A set of recommendation and dynamic incentive subsystems for a retail advertising placement and billing system based on physical behavior triggers and real-time bidding

Abstract

The application discloses a set of joint recommendation and dynamic incentive subsystem and method, integrated in the retail advertising delivery charging system of claim 1 of the original application (application number 2026109017636). The subsystem multiplexes shelves, carrier perception channels and adds trial perception branches; the scene adaptation engine locally caches multiple industry templates, supports offline full-service operation; the set bidding unit sets up competitive product shunt attenuation and transaction backtracking correction logic; the dynamic settlement unit is built-in exposure / diversion / transaction three-layer independent charging and refund reverse settlement module, matched with double-channel code scanning, weighing and verification, multi-level account distribution and secondary sales mechanism; the edge removes original audio and video, and desensitizes logs through SHA-256 hash, Merkle tree and tenant HMAC key to complete alliance chain isolation audit; the cloud end configures cross-domain clearing unit, and the edge carries behavior risk control unit. The scheme realizes automatic identification of goods to refund full-link, achieves advertisement budget control, store revenue increase and data compliance multiple effects.

Description

Technical Field

[0001] This invention relates to the fields of smart offline retail, edge IoT multimodal sensing, real-time programmatic advertising bidding, conversion effect-based differentiated billing, and automatic revenue sharing for offline stores. In particular, it relates to a packaged recommendation and dynamic incentive subsystem and method for a retail advertising placement and billing system based on physical behavior triggering and real-time bidding. Background Technology

[0002] Traditional offline brick-and-mortar retail advertising mainly relies on static posters, fixed electronic price tags, and large screen carousels above the checkout counter. This presents several critical technical shortcomings that urgently need to be addressed: a severe disconnect between ad triggering and customers' actual purchasing intentions, making it impossible to accurately push relevant ads when customers exhibit a strong intention to "pick up the product"; a crude billing model, generally using a uniform CPM deduction, making it impossible to quantify and guarantee advertiser ROI; a lack of data trust, with no tamper-proof third-party evidence storage mechanism; significant privacy and compliance risks, as the continuous uploading of raw audio and video to the cloud violates personal information protection laws; and the inability to achieve multi-intent signal fusion and inventory linkage. While existing technologies include fragmented solutions such as shelf weighing, visual recognition, online bidding, and blockchain evidence storage, no solution has yet integrated "shelf + checkout dual-mode synchronous time window multi-sensor cross-verification," "near-expiry / low inventory hard constraint bidding," "display-add-to-purchase three-level differential billing," and "edge raw data destruction + tenant isolation evidence storage" into a unified system. Furthermore, although the applicant's original application (application number 2026109017636) disclosed a basic system for retail advertising placement and billing based on physical behavior triggering and real-time bidding, the original application did not solve the following technical problems: how to realize cross-business bundled recommendations, how to provide brands with independent SKU budget pool management, how to realize differentiated incentives and automatic revenue sharing for store staff / robots, how to trigger a secondary sales loop after a transaction, how to handle reverse settlement of refunds and cross-store fund clearing, and how to solve the problem of assist attribution and three-tier billing in cross-regional roaming scenarios. This invention is an improved solution proposed to address the aforementioned technical problems that the original application failed to solve. Summary of the Invention

[0003] This invention aims to address the shortcomings of existing technologies, such as high false triggers in trial recognition, incompatibility across multiple business formats, uncontrollable advertising budgets, lack of automatic settlement of sales incentives, absence of follow-up promotions after sales, missing reverse settlement for refunds, gaps in cross-store clearing, gaps in anti-fraud risk control, and gaps in billing for cross-regional roaming attribution and price comparison scenarios. It provides a packaged recommendation and dynamic incentive subsystem and method for retail advertising placement and billing based on physical behavior triggering and real-time bidding. Relying on edge computing and local closed-loop processing, it achieves accurate hard judgment of trial behavior, deployment of a single device across multiple business formats, differentiated pricing through human-machine collaborative recommendations, controllable consumption of brand advertising budgets, fully automatic revenue sharing among multiple parties, follow-up promotions after sales to increase average order value, closed-loop reverse settlement for refunds, cross-domain fund clearing, and end-to-end risk control and privacy compliance. To achieve the above objectives, the present invention adopts the complete technical solution of claim 1. The system includes an edge computing device, a cloud management platform, multi-level sensing channels, multiple types of human-computer interaction terminals, and a customer display terminal. The edge device has ten core functional units built-in, including NTP clock synchronization, offline caching, dual-mode trial verification, human-computer recommendation arbitration, multi-scenario bidding, dynamic tiered settlement, automatic revenue sharing, secondary sales, reverse settlement, privacy desensitization, and behavioral risk control, forming a complete closed loop for offline retail revenue growth.

[0004] The core beneficial effects of this invention are: 1. Dual-mode synchronous time window multi-source verification mechanism: Integrating three signals from shelf vision, trial location sensing, and shelf weighing, the false touch recognition rate of traditional single-camera is reduced to less than 2%, accurately distinguishing between random touches and genuine trial with strong consumer intent; compared to the original application that only relied on shelf single-mode recognition, a new trial auxiliary mode is added to cover try-on / trial / tasting scenarios, improving the recognition coverage rate by about 60%; 2. Multiple standardized retail format templates are cached locally: Supports complete operation of stores offline. One set of universal hardware is compatible with large supermarkets / warehouse membership stores, clothing, 3C digital products, trendy toys, electrical appliances, pharmacies, convenience stores, KTV beverage shops, which greatly reduces the transformation and deployment costs of chain brands for multiple stores compared with the original single-format universal model. 3. Human-machine collaborative recommendation priority arbitration + differentiated pricing: Store staff recommend higher prices than robots and allow for independent purchases, while also incentivizing sales staff to actively promote products; when multiple recommendation sources are triggered simultaneously, the primary and secondary recommendation subjects are automatically assigned to eliminate conflicts between multiple terminals; compared to the original application without differentiated billing based on recommendation source, sales staff performance can be quantified. 4. SKU-specific ad budget pre-verification: Automatically lower bids when budget is exhausted and periodically check recharge status, preventing uncontrolled and ineffective spending on brand advertising from the source; compared to applications without budget control mechanisms, advertisers' ROI increases by approximately 30%-40%; 5. Closed-loop post-sale promotion: Based on real-time order amount and remaining brand advertising budget, push low-priced, high-margin small items, confirm additional purchases through weighing verification, significantly increasing the average order value and bundled sales rate in stores; this function is a completely new module that was not disclosed in the original application; 6. After the transaction is completed, revenue is automatically distributed across multiple levels based on a pre-configured dual-rule table in the cloud: advertising fees are directly allocated from the brand's budget pool to stores, sales associates, and robot operators, eliminating the labor costs and accounting errors associated with manual reconciliation; compared to applications without automatic revenue distribution capabilities, the number of disputed work orders can be reduced by approximately 95%; 7. Three-layer independent billing loop (exposure / traffic redirection / transaction): For cross-regional roaming and price comparison scenarios, it is the first to create a three-layer independent billing model with a display layer, a competitor intent traffic redirection layer, and an actual transaction conversion layer. Each layer deducts from the corresponding SKU's independent Budget_pool, with physical isolation of funds, completely solving the attribution and billing disputes of "A exposure B transaction". 8. Competitor-driven traffic attenuation and transaction retrospective correction: In scenarios with multiple competing product categories, the attenuation coefficient γ is automatically applied to balance exposure opportunities, and the pCVR estimated weight is dynamically adjusted based on the transaction results within the session to achieve dynamic traffic balance and maximize system ROI. 9. Closed-loop reverse settlement and cross-domain clearing: Supports Budget_pool fund rollback and three-level reverse accounting in refund scenarios, supports the chain headquarters to periodically aggregate and allocate funds from multiple stores, and fills the gap in the original application for after-sales cash flow; 10. All original images and audio streams are destroyed locally at the edge: Only desensitized behavior logs without biometric features are uploaded to the cloud, fully meeting the compliance requirements of relevant domestic personal information protection laws and regulations; the level of privacy protection remains consistent with the original application.

[0005] Exemplary parameter description: All values ​​in this specification (ΔT range 200~500ms, weight error ±8%, matching threshold 0.85, various bidding weighting factors, etc.) are the optimal engineering range measured in multiple offline scenarios; those skilled in the art can make ±20% adjustments to the parameters according to store lighting, sensor accuracy, and environmental noise, and the technical solution after the adjustment still falls within the protection scope of this invention.

[0006] Distinction from the original application: The applicant's original application (application number 2026109017636) only disclosed a basic system for retail advertising placement and billing based on physical behavior triggering and real-time bidding, protecting the multimodal hardware acquisition architecture including shelf vision, weighing, and microphones, and the advertising placement and billing logic based on this architecture. This invention does not improve the hardware structure; all innovations are concentrated on the upper-level business decision-making closed loop after hardware data acquisition: that is, based on the parent advertising placement system, it adds a multi-business scenario adaptation engine, bundled recommendation logic, brand-specific independent budget pool pre-verification, differentiated incentives for store staff / robots, a post-sale secondary sales closed loop, a refund reverse settlement module, a cross-domain clearing unit, a behavioral risk control unit, a three-layer independent billing model, and competitor diversion attenuation and transaction retrospective correction logic. The two patent applications have completely distinct protected subjects, belonging to different inventions under the same overall inventive concept, and there is no risk of duplicate authorization.

[0007] Supporting Explanation for this Sub-case: All new features in this sub-case—including the configuration of the trial experience area perception sub-channel, ΔT dual-mode decision logic, dual-channel barcode scanning and input via store clerk handheld POS terminals and shopping guide robot terminals, multi-format package combination rules for large-scale integrated supermarkets / warehouse membership stores / trendy toys / appliances, SKU independent budget pool pre-verification, adding products as independent items to the transaction session in the secondary promotion closed loop, reverse settlement of refunds, cross-domain clearing, behavioral risk control, three-layer independent billing (exposure / traffic redirection / transaction), competitor diversion attenuation coefficient γ, and transaction retrospective correction logic—are all specific implementation schemes under the original publicly disclosed framework of "physical behavior triggering → multimodal perception → real-time bidding → billing" in the parent case, and do not introduce hardware structures not recorded in the parent case. The trial experience area perception sub-channel is a regionalized configuration extension of the parent case's shelf area perception channel in the trial scenario. The activation signals output by pressure sensing devices such as shoe benches / clothing chairs / tasting tables can be used as one of the auxiliary verification conditions for shelf mode ΔT decision, and do not constitute a completely new hardware architecture. The specific technical features of this divisional application relative to the parent application are: NTP synchronous multimodal perception to determine physical behavior within a unified time window + bidding score formula to drive billing. This divisional application narrows its scope from "advertising placement and billing" to "bundle-related recommendations + differentiated incentives for store staff / robots + multi-level revenue sharing from brand budget pools + after-sales reverse settlement + three-layer attribution for cross-regional roaming", which meets the unity requirement of divisional applications under Article 31 of the Patent Law. Attached Figure Description

[0008] Figure 1System overall architecture diagram. The diagram shows: Edge computing device 10, shelf area sensing channel 20 (including shelf wide-angle camera 21, shelf load-bearing sensor 23, shelf microphone matrix 24), shopping vehicle sensing channel 30 (including vehicle weight sensor 22, positioning tag 32), trial experience area sensing sub-channel 40 (including shoe-fitting stool / clothing-fitting chair / tasting station pressure sensor 41, infrared proximity sensor 42), recommendation output terminal 50 (including third-party advertising display area 51), human-computer interaction terminal 60 (including shopping guide robot terminal 61, staff handheld POS terminal 62), and cloud management platform 70 (including cross-domain clearing unit 71).

[0009] Figure 2 : Flowchart of the dual-mode verification logic triggered by physical trial. The diagram is labeled: Shelf high-precision mode branch (S201→S202→S203), trial auxiliary mode branch (S204→S205→S203), and ΔT synchronization time window decision node.

[0010] Figure 3 Flowchart of Human-Machine Collaborative Recommendation Recognition and Priority Arbitration. The diagram is labeled: Staff Recommendation Subunit (S301→S302), Robot Recommendation Subunit (S303→S304), Priority Arbitration Module (S305), and Recommendation Source Identifier Generation.

[0011] Figure 4 Flowchart for multi-business scenario adaptation, package bidding, and budget verification. The flowchart shows: Scenario adaptation engine loading template (S401), Package combination library query (S402), Inventory hard constraint correction (S403), Bidding score calculation (S404), Budget pool pre-verification and polling (S405), Application of competitor diversion attenuation coefficient γ (S406), Transaction backtracking correction (S407).

[0012] Figure 5 Flowchart for dynamic tiered billing, multi-level automatic revenue sharing auditing, and reverse billing. The flowchart is labeled as follows: Recommendation source identifier matching (S501), application of differentiated unit prices (S502), tiered weight locking (S503), loading of R_table / B_table (S504), multi-level revenue sharing execution (S505), reverse billing triggering (S506), Budget_pool fund rollback (S507), and three-tier independent billing execution (S508).

[0013] Figure 6 : Complete logic flowchart of the closed loop of secondary promotion after a transaction. The diagram is marked as follows: Transaction confirmation trigger (S601), additional SKU budget verification (S602), POS / robot scanning and entry (S603), dual verification of weighing unit (S604), generation of independent settlement sub-item for additional products (S605). Detailed Implementation

[0014] Example 1: Anta Apparel Store (High-Precision Shelf Mode + Robot Scanning + Display Terminal Interaction + Secondary Sales): The store is fully equipped with shoe-fitting benches and fitting chairs with pressure sensors at the fitting areas (perception sub-channels in the trial experience area). Interactive electronic price tags and large-screen displays are fixed next to the shoe-fitting benches as customer-facing display terminals. The scene engine loads the apparel store template and caches it locally. When a customer sits down on the running shoe shelf holding SKU:ANT-001 trail running shoes, the system visually recognizes the complete trial action, the shoe-fitting bench outputs a standardized activation signal, and the shelf load-bearing detection measures the weight error of the removed item (≤8%). ΔT=350ms determines the first physical retrieval event (high-precision shelf mode). At the same time window, a store clerk wearing a smart badge proactively recommends matching trousers (SKU:ANT-002), and a shopping robot patrols nearby to simultaneously push out matching outfit options. The priority arbitration module determines that the store clerk is the primary recommendation source. Within 3 seconds, the store clerk approaches to interact with the customer, and the system maintains the priority display of the store clerk's recommendation at the unit price, and synchronizes the recommended content to the customer-facing display terminal. Product Placement and Settlement Interaction: 1. Display Terminal Guidance: Refresh "Shoes + Pants Set 8.5% Off"; 2. Robot Scanning and Input: Store clerk or robot scans the barcodes of running shoes and pants, SKU information is synchronized to the edge computing device; 3. Weighing Verification Confirmation: After the edge device verifies the consistency, the instruction is to place the product into a tray-type weighing carrier. If the weight increment error is ≤±8%, it is confirmed as valid, and the discount coefficient of 0.90 for 2 items is locked; Settlement Confirmation: The customer confirms payment on the POS terminal or the display terminal touch screen. Secondary Promotion Closed Loop: After settlement, the display terminal pushes an additional recommendation "Sports Socks Single Pair 29 Yuan"; after verifying that the Budget_pool balance of the sports socks is sufficient, the store clerk or robot scans the barcode of the sports socks, and the edge device instructs to place the socks into the tray. If the weighing verification is successful, the additional product is added as an independent settlement sub-item and included in the transaction session payable amount without modifying the original completed order; the main product Budget_pool and the additional product Budget_pool are separated, and the additional product generates an independent billing log entry. Revenue Sharing and Privacy: After a transaction, the revenue sharing module automatically deducts advertising fees from the pants brand's independent budget pool and settles store commissions, employee performance bonuses, and robot operator rewards proportionally; edge devices simultaneously destroy all original videos and audio recordings from the fitting area, do not collect or store customer facial or voiceprint biometric information, and only upload desensitized SKUs, times, and recommendation source logs to the cloud; billing logs are uploaded to the consortium blockchain via SHA-256 hash chain and Merkle root hash, and tenant HMAC is used for isolated auditing.

[0015] Example 2: Trendy Toy Blind Box Collection Store (Trial Assistance Mode + Robot Scanning Dominance + Limited Quantity Weighting): The store is equipped with a trial opening table and trial position sensors, loading local templates for trendy toy blind boxes; when a customer picks up a hidden blind box, shakes it, and examines it, the visual system recognizes the "reaching out → picking up → examining" action sequence, the trial position sensor outputs an activation signal, and the first physical picking event is determined within ΔT (trial assistance mode); the shopping robot stays next to the customer for 12 seconds, uses the onboard scanning module to scan the blind box barcode, and determines that the robot's recommendation is valid. The hidden SKU is marked as limited stock, DOS=2<3, skipping the out-of-stock removal rule, the bidding score is multiplied by a limited quantity factor of 1.3, and the electronic price tag continuously displays "Only 3 pieces left"; the customer takes three blind boxes and puts them into the weighing shopping basket, and the settlement weight is locked at 0.85; after the transaction, a cartoon pendant is pushed for additional promotion, after verifying the budget for the additional product, the robot scans the pendant barcode, weighs and confirms, and the additional product is added as an independent settlement sub-item to update the order amount, automatically completing the three-way revenue sharing among the brand, store, and robot operator.

[0016] Example 3: Large appliance store (high-precision shelf mode + robot scanning + trade-in): TVs cannot be removed from the shelf; only infrared trial position sensors are deployed. When a customer lingers and stares at the TV for an extended period, the infrared trial position sensor is activated, generating a trial trigger event. When a customer mentions the keyword "rice cooker," the microphone array identifies the structured audio event, increasing the bidding confidence by 1.15 times. An additional 1.15 replacement factor is added to the rice cooker product bidding. The relevant bidding results are displayed on the system's terminal. Store staff use a handheld POS terminal or a shopping guide robot to scan the product barcode to confirm, increasing the terminal weight ω_t=1.25 and improving exposure priority. After a sale, additional recommendations for power strips and cleaning kits are pushed. The robot scans the additional product barcodes, and the scale is reset to verify the weight of the additional products, completing the secondary sales loop.

[0017] Example 4: A 3C digital specialty store (mandatory accessory bundle + robot scanning) loads a local 3C digital template. The bundle rule is main unit + matching accessories. When a customer picks up the main unit, the visual recognition of the user's attempt to use the device activates the infrared sensor at the display stand, determining the pick-up event. The system automatically retrieves three types of accessories and adds them to the bidding queue, prioritizing the push of bundled options through the shopping guide robot. After the customer confirms, the robot scans the barcodes of the main unit and accessories, adds the three items to the weighing shopping cart, and applies a tiered discount of 0.85. After checkout, a Bluetooth headset recommendation is pushed. After verifying the budget for the additional item, the robot scans the barcode of the additional item, weighs and verifies it, and updates the settlement amount as an independent settlement item.

[0018] Example 5: KTV Beverage Shop (Weighing Cart + Robot Scanning + Delivery): A weighing cart is deployed as the shopping vehicle (exemplarily a wheeled weighing cart). A KTV beverage shop template is loaded, and the beer shelf layer is bound to a downstream snack mapping table for fruit platters, peanuts, and duck necks. When a customer reaches for a beer, visual motion recognition and shelf load detection detect a 6% weight loss error in the beer, triggering a decision. The microphone recognizes the customer's keyword "take three bottles of beer," and audio confidence weighting increases the bid. The robot pushes a beer + fruit platter 20% off set. The robot scans the barcodes of the beer and fruit platter using a scanning module. The clerk puts the two items into the weighing cart. After the weight verification passes, the discount coefficient of 0.90 for the two items is locked. After the customer completes one-click checkout, the system reads the order amount and pushes additional recommendations for peanuts and hangover soup. After verifying the budget of the additional items, the robot scans the barcode of the additional items. After the weighing cart verification is confirmed, the additional items are added as independent settlement sub-items into the transaction session. After the billing is completed, the system generates a delivery task and delivers the entire set of items in the weighing cart to the corresponding private room.

[0019] Example 6: A large supermarket / warehouse membership store (Bread section triggers ad A, final purchase of milk B). The store deploys weighing shopping carts as shopping vehicles, loads a large supermarket template, and configures a competitor attenuation coefficient γ=0.8 (effective when there are 3 or more competing SKUs in the bread category). When a customer pushes their cart to the bread shelf, visual recognition indicates the customer has picked up brand A toast bread. The shelf's high-precision mode triggers the first physical retrieval event (visual outline removal + shelf load ΔW≤±8%). The set bidding allocation unit retrieves the bread-matching beverage set combination library. Multiple brands, including milk A, milk B, and yogurt C, participate in the bidding. Milk A wins due to its higher Bid and pCVR. The interactive electronic price tag on the shelf fully displays the "A toast + A milk" set advertisement, and the customer's 3-second dwell time meets the effective exposure condition. Three-tier independent billing execution: S1, Supporting Display Layer: The system deducts 0.5 yuan from the exposure-level advertising fee of Milk A's independent Budget_pool, generates an independent SHA-256 hash log, and marks the billing level as exposure, triggering product as bread, and advertising SKU as Milk A; S2, Competitor Intent Traffic-Driven Layer: If a customer does not take Milk A and walks to the dairy product shelf, the visual system continuously captures Milk B for 12 seconds, generating a low-weight competitor attention intent. The system deducts a small traffic-driving fee of 0.2 yuan from Milk B's independent Budget_pool (of which 30%, or 0.06 yuan, is returned to the beverage category's public advertising pool); S3, Actual Transaction Conversion Layer: If a customer takes Milk B and puts it into the weighing shopping cart, and the weight increase is ≤±8% and passes the verification, the POS scan completes the transaction. The system deducts a transaction conversion fee of 3.0 yuan from Milk B's Budget_pool, which is allocated to the merchant and the sales guide / bot operator according to a preset ratio. Reverse Settlement and Retrospective Correction: If a customer subsequently requests a refund for Milk B, the reverse settlement submodule only rolls back the transaction level of 3.0 yuan; the already consumed exposure fee of 0.5 yuan and traffic referral fee of 0.2 yuan are not revoked (the traffic has actually been reached). Because A received exposure and B made a transaction in this session, the transaction retrospective correction logic takes effect, and the estimated conversion rate of Milk B's pCVR will increase by 15% in the next bread category bidding, improving its exposure competitiveness. All exposure, traffic referral, and transaction de-identified logs are constructed into a Merkle tree on the blockchain by type. Brands A and B can retrieve the advertising consumption details layer by layer through their respective HMAC keys, and the accounts are completely traceable independently.

[0020] Example 7: Chain Convenience Stores & Community Pharmacies (Basic Judgment Mode + Accelerated Turnover of Near-Expiry Products): Convenience stores lack dedicated hardware for trial use, relying solely on refrigerated shelf sensors and enabling a high-precision shelf mode; near-expiry yogurt with T_rem / T_total=0.10≤0.15 is forcibly placed at the top, automatically paired with a potato chip set recommendation; store clerks or robots scan product barcodes, place two items in a weighing basket, apply a discount coefficient of 0.90, and upon completion, push a follow-up promotion of bottled water. Pharmacies load a dedicated template, tightening the threshold for near-expiry drugs to 0.10; when a customer takes antihypertensive medication, a robot scans the drug barcode, and the visual confirmation unit matches the drug's outer packaging features to the 0.85 threshold, passing the test; the privacy compliance unit immediately destroys the original images captured in the drug area, without collecting or storing customer facial or voiceprint biometric information, and the desensitized log contains no customer biometric information; the system automatically pushes matching vitamin supplementary medication sets, supporting follow-up promotions of cotton swabs and pill boxes after the transaction.

[0021] Example 8: Fully Offline Store Operation Example: When the store's external network is interrupted, the edge device locally and persistently caches all 8 sets of business format templates, the advertising budget balance of each SKU, the merchant revenue sharing rules of R_table, and the brand revenue sharing rules of B_table; dual-mode trial identification, multi-source recommendation arbitration, package bidding, tiered settlement, secondary promotion, and revenue sharing log generation are all executed locally in a closed loop, with no business interruption; after the network is restored, SHA-256 hash encrypted and desensitized logs are uploaded to the cloud in batches, completing the end-to-end tamper-proof audit archiving.

[0022] Example 9: Budget Pool Exhaustion Polling and Independent Settlement for Additional Products Example: A brand's advertising budget pool balance for SKU:SNACK-001 drops to a critical value after multiple consecutive bids. When the Budget_pool (SNACK-001) balance is insufficient to cover the expected reward amount for the next round, the bundled bidding allocation unit automatically marks the SKU as "budget exhausted" and initiates a timed polling refresh mechanism, querying the recharge status from the cloud management platform every preset interval (exemplarily 60 seconds). If the brand completes the recharge within the polling period, the system automatically removes the "budget exhausted" mark and restores the SKU's bidding ranking. In the secondary promotion scenario, the Budget_pool balance of the additional product SKU:DRINK-002 is independent of the budget pool of the main product SKU:SNACK-001. Before triggering the additional recommendation, the secondary promotion closed-loop submodule first checks whether the Budget_pool balance of DRINK-002 itself is sufficient. If the balance is insufficient, the additional product recommendation is skipped, and the next candidate SKU in the additional recommendation pool is checked. After the additional goods are weighed and verified, they are included in the transaction session payable amount as an independent settlement sub-item, generating an independent billing log entry. The billing records of the original completed order are not modified, ensuring that the billing of the main transaction and the additional transaction can be traced independently.

[0023] Example 10: Reverse Refund Settlement and Cross-Store Clearing Closed Loop: Within 30 minutes of a customer making a purchase at an Anta apparel store, a refund request is made for an additional item, sports socks. The POS interface at the cashier receives the refund request and synchronizes it to the edge computing device. The reverse settlement submodule verifies that the Budget_pool frozen amount corresponding to the sports socks SKU is sufficient and transfers funds in reverse according to the original revenue sharing path (merchant 70% / salesperson 20% / robot 10%). A negative billing log entry (amount: -29 yuan) is generated and marked as "refund". The log is then processed via a SHA-256 hash chain and a Merkle root hash chain to ensure that the audit is tamper-proof. This store belongs to the Beijing region of a chain brand. The cross-domain clearing unit of the cloud management platform aggregates the billing logs of all stores in the Beijing region every morning. According to the "headquarters takes a 5% cut" rule in R_table, funds are collected from the Beijing region to the brand headquarters account, and the remaining 95% is retained in the store account, completing the three-level clearing of headquarters-region-store.

[0024] Example 11: Behavioral Risk Control and Hardware Failure Degradation: If a customer repeatedly takes 10 items of clothing in the fitting room (more than 5 times / 10 minutes), the behavioral risk control unit freezes the customer's session ID's set recommendation permissions and marks it as "suspicious"; the permissions are restored after the store staff manually unfreezes them. When the weighing device sensor malfunctions (reading mutation rate exceeds ±50% / second), the physical verification trigger unit automatically degrades to a pure visual judgment mode, relying solely on the visual confirmation unit to match the product anchor point, ensuring uninterrupted business operations.

[0025] The above are merely preferred embodiments of the present invention and are not intended to limit the technical solutions of the present invention. Any equivalent hardware replacement, reasonable parameter fine-tuning, simple process modification, or business scenario expansion within the protection scope of the core concept of the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A bundled recommendation and dynamic incentive subsystem integrated into a retail advertising placement and billing system, characterized in that, Applied to the retail advertising placement and billing system described in claim 1 of the original application (application number 2026109017636); The subsystem is activated by loading a multi-business scenario adaptation engine onto a retail location edge computing device, including: the edge computing device is equipped with an NTP clock synchronization module, which uniformly calibrates the system clock across all channels via the NTP protocol; the edge computing device is also equipped with a behavior risk control unit, which counts the frequency of physical retrieval events within the same session ID, and freezes the package recommendation permission for that session and marks it as suspicious behavior when the frequency exceeds a preset threshold; it verifies the abrupt change rate of the weighing unit reading, and determines it as abnormal interference and triggers an alarm when the rate exceeds a preset threshold; the edge computing device reuses the shelf area sensing channel and shopping vehicle sensing channel of claim 1 of the prior application, and adds a trial experience area sensing sub-channel: including at least one of a shoe-fitting stool / clothing-fitting chair pressure sensing device and an infrared proximity sensor, outputting a trial position activation signal containing an activation timestamp and a device ID to assist ΔT∈[200ms, Within 500ms, a first physical retrieval event is determined; the physical verification triggering unit has two parallel determination branches that can be triggered independently, namely the high-precision shelf mode and the trial assistance mode. A first physical retrieval event is generated when either mode meets the conditions; High-precision shelf mode: When the visual acquisition device recognizes the outline of the product moving out of the shelf display surface, and the shelf load-bearing sensor detects a weight reduction ΔW that satisfies |ΔW−m_item| / m_item ≤ 8%, a first physical retrieval event is determined; Trial assistance mode: When the visual acquisition device recognizes the user's "reaching out → picking up → examining / trying out" action sequence towards the product, and the trial experience area sensing sub-channel outputs a trial position activation signal, a first physical retrieval event is determined; The edge computing device is equipped with a scenario adaptation engine, which has multiple sets of retail scenario parameter templates built-in. Configurations are loaded and cached locally on the edge computing device based on store type. In offline conditions, dual-mode verification, bundled bidding, tiered settlement, revenue sharing auditing, secondary sales, and reverse settlement are all executed in a closed loop locally on the edge side, without relying on the cloud. The templates define differentiated bundled combination rules and inventory constraint thresholds. The template set includes at least: convenience store templates, large supermarket templates, clothing specialty store templates, pharmacy templates, 3C digital templates, trendy toy blind box store templates, electronics store templates, and KTV convenience store shelf templates. A human-machine collaborative recommendation and recognition unit includes a staff recommendation subunit, a robot recommendation subunit, and a recommendation priority arbitration module. The staff recommendation subunit's judgment logic is: matching the employee badge microphone with a pre-recorded brand recommendation word list, and detecting wrist displacement vectors using the employee badge motion sensor. The system determines a valid recommendation from a salesperson if the customer's movement within a 10cm-60cm range, the shopping cart showing no corresponding weight increase within the same ΔT, and all three conditions are met simultaneously. The robot recommendation sub-unit's logic is as follows: the robot's onboard vision detects that the customer's dwell time in front of a product exceeds a 10s threshold, and the robot's voice announcement, screen advertisement, and the current product SKU are all matched. The recommendation priority arbitration module prioritizes recommendations based on the following: when both salesperson and robot recommendations are triggered simultaneously within the same ΔT, the salesperson recommendation becomes the primary source, with the robot serving only as an auxiliary verification. If there is no salesperson interaction within 3 seconds of a triggered recommendation, the system automatically switches to the robot as the primary source. The bundle bidding allocation unit responds to the first physical retrieval event, queries the bundle combination library loaded by the scenario adaptation engine, and performs inventory hard constraint correction: the ratio of the remaining shelf life of the associated products, T_rem / T_total. ≤ 0.15 will be forcibly placed at the top; inventory available days (DOS) < 3 will be directly removed; and a comprehensive bidding score will be calculated: Score = Bid ​​× (pCVR)α × ωt × γ where: α ∈ [0.1, 2.0] is an adjustable weight coefficient; ω_t ∈ [0.5, 1.5] is the terminal weight factor of the recommended output terminal; Bid comes from the brand's pre-charged independent advertising budget pool Budget_pool(SKU) for the corresponding SKU; γ is the competitor diversion attenuation coefficient, which is applied when there are N or more competitor SKUs (N≥2) in the associated product pool of the triggered product category, γ ∈ [0.7, 0.9]; If the first physical take-up event is accompanied by a structured audio event, that is, the microphone matrix output is bound to the trigger keyword mark of the product and the DOA falls within the product display sector, then the comprehensive bidding score is multiplied by an enhancement value of 1.1~1.2; The package bidding allocation unit also has built-in transaction retrospective correction logic: if the first SKU that wins the bid does not sell within the same session ID, but the associated competitor SKU completes the transaction, then in the next bidding for the same trigger product, the pCVR estimated conversion rate weight of the sold competitor SKU is increased (exemplarily increased by 10%~20%).The dynamic incentive settlement unit marks the recommendation source identifier {staff recommendation / robot recommendation / self-pickup}, and applies differentiated display unit prices: staff recommendation > robot recommendation > self-pickup; the placed carriers are matched with tiered weights according to quantity: 1 item 0.95, 2 items 0.9, ≥3 items 0.85; the unit has a built-in reverse settlement sub-module: responding to the refund request from the cashier's sensing channel, it verifies the Budget_pool frozen amount corresponding to the refunded product SKU, and reverses the fund allocation path to the merchant's settlement account, staff's personal account, and robot operation account, and generates a negative billing log entry to be synchronously uploaded to the blockchain; the dynamic incentive settlement unit has a built-in hierarchical billing sub-module, which independently executes the following three-layer billing logic based on the recommendation source identifier and conversion result, with each layer's fee deducted from the corresponding SKU's independent Budget_pool, and funds physically isolated: matching display layer: when the recommended advertisement is displayed in front of the customer When the terminal meets the effective exposure conditions (e.g., a dwell time of ≥2 seconds and complete rendering without obstruction), the display-level advertising fee is deducted from the Budget_pool corresponding to the associated product SKU of the recommended advertisement; Competitor intent referral layer: When a low-weight gaze intent targeting a competitor is detected within the same session ID (e.g., visual recognition of a customer dwelling in front of a competitor's shelf for ≥10 seconds), a small referral fee is deducted from the Budget_pool corresponding to the competitor SKU of the same category as the recommended product, and the referral fee is returned to the corresponding category public advertising pool according to a preset ratio (e.g., 30%); Actual transaction conversion layer: When the product is placed in a vehicle and transacted via POS scanning, the transaction-level advertising fee is deducted from the Budget_pool corresponding to the transacted SKU; The dynamic incentive settlement unit has a built-in revenue sharing audit module and a secondary promotion closed-loop sub-module; The revenue sharing audit module loads the merchant rule table R_table. After a transaction is confirmed, the amount deducted from the Budget_pool in one go, and allocated to the merchant's settlement account, the store clerk's personal account, and the robot's operation account according to a preset ratio, is used in conjunction with the brand rule table B_table. The secondary promotion closed-loop sub-module triggers an additional recommendation event after the transaction is completed, based on the transaction amount and the budget balance of each additional SKU, and verifies the Budget_pool balance of the additional product itself. The package bidding allocation unit retrieves the additional recommendation pool to calculate the additional score (also applicable to structured audio enhancement value), and sends the additional recommendation instruction to the display terminal. After the customer confirms, the POS interface of the cashier's sensing channel receives the barcode of the additional product scanned by the store clerk's handheld POS terminal or the shopping guide robot terminal and synchronizes the SKU information to the edge computing device. The weighing unit detects that the weight increment of the vehicle is ≤±8%. The additional product is included in the transaction session payable amount as an independent settlement sub-item, without modifying the original completed order. The main product Budget_pool and the additional product Budget_pool are separated, and the additional product generates an independent billing log entry.Privacy Compliance Unit: Immediately destroys the original audio and video frames cached locally at the edge after generating the recommendation instruction. It does not collect or store customer facial or voiceprint biometric information, only anonymizing {SKU, timestamp, session ID, recommendation source} before uploading it to the cloud. Billing Data Storage Unit: Collects structured field groups for each billing event. These fields only include tenant ID, timestamp, product SKU hash, trigger type, billing level, and recommendation source identifier, without any customer biometric information. Each billing event is encapsulated as a SHA-256 hash chain entry. A Merkle tree is automatically constructed and a unique root hash is generated every 100 accumulated log entries or every 10 minutes. The root hash is synchronously uploaded to the cloud consortium blockchain for tamper-proof storage. Each tenant is assigned an HMAC signature field derived from an independent key. Any tenant can only access its own dedicated log subtree and integrity verification path using HMAC credentials generated by its own key. The cloud management platform is equipped with a cross-domain clearing unit, which periodically aggregates the anonymized billing logs reported by each edge node and executes the three-level fund clearing and transfer instructions of headquarters-region-store according to the tenant hierarchy rules in R_table / B_table.

2. The subsystem according to claim 1, characterized in that, The edge computing device is also equipped with a visual verification unit: Before generating a package recommendation instruction, the shelf visual acquisition device is used to photograph the product display surface, extract the product appearance color histogram, texture gradient feature vector and barcode / RFID tag feature vector, and match them with the locally stored visual template; if the matching degree is greater than a preset threshold of 0.85, the product anchor point is confirmed to be valid and a recommendation instruction is issued; if the matching degree is ≤0.85, a package recommendation instruction for that SKU is not generated.

3. The subsystem according to claim 1, characterized in that, The scenario adaptation engine has built-in differentiated rules specific to each business format: Large-scale supermarket template: Built-in tasting and trial slot association mapping table; pressure sensors or load sensors on tasting stations / baking islands / deli cabinets output trial slot activation signals to trigger trial assistance mode; bundle rules are mandatory binding of large package main products + matching consumables (laundry detergent + fabric softener, beef rolls + dipping sauce, baked bread + jam); near-expiry baked goods / deli / pre-made dishes are forced to be placed at the top if T_rem / T_total ≤ 0.15, and directly removed if DOS < 2; frozen glutinous rice ball shelf scenario is configured with a competitor diversion attenuation coefficient γ=0.

8. When there are 3 or more competitor SKUs in the glutinous rice ball category, the bidding score of the top brand is automatically reduced to reserve exposure opportunities for small and medium-sized brands; Template for trendy toy blind box stores: Add a whitelist of exceptions for limited editions; when the SKU is marked as limited edition / hidden edition / collaboration overstock and DOS < 3, skip the out-of-stock removal rule, implement the sense of urgency to the top, and multiply the bidding score by 1.3 limited edition factor; Appliance Store Template: Inventory constraints are adjusted to old models being phased out (DOS < 5), forcibly pinned to the top; SKUs participating in the trade-in program have their bidding scores multiplied by an additional 1.15 replacement factor; 3C Digital Template: The fixed rule for the package is main unit + matching accessories; after the shelf recognizes the main unit being picked up and triggers the event, the charger, protective case, and headphone accessories are forced to enter the bidding queue; KTV snack bar shelf template: Built-in beverage-snack association mapping table, the mapping table is configured by the brand cloud and bound to the shelf layer SKU association configuration parameters; after the shelf detects the beverage taking event, it automatically retrieves the downstream SKUs of fruit plate, peanuts, and duck neck bound on the same layer to participate in the bidding.

4. The subsystem according to claim 1, characterized in that, The human-computer interaction terminal includes a shopping guide robot terminal and a staff handheld POS terminal; the shopping guide robot terminal integrates a barcode scanning module, configured to scan product barcodes and synchronize SKU information to an edge computing device; the staff handheld POS terminal is configured to scan product barcodes and synchronize SKU information to an edge computing device; the customer-facing display terminal includes an interactive electronic price tag and a large-screen display fixed to the shelf / checkout counter, supporting touch and voice input confirmation of recommendation schemes.

5. A bundled recommendation and dynamic incentive method integrated into a retail advertising placement and billing system, applied to the subsystem described in any one of claims 1 to 4, characterized in that, It includes the following steps: S1. Reuse the sensing channel of claim 1 of the prior application, and within the synchronization time window of ΔT=200~500ms, select the shelf high-precision mode or the trial assistance mode to complete the trial behavior verification and generate the first physical retrieval trigger event. S2: The scene engine loads the corresponding business template and caches it locally to support offline operation. The bidding unit performs inventory constraint correction and pre-verification of the advertising budget balance of each SKU before generating a package recommendation instruction and sending it to each terminal. S3. In response to the scanning and input event of the store clerk's handheld POS terminal or shopping guide robot terminal, receive SKU information and verify its consistency with the recommended set, and instruct the weighing unit to perform weight increment verification on the goods placed in the carrier; if the weight increment error is ≤±8%, confirm that the placement is valid and update the settlement amount. S4. After an order is completed, multi-level automatic revenue sharing is executed; at the same time, a secondary promotion process is initiated, and after verifying the budget of the additional product, high-margin small items are pushed; in response to the POS or robot scanning and entry of the additional product, after the weighing verification is passed, the additional product is added as an independent settlement sub-item and incorporated into the transaction session without modifying the original completed order; S5. In response to the refund request from the cashier sensing channel, execute the reverse settlement process: verify the Budget_pool frozen amount corresponding to the refunded product SKU, transfer funds in reverse according to the original accounting path, and generate a negative billing log entry to be synchronously uploaded to the chain. S6. At the edge, all original audio and video are destroyed. No customer facial or voiceprint biometric information is collected or stored. Only desensitized behavior logs are uploaded to the cloud for auditing and archiving. Billing logs are synchronized to the consortium chain via SHA-256 hash chain and Merkle root hash, and tenant HMAC is isolated for auditing.

Images