fresh produce categories. These are mapped to the "column dimension" of the output matrix, providing multi-dimensional decision support for subsequent automated and precise pricing and coupon issuance.

[0080] Understandably, firstly, the cloud server, based on the different data update frequencies or time stability, directly segments and divides the multi-dimensional member profile features into static attribute feature subsets and dynamic preference feature subsets in the underlying processing module (for example, forcibly separating the fixed member level data from the recently frequently changing page click records at the field level). This is done in order to physically separate the long-term background from the short-term trends so that different model calculation channels can be scheduled for specialized mining later. Secondly, the cloud server directs the segmented static attribute feature subset into the residual network layer. Through matrix multiplication and jump connection operations within this code layer, nonlinear semantic mapping is performed (e.g., using deep computational logic to directly extract the true consumption power hidden under basic labels). This is done to strengthen and extract the deep, solidified foundation that is not affected by short-term environmental fluctuations. Simultaneously, the cloud server feeds the dynamic preference feature subset into a parallel gated recurrent unit network layer. Using its internal update and reset gate components, state memory calculations are performed on the input sequence (e.g., based on the order of timestamps, the evolution of customers' recent interest from pork to seafood is deduced). This is done specifically to capture the temporal dependencies of feature data over time.

[0081] Then, in the underlying tensor computation graph, the cloud server performs a strict element-wise addition operation on the static attribute feature matrix and the dynamic temporal feature matrix extracted independently from the two branches mentioned above (for example, directly adding the real values ​​at corresponding positions within two multidimensional arrays of identical size). This is done to forcibly mathematically bind and integrate the user's inherent historical habits and recent behavioral trends at the lowest level of the algorithm. Finally, the cloud server calls the tensor reconstruction layer component at the end of the network to perform hard spatial stretching and row and column data rearrangement on the added and merged mixed prediction features (for example, forcibly converting the originally complex abstract feature package into a two-dimensional data grid with fixed length and width dimensions). This directly outputs a long-term marketing preference matrix with row dimensions corresponding to each candidate fresh food category and column dimensions corresponding to each multi-dimensional marketing attribute. This is done to completely convert the complex high-dimensional tensors inside the neural network into a standardized business dashboard that is format-aligned and can be directly read by downstream automated marketing systems and used for coupon pricing.

[0082] As an example, the step of dynamically sorting discount rules in a preset instruction rule base based on the real-time cashier preference vector to generate a real-time cashier control instruction includes: obtaining the target member level identifier and the current purchase category identifier transmitted by the cashier terminal, and concatenating them into a basic constraint query key value; performing precondition filtering in the hash index tree of the preset instruction rule base based on the basic constraint query key value to obtain a candidate discount strategy set; extracting the confidence weights corresponding to the candidate discount strategy set from the real-time cashier preference vector, and scoring and sorting the candidate discount strategy set according to the confidence weights, selecting the strategy ranked first as the target discount calculation strategy; and generating a real-time cashier control instruction containing discount amount deduction parameters based on the target discount calculation strategy.

[0083] It's important to note that the target member level identifier is a unique identification code assigned by the system to a customer within the current membership operation system's hierarchical structure (e.g., a numerical number representing "Ordinary Member," "Silver Card," or "Gold Card"). This directly determines the basic level of discounts the customer can enjoy. The current purchase category identifier is a unique system code assigned by the system to the category or subcategory of the product the customer is currently checking out (e.g., an identification code representing "local fresh leafy vegetables" or "imported seafood"). This code is used to precisely define the physical attributes of the goods in this transaction within the underlying data. The basic constraint query key is a composite key created by combining the level code representing the identity with the category code representing the product using specific underlying string concatenation rules. It's like a special "digital key," specifically used to quickly retrieve discount thresholds from a vast amount of rules that only apply to customers with a specific identity who purchase specific products.

[0084] A hash index tree is a high-speed data structure used in the pre-set instruction rule base in the background. It maps complex promotional rule preconditions to short hash values ​​and arranges them layer by layer in a tree-like branching structure, allowing the system to instantly intercept and filter out invalid activity rules that do not meet the current transaction conditions within milliseconds. A candidate discount strategy set refers to the theoretically available discount schemes that meet the triggering conditions and are selectable for a specific customer level when purchasing a specific batch of goods, after precondition filtering and preliminary screening. It is equivalent to a "menu of valid alternative discounts" provided by the system. Confidence weight is a probability assessment value calculated and assigned to each specific discount scheme by the neural network model in the preceding operations. It precisely quantifies the degree to which the artificial intelligence model is confident that "if this discount is adopted immediately, the customer's conversion intention is strongest and it best meets the business requirements."

[0085] The target discount calculation strategy refers to the pricing logic that stands out from all alternative discount schemes after being scored and ranked by the model using probability values ​​(e.g., "10 yuan off for purchases over 100 yuan"), and it is the only execution scheme ultimately adopted for this order checkout. The discount amount deduction parameter refers to the specific value or floating-point discount ratio field that the system directly calculates based on the final selected pricing logic, and which needs to be physically reduced from the total amount payable in the current order. It is a rigid machine instruction variable that the POS hardware can directly read, parse, and execute the deduction action accordingly.

[0086] Understandably, firstly, the cloud server receives the target member level identifier and the current purchase category identifier directly transmitted from the underlying POS terminal in real time. Then, in the underlying code, these two independent code fields are forcibly concatenated into a unique basic constraint query key value according to a preset character order (e.g., concatenating "V3" representing gold card members with "C08" representing fresh seafood into the combined string "V3-C08"). This is done to bind and lock the discrete identity access thresholds and product restrictions in the current transaction into a precise composite search index. Secondly, the cloud server carries this concatenated basic constraint query key value directly into the preset instruction rule base in the backend. In the underlying hash index tree data structure, it performs extremely fast key-value matching and pre-condition filtering (e.g., instantly comparing and eliminating all invalid activity rules that only apply to ordinary members or non-seafood categories along the tree structure). This accurately screens and derives a set of candidate discount strategies containing all currently compliant and available rules. This is done to leverage the efficient query characteristics of hash trees to lock in a pool of real alternative solutions from a massive amount of activity within milliseconds.

[0087] Then, the cloud server accurately extracts the confidence weight values ​​corresponding to each specific strategy in the aforementioned candidate discount strategy set from the real-time checkout preference vector derived from the neural network. Based on these values ​​representing the probability of AI recommendation, the server performs mandatory scoring and descending sorting on all alternative solutions within the candidate discount strategy set. The top-ranked strategy is then directly selected as the final target discount calculation strategy for this transaction (for example, in the two compliant alternatives of "spend more and save" and "discount", the "discount" rule, which has been assigned the highest confidence score by the model, is directly chosen). This is done to ensure that the soft probability assessment output by the AI ​​model truly dominates and determines the final hard business decision. Finally, the cloud server strictly follows the locked target discount calculation strategy, parses and extracts the specific reduction value at the underlying level, and encapsulates it according to the communication protocol to generate a real-time cashier control instruction containing specific discount deduction parameters (for example, generating a machine communication message carrying the variable "total price reduced by 20 yuan"). This is done in order to completely transform the abstract algorithm decision into rigid hardware execution code that can be directly issued and drive the physical cash register to perform accurate deduction actions.

[0088] As an example, the step of extracting recommended product identifiers from the preset instruction rule base based on the long-term marketing preference matrix includes: extracting the comprehensive preference score of each candidate fresh food category from the long-term marketing preference matrix; selecting the target fresh food category with the highest comprehensive preference score and extracting the target marketing attribute vector corresponding to the target fresh food category from the long-term marketing preference matrix; jointly encoding the target fresh food category and the target marketing attribute vector to generate a product retrieval feature vector; calculating the spatial similarity between the product retrieval feature vector and each preset product profile vector in the product mapping space included in the preset instruction rule base; selecting the product profile vector with the highest spatial similarity and using the unique data code bound to the product profile vector as the recommended product identifier.

[0089] It's important to note that the comprehensive preference score is a quantifiable score assigned to each candidate product category by the system after integrating and calculating the massive matrix data output by the behavioral prediction network. It comprehensively reflects the strength of a customer's overall potential willingness to purchase this type of fresh produce in the future. The target fresh produce category refers to the category that stands out from all candidate product types after system ranking and comparison, ranking first in comprehensive preference score (e.g., "high-end imported fruits" or "deep-sea fish"). It represents the macro-level product direction the system will focus on in this asynchronous marketing push. The target marketing attribute vector is a set of one-dimensional mathematical arrays specifically extracted from the long-term prediction matrix for the target fresh produce category. It details the customer's multi-dimensional consumption tendencies unique to this specific category (e.g., price sensitivity coefficient and preferred packaging specifications for high-end fruits). The product retrieval feature vector refers to a composite tensor structure generated by the system after mathematically forcibly packaging and jointly encoding the selected macro-category information and micro-attribute tendency values. It is equivalent to a high-dimensional digital retrieval key specially forged by the system to accurately retrieve physical products from a huge product database.

[0090] The product mapping space refers to a high-dimensional mathematical virtual coordinate system pre-constructed using vectorization technology in the system's backend instruction rule base. In this space, every specific physical product sold in a real store (such as cherries from a specific origin or weighing a specific gram) is transformed and placed at mathematical points with precise geometric coordinates. The preset product profile vector refers to the ready-made coordinate data package representing each real physical product (i.e., a specific SKU) in the aforementioned high-dimensional mapping space. It objectively and digitally quantifies the inherent commercial and physical attributes of the specific product (such as purchase cost, current inventory, and shelf life). Spatial similarity refers to the geometric distance or directional angle between two multi-dimensional coordinate points in the aforementioned high-dimensional mathematical coordinate system: the product retrieval feature vector representing the customer's abstract needs and the preset product profile vector representing the physical product. The larger this value (i.e., the closer the distance and the smaller the angle), the higher the perfect match between the physical product and the customer's potential needs.

[0091] Understandably, firstly, the cloud server directly reads the previously derived long-term marketing preference matrix and precisely extracts the one-dimensional comprehensive preference score corresponding to each candidate fresh food category from this two-dimensional array (for example, directly extracting the quantitative values ​​of purchase intention for various business categories such as apples and beef from the prediction results). This is done to establish a unified comparison benchmark to measure the potential customer conversion probability of each macro category. Secondly, the cloud server performs extreme value optimization and descending order comparison on all extracted scores, directly selecting the target fresh food category with the absolute highest comprehensive preference score. Then, it immediately extracts the target marketing attribute vector that strictly corresponds to the target fresh food category from the long-term marketing preference matrix (for example, after locking in "deep-sea fish" with the highest score, it specifically extracts detailed feature arrays such as price sensitivity elasticity and preferred weight for this category). This is done to further obtain a micro-level three-dimensional profile of this direction after determining the macro-level main product direction.

[0092] Then, the cloud server invokes the underlying network coding logic to perform forced assembly and joint coding operations on the selected target fresh produce category data and the extracted target marketing attribute vector at the mathematical tensor level. This directly generates a product retrieval feature vector with a specific dimensional structure (for example, packaging the category code and various prediction coefficients into a high-dimensional digital retrieval key). This completely transforms the abstract business prediction results into a standard computational payload that can be used for underlying spatial comparison. Next, the cloud server, carrying the generated product retrieval feature vector, directly accesses the pre-built high-dimensional product mapping space within the preset instruction rule base. It invokes specific distance or angle mathematical formulas to calculate the spatial similarity between the product retrieval feature vector and each existing preset product profile vector scattered in the space (for example, using the cosine formula to calculate the geometric proximity between the high-dimensional coordinate points representing customer demand characteristics and the high-dimensional coordinate points representing each physical product). This is done to move beyond the traditional rigid keyword filtering and utilize the spatial positional relationships of deep features to achieve precise anchoring between user intent and physical goods.

[0093] Finally, the cloud server sorts all the calculated similarity values ​​and directly selects the preset product profile vector with the highest absolute value of spatial similarity (i.e., the closest spatial location). It then extracts and outputs the unique data code that is strongly bound to the underlying layer as the recommended product identifier for this asynchronous marketing push (for example, locking the salmon profile of a specific origin and specification that is closest in the mapping space and directly extracting its unique SKU barcode). This is done in order to reliably translate the extremely complex high-dimensional vector calculation at the underlying layer into a concrete physical product credential that can be directly called and rendered by the downstream display system.

[0094] This embodiment first synchronously inputs multi-dimensional member profile features and real-time transaction context features into a shared feature extraction layer for underlying feature fusion and mapping calculations, generating a unified globally shared semantic feature. This extracts a common basic data payload, avoiding computational redundancy during subsequent multi-task processing. Second, the intent recognition branch performs real-time intent inference on this globally shared semantic feature, deriving a real-time checkout preference vector representing the probability distribution of various strategies. This accurately quantifies the applicability of different discount schemes at the moment of checkout. Simultaneously, the behavior prediction branch performs long-term temporal inference on the same globally shared semantic feature in parallel, deriving a long-term marketing preference matrix representing multi-dimensional marketing attributes. This allows for a comprehensive prediction of customers' future long-term consumption trends using a single feature source. Then, based on the values ​​in the real-time checkout preference vector, various discount rules are dynamically scored and reordered in a preset instruction rule base, directly generating real-time checkout control instructions. This rapidly transforms the abstract probabilities output by the model into rigid execution logic for controlling instant deductions. Finally, based on the generated long-term marketing preference matrix, the most suitable recommended product identifier is compared and extracted from the preset instruction rule base, and then encapsulated to generate asynchronous marketing push instructions. This ensures that future precise marketing actions are perfectly taken into account and implemented in the same business simulation flow.

[0095] For example, to help understand the implementation process of the integrated intelligent control method for fruit and fresh produce payment after combining this embodiment with the above embodiment one, please refer to... Figure 4 , Figure 4 This is a schematic diagram of the system architecture of the integrated intelligent control method for fruit and fresh produce payment provided in Embodiment 2 of this application. Specifically: The core architecture of this diagram uses a cloud server as the central processing node. A data acquisition module synchronously receives fresh food transaction logs transmitted from POS terminals containing visual recognition cameras, as well as member interaction logs uploaded from member terminals such as mobile apps. The collected raw logs enter a data preprocessing module for cleaning and structure alignment, and are then passed to a feature mining module for cross-domain feature fusion and profile mining. During processing, this module performs historical record storage and retrieval and profile alignment with the historical database at the bottom, generating multi-dimensional member profile features and real-time transaction context features. A multi-task collaborative network, acting as the system's decision-making brain, receives these features and interacts with a preset instruction rule base for dynamic discount sorting, etc. Through bidirectional inference, the prediction results are output to the instruction generation module. The instruction generation module generates specific business code under the rule mapping of the preset instruction rule base and delivers it to the instruction distribution module for execution and distribution. Finally, real-time POS control instructions containing discount parameters are sent to the POS terminals, while asynchronous marketing push instructions containing recommended product identifiers are sent to the member terminals, driving corresponding cross-module collaborative actions on both ends.

[0096] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the integrated intelligent control method for fruit and fresh produce cashier in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0097] This application also provides an integrated intelligent control device for fruit and fresh produce payment; please refer to [reference needed]. Figure 5 The integrated intelligent control device for fruit and fresh produce checkout includes: The data acquisition module 10 is used to acquire fresh food transaction logs containing visual recognition information uploaded by the POS terminal, and member interaction logs uploaded by the member terminal; Data preprocessing module 20 is used to perform multi-source data cleaning and structure alignment on the fresh food transaction log and the member interaction log to obtain fresh food transaction feature set and member interaction feature set; The profile fusion module 30 is used to perform cross-domain feature fusion and profile mining on the fresh food transaction feature set and the member interaction feature set to obtain multi-dimensional member profile features and real-time transaction context features. The intelligent deduction module 40 is used to synchronously input the multi-dimensional member profile features and the real-time transaction context features into the multi-task collaborative network for bidirectional instruction deduction, and generate real-time cashier control instructions and asynchronous marketing push instructions. The collaborative delivery module 50 is used to send the real-time cashier control command to the cashier terminal and the asynchronous marketing push command to the member terminal, so that the cashier terminal and the member terminal can perform corresponding cross-module collaborative actions.

[0098] The integrated intelligent control device for fruit and fresh produce checkout provided in this application, employing the integrated intelligent control method for fruit and fresh produce checkout in the above embodiments, can solve the technical problem of how to achieve full-link automated collaboration between checkout and membership management in fresh produce retail scenarios and reduce the reliance on manual intervention in the operational process. Compared with the prior art, the beneficial effects of the integrated intelligent control device for fruit and fresh produce checkout provided in this application are the same as those of the integrated intelligent control method for fruit and fresh produce checkout provided in the above embodiments, and other technical features in the integrated intelligent control device for fruit and fresh produce checkout are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0099] This application provides an integrated intelligent control device for fruit and fresh produce checkout, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the integrated intelligent control method for fruit and fresh produce checkout in the first embodiment described above.

[0100] The following is for reference. Figure 6The diagram illustrates a structural schematic suitable for implementing the integrated intelligent control device for fruit and fresh produce checkout in the embodiments of this application. The integrated intelligent control device for fruit and fresh produce checkout in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Android Devices), PMPs (Portable Media Players), and vehicle terminals (e.g., vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 6 The integrated smart control device for fruit and fresh produce checkout shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments in this application.

[0101] like Figure 6 As shown, the integrated intelligent control device for fruit and fresh produce checkout may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to the program stored in ROM 1002 (Read Only Memory) or the program loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the dual-model inversion oil depot fire simulation scenario construction device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the integrated smart POS system for fresh produce to exchange data with other devices wirelessly or via wired communication. Although the figure shows an integrated smart POS system for fresh produce with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0102] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0103] The integrated intelligent control device for fruit and fresh produce checkout provided in this application, employing the integrated intelligent control method for fruit and fresh produce checkout in the above embodiments, can solve the technical problem of how to achieve full-link automated collaboration between checkout and membership management in fresh produce retail scenarios and reduce the reliance on manual intervention in the operational process. Compared with the prior art, the beneficial effects of the integrated intelligent control device for fruit and fresh produce checkout provided in this application are the same as those of the integrated intelligent control method for fruit and fresh produce checkout provided in the above embodiments, and other technical features in this integrated intelligent control device for fruit and fresh produce checkout are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0104] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the integrated intelligent control method for fruit and fresh produce cashier in the above embodiments.

[0105] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0106] The aforementioned computer-readable storage medium may be included in the integrated intelligent control device for fruit and fresh produce checkout; or it may exist independently and not be assembled into the integrated intelligent control device for fruit and fresh produce checkout.

[0107] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the integrated intelligent control device for fruit and fresh produce checkout, the integrated intelligent control device for fruit and fresh produce checkout performs the following actions: acquires fresh produce transaction logs containing visual recognition information uploaded by the checkout terminal, and member interaction logs uploaded by the member terminal; performs multi-source data cleaning and structured alignment on the fresh produce transaction logs and the member interaction logs to obtain a fresh produce transaction feature set and a member interaction feature set; performs cross-domain feature fusion and profile mining on the fresh produce transaction feature set and the member interaction feature set to obtain multi-dimensional member profile features and real-time transaction context features; synchronously inputs the multi-dimensional member profile features and the real-time transaction context features into a multi-task collaborative network for bidirectional instruction deduction to generate real-time checkout control instructions and asynchronous marketing push instructions; and sends the real-time checkout control instructions to the checkout terminal and the asynchronous marketing push instructions to the member terminal, so that the checkout terminal and the member terminal perform corresponding cross-module collaborative actions.

[0108] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0109] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0110] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0111] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the above-described integrated intelligent control method for fruit and fresh produce checkout. This solves the technical problem of how to achieve fully automated collaboration between checkout and membership management in fresh produce retail scenarios and reduce reliance on manual intervention in the operational process. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the integrated intelligent control method for fruit and fresh produce checkout provided in the above embodiments, and will not be repeated here.

[0112] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described integrated intelligent control method for fruit and fresh produce cashier.

[0113] The computer program product provided in this application solves the technical problem of how to achieve full-link automated collaboration between checkout and membership management in fresh food retail scenarios and reduce the reliance on manual intervention in the operational process. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the integrated intelligent control method for fruit and fresh food checkout provided in the above embodiments, and will not be repeated here.

[0114] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A smart control method for integrated cashier system for fresh fruit, characterized in that, The method includes: Obtain fresh food transaction logs containing visual recognition information uploaded by the POS terminal, and member interaction logs uploaded by the member terminal; The fresh food transaction logs and the member interaction logs are subjected to multi-source data cleaning and structure alignment to obtain the fresh food transaction feature set and the member interaction feature set. Cross-domain feature fusion and profile mining are performed on the fresh food transaction feature set and the member interaction feature set to obtain multi-dimensional member profile features and real-time transaction context features. The multi-dimensional member profile features and the real-time transaction context features are synchronously input into a multi-task collaborative network for bidirectional instruction deduction, generating real-time cashier control instructions and asynchronous marketing push instructions. The real-time cashier control command is sent to the cashier terminal, and the asynchronous marketing push command is sent to the member terminal, so that the cashier terminal and the member terminal can perform corresponding cross-module collaborative actions.

2. The method as described in claim 1, characterized in that, The steps of performing cross-domain feature fusion and profile mining on the fresh food transaction feature set and the member interaction feature set to obtain multi-dimensional member profile features and real-time transaction context features include: Update the member consumption profile in the preset database based on the fresh food category data, weight data, and consumption frequency data in the fresh food transaction feature set; Based on the member consumption profile and the member interaction feature set, behavioral correlation analysis is performed to obtain a multi-tag set; The multi-tag set is clustered and integrated according to the preset number of tag categories to obtain multi-dimensional member profile features; The real-time interaction sequence in the member interaction feature set is concatenated with the current order transaction feature in the fresh food transaction feature set to obtain the real-time transaction context feature.

3. The method as described in claim 2, characterized in that, The step of performing behavioral correlation analysis based on the member consumption profile and the member interaction feature set to obtain a multi-tag set includes: Extract historical purchase frequency and average order value features from the member consumption profile, and extract page dwell time and activity click conversion features from the member interaction feature set; The historical purchase frequency feature, the average order value feature, the page dwell time feature, and the activity click conversion feature are aligned to construct a multimodal user feature sequence; The multimodal user feature sequence is input into a hidden Markov model for behavior sequence decoding to obtain behavior-related feature vectors; The behavior-related feature vectors are mapped to a preset preference label space for threshold filtering to obtain a multi-label set.

4. The method as described in claim 1, characterized in that, The multi-task collaborative network includes a shared feature extraction layer, an intent recognition branch, and a behavior prediction branch; The step of synchronously inputting the multi-dimensional member profile features and the real-time transaction context features into a multi-task collaborative network for bidirectional instruction deduction to generate real-time cashier control instructions and asynchronous marketing push instructions includes: The multi-dimensional member profile features and the real-time transaction context features are input into the shared feature extraction layer for global feature mapping to obtain global shared semantic features. The intent recognition branch performs intent mapping on the globally shared semantic features to obtain a real-time cashier preference vector representing the probability distribution of the strategy; By performing temporal inference on the globally shared semantic features through the behavior prediction branch, a long-term marketing preference matrix representing multi-dimensional marketing attributes is obtained; Based on the real-time cashier preference vector, discount rules are dynamically sorted in a preset instruction rule base to generate real-time cashier control instructions. Based on the long-term marketing preference matrix, recommended product identifiers are extracted from the preset instruction rule base and encapsulated to generate asynchronous marketing push instructions.

5. The method as described in claim 4, characterized in that, The intent recognition branch includes a multi-head attention mechanism layer, a fully connected network layer, and a normalized activation function; The step of mapping the globally shared semantic features through the intent recognition branch to obtain a real-time cashier preference vector representing the strategy probability distribution includes: The globally shared semantic features are input into the multi-head attention mechanism layer for weight allocation to obtain the intent-focusing features; The intent-focusing features are input into the fully connected network layer for dimensionality reduction to obtain a low-dimensional representation of the intent features. The low-dimensional feature representation of the intent is probabilistically normalized by the normalized activation function to obtain a real-time cashier preference vector that represents the probability distribution of the strategy.

6. The method as described in claim 4, characterized in that, The behavior prediction branch includes a residual network layer, a gated recurrent unit network layer, and a tensor reconstruction layer; The step of inputting the multi-dimensional member profile features into the behavior prediction branch of the multi-task collaborative network for time-series inference to obtain the long-term marketing preference matrix includes: The multi-dimensional member profile features are divided into a subset of static attribute features and a subset of dynamic preference features; The subset of static attribute features is input into the residual network layer for nonlinear semantic mapping to obtain static attribute features; The dynamic preference feature subset is input into the gated recurrent unit network layer to calculate the temporal dependency relationship, thus obtaining the dynamic temporal features; The static attribute features and the dynamic temporal features are added element by element to obtain the hybrid prediction features; The hybrid prediction features are spatially expanded and reconstructed by the tensor reconstruction layer to obtain a long-term marketing preference matrix in which the row dimension represents candidate fresh food categories and the column dimension represents multiple marketing attributes.

7. The method according to any one of claims 1 to 6, characterized in that, The steps of sending the real-time cashier control command to the cashier terminal and the asynchronous marketing push command to the member terminal, so that the cashier terminal and the member terminal can perform corresponding cross-module collaborative actions, include: Obtain the hardware driver layer protocol parameters of the real-time cashier control command, and convert the asynchronous marketing push command into an application layer network message; The real-time cashier control command, containing the hardware driver layer protocol parameters, is sent to the cashier terminal through a preset transport layer security encryption protocol channel, so that the cashier terminal can perform pricing and discount deduction operations. The application layer network message is sent to the member terminal through a preset program push interface, so that the member terminal can display the preferred product discount information corresponding to the asynchronous marketing push instruction.

8. A smart control device for integrated cashier system for fresh fruit, characterized in that, The device includes: The data acquisition module is used to acquire fresh food transaction logs containing visual recognition information uploaded by the POS terminal, and member interaction logs uploaded by the member terminal; The data preprocessing module is used to perform multi-source data cleaning and structure alignment on the fresh food transaction logs and the member interaction logs to obtain fresh food transaction feature sets and member interaction feature sets. The profile fusion module is used to perform cross-domain feature fusion and profile mining on the fresh food transaction feature set and the member interaction feature set to obtain multi-dimensional member profile features and real-time transaction context features. The intelligent deduction module is used to synchronously input the multi-dimensional member profile features and the real-time transaction context features into the multi-task collaborative network for bidirectional instruction deduction, generating real-time cashier control instructions and asynchronous marketing push instructions; The collaborative delivery module is used to send the real-time cashier control command to the cashier terminal and the asynchronous marketing push command to the member terminal, so that the cashier terminal and the member terminal can perform corresponding cross-module collaborative actions.

9. A smart control device for integrated cashier system for fresh fruit, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the integrated intelligent control method for fruit and fresh produce cashier as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the integrated intelligent control method for fruit and fresh produce cashier as described in any one of claims 1 to 7.

Images