An idle time period marketing intelligent decision system based on payment voucher uploading

Abstract

The application discloses an idle time period marketing intelligent decision system based on payment voucher uploading, and the system comprises a payment voucher processing module, an AI decision engine module, a dynamic free coupon generation module, a store activation management module and a self-evolution operation module. The payment voucher processing module receives the payment voucher image uploaded by a consumer, completes authenticity verification and extracts structured payment information. The AI decision engine module adopts a five-dimensional weighted scoring system including idle time urgency, demand prediction, free coupon strategy, store priority and risk assessment. After the system is simulated and verified by digital twinning, the strategy is put into place. Cross-brand privacy protection collaborative optimization is realized through federated learning. Decision explanation is generated through an interpretable AI module. The system does not rely on payment gateway application program interface docking, independent software vendor cooperation, special-purpose payment collection QR codes and pre-merchant registration, and realizes intelligent and automated guidance of offline store idle passenger flow and batch access of store networks.

Description

Technical Field

[0001] This invention relates to the field of off-peak store marketing strategy technology, specifically to an intelligent decision-making system for off-peak marketing based on payment voucher uploads. Background Technology

[0002] Existing offline store operation and marketing technology solutions generally have the following limitations:

[0003] Regarding channel dependence, food delivery platforms require stores to pay 15% to 25% of sales revenue as commission, while in-store group buying models force discounts of 20% to 30% accompanied by 60 to 90-day payment terms, severely compressing store profit margins. In terms of technical integration, connecting to payment gateway companies' systems requires individual application programming interface (API) integration, a lengthy process with the risk of rejection. Relying on independent software vendors necessitates signing contracts, paying transaction fees, and replacing QR codes with dedicated payment methods, resulting in high transformation costs for merchants. Regarding store expansion, franchise stores require individual negotiation, signing agreements, and training, leading to a slow and linear expansion pace. In terms of marketing… In terms of strategy, existing systems mostly adopt static strategies with fixed amounts, fixed time periods, and fixed validity periods. They lack specific designs for guiding customer flow during off-peak hours and the ability to dynamically adjust based on real-time data. Using a uniform approach for different stores and consumers fails to achieve differentiated operations. Regarding data utilization, a large number of historical payment vouchers held by consumers are discarded, and transaction data generated during store operations is not fully accumulated into data assets that can be used for business decision-making. In terms of continuous optimization, existing systems lack the ability to self-learn, self-optimize, and self-evolve strategies, making it difficult for marketing effectiveness to improve over time.

[0004] In addition, existing solutions generally separate brand headquarters from franchise stores, lacking a unified service activation channel, and brand headquarters cannot quickly incorporate all its stores into the same marketing system through a one-time decision. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent marketing decision-making system based on payment voucher upload during off-peak hours, in order to solve the problems mentioned in the background art.

[0006] A smart marketing decision-making system for off-peak hours based on payment voucher uploads includes a payment voucher processing module, which is used to acquire payment voucher images uploaded by consumer terminals, complete voucher authenticity verification, and extract structured payment information containing merchant name, payment amount, and payment time from the images;

[0007] The AI ​​decision engine module is used to access data on store status, time attributes, consumer characteristics, operational history, and external environment, and generate marketing decision results for off-peak hours through a multi-dimensional weighted scoring system.

[0008] The dynamic free coupon generation module is used to generate free coupons whose face value, usage period, and validity period can be dynamically adjusted based on the marketing decision results.

[0009] The store activation management module is used to provide at least two store activation paths;

[0010] The self-evolving operations module is used to collect free coupon redemption data and store operation performance data, feed them back to the AI ​​decision engine module to iteratively optimize decision-making strategies, and automatically generate operational digital assets.

[0011] Furthermore, the payment voucher processing module acquires payment voucher images including real-time payment vouchers and historical payment vouchers. This module performs forgery detection on the images using a generative adversarial network-based algorithm and outputs a forgery probability value. ,when It was determined to be counterfeit at the time, among which The preset threshold for determining forgery;

[0012] The system employs at least one of optical character recognition, visual artificial intelligence, and multimodal artificial intelligence to extract structured payment information from payment voucher images. The structured payment information includes merchant name, payment amount, payment time, and consumption details.

[0013] Furthermore, the multi-dimensional weighted scoring system of the AI ​​decision engine module includes an off-peak urgency scoring submodule, a demand forecasting scoring submodule, a free coupon strategy scoring submodule, a store priority scoring submodule, and a risk assessment scoring submodule. The off-peak urgency scoring submodule normalizes seat occupancy rate, weather impact index, traffic impact index, historical sales figures for the same period, holiday coefficient, store basic score, and current time-of-day customer traffic to the [0,1] interval, denoted as... , ,..., Based on the dynamic weight vector =[w1,w2,...,w7] Calculate the idle time urgency score using the following formula: ; The demand forecasting and scoring submodule normalizes the historical sales volume, historical customer traffic, weather-sales correlation coefficient, holiday-sales correlation coefficient, historical free coupon effectiveness index, and business district dynamic index for the same period, and denotes them as y1, y2, ..., y6. The demand forecasting score is then calculated using the following formula: ;

[0014] The lower the value, the greater the predicted demand gap, and the higher the marketing priority should be. The free coupon strategy scoring submodule will , The following factors were used as inputs: consumer price sensitivity (c1), store historical conversion rate (c2), remaining budget percentage (c3), and business district competition index (c4). The free coupon strategy score was calculated using the following formula: ; in =[ , [c1,c2,c3,c4] For the corresponding weights; the store priority scoring submodule normalizes the off-peak urgency level u1, store historical activity response rate u2, store geographical location value score u3, brand strategy consistency u4, and store long-term growth potential value u5, and then calculates the store priority score using the following formula: ; The risk assessment scoring submodule normalizes the voucher forgery detection score r1, the number of historical anomalies in stores and consumers r2, the deviation of free coupon amount r3, and the abnormal usage pattern detection value r4, and then calculates the risk assessment score using the following formula: ; when When the risk threshold τ is exceeded, the system will automatically reduce the amount of free coupons or suspend their distribution.

[0015] Furthermore, the dynamic free coupon generation module generates the free coupon in at least one of the following ways: face value Based on the payment amount in the payment information Multiply by dynamic scaling factor Determined, that is ,in Calculate using the following formula: ; As the benchmark scaling factor, For consumers' price sensitivity, The expected conversion rate for the time period. The remaining percentage of the budget, to For coefficients; Use time periods through time period decision functions Output, in For real-time passenger flow data, function g calculates the initial off-peak time period based on the real-time passenger flow. , Adaptive adjustment is performed using the following formula: ; Adjust the step size for the time period; The validity period d is dynamically set according to the consumer return probability p_return, the average historical response period μ of the consumer, and the strategy objective parameter λ, using the following formula: ;

[0016] in Based on the number of valid days, This is the attenuation coefficient.

[0017] Furthermore, the store activation management module includes a first activation path and a second activation path;

[0018] Under the first activation path, after the brand headquarters completes the signing, it makes a one-time collective decision. The system marks all stores under the brand as service targets in batches within a preset time. The marked stores do not need to change the payment QR code, modify the existing system, or conduct additional training.

[0019] Under the second activation path, when the store corresponding to the payment voucher uploaded by the consumer is not registered in the system, the system automatically identifies the store information and sends a registration invitation to the store operator. The store operator completes the registration with a single confirmation operation, and all consumer incentive requests initiated before registration are processed normally as temporary requests.

[0020] Furthermore, it also includes a digital twin simulation module, used to retrieve historical operational data of target stores before the official launch of marketing strategies, and construct a probabilistic model that includes the consumer arrival process, redemption probability distribution, and average transaction value variation patterns as a virtual digital twin model of the store; candidate strategies are run in this virtual model, and the customer traffic change rate is output. Free coupon redemption rate Average order value change rate and follow-up rate At least one of the indicators; weight the indicators according to the following formula to obtain the effect score. : ; in to This is the weighting coefficient, which can take positive or negative values; when When this happens, the candidate strategy is pushed to the actual store for execution; if If the conditions for deployment are met, the strategy parameters will be automatically adjusted and the simulation will be rerun until the maximum number of simulations is reached. .

[0021] Furthermore, it also includes a federated learning module for scenarios involving collaborative operations across multiple brands, where each brand's original operational data is stored independently on a local server without centralized transmission; Each brand trains a local model using its own data and updates the local model parameters using a stochastic gradient descent algorithm. Only the updated The encrypted data is uploaded to the central aggregator; the central aggregator uses a federated averaging algorithm to calculate the global model parameters using the following formula: ;

[0022] in For the first Local data volume for each brand , Total data for all brands To determine the number of participating brands; to obtain Then it is distributed to each brand's local machine to update the local model, so as to achieve cross-brand model collaborative optimization without the original data leaving the local machine.

[0023] Furthermore, it also includes an explainable artificial intelligence module, used to automatically generate decision explanation text for each marketing decision output by the AI ​​decision engine module by calculating the contribution value of each input feature to the decision result; for each feature... , Calculate using the following formula: ;

[0024] in For the set of all features, For features not included Feature subset, For use only The output value of the model when the feature is in the middle; Reflecting characteristics Based on the marginal contribution of the decision-making outcome, a decision explanation text is generated, which includes the ranking of core influencing factors, the contribution ratio of each score, and a comparative analysis with the historical strategies of the same store.

[0025] Furthermore, the self-evolving operation module incorporates a reinforcement learning unit, which will process the current state of the store. Defined as a state space, it contains optional marketing strategies. Defined as the action space, it represents the monitored increase in customer traffic after each strategy execution. Free coupon redemption rate Changes in follow-up visit rate and changes in average order value Calculate the reward value using the following formula: : ;in to The weights are coefficients, which can be positive or negative; the action value function is updated using the Q-learning algorithm as follows: ;

[0026] In the formula For learning rate, As a discount factor, To perform the action The state is then transitioned to the next state; and adopted - Greedy strategy selects actions based on probability. Randomly explore new strategies, using probability Select the current optimal strategy, and gradually reduce its effectiveness as the number of executions increases. Value is enhanced to continuously improve the accuracy of decision-making; the self-evolving operation module automatically generates operational digital assets including at least one of the following: regional consumption heat map, time period consumption feature map, brand competitiveness index, consumer migration path map, marketing strategy effect library, industry benchmark dataset, and AI-generated business suggestion report. All types of digital assets are automatically updated and archived, and support independent commercial authorization.

[0027] It also includes a real-time settlement module for handling fees, which supports at least one of three settlement methods: deduction from the brand headquarters' pre-deposited account, deduction from the virtual account of independent stores, and automatic deduction from bank accounts; when the dynamic free coupon generation module generates a free coupon, a settlement operation is triggered simultaneously, and the amount of the free coupon is deducted from the corresponding account. preset ratio Real-time deduction of handling fees : ; The free coupon will only be sent to the consumer's terminal after the deduction is successfully confirmed.

[0028] A smart decision-making method for off-peak marketing based on payment voucher upload, applied to an off-peak marketing smart decision-making system based on payment voucher upload, includes the following steps: Collect data on store status, time attributes, consumer characteristics, product information, external environment, historical operations, and brand policies; The collected data is input into a multi-dimensional weighted scoring model. After normalizing each variable, the data is weighted and summed using a weight vector to generate candidate marketing strategies for off-peak periods. The candidate marketing strategies are simulated in a virtual digital twin model of the store built based on historical data, and the effectiveness score is calculated. : when Then it enters the execution phase; Obtain the payment voucher image uploaded by the consumer, verify its authenticity using a generative adversarial network, and output a forgery probability value. and with threshold The comparison process extracts the merchant name, payment amount, and payment time, matches the corresponding store, and confirms the store's activation status. Based on the candidate marketing strategies and payment information, according to the formula Free coupons with dynamically adjustable face value are issued to consumers, among which... A dynamic proportional coefficient related to off-peak urgency score, consumer price sensitivity, expected conversion rate during time period, and remaining budget percentage; Collect data on free coupon redemption and store operation performance, and apply the formula. Calculate reward values, use Q-learning algorithms to optimize the parameters of the decision-making model, and automatically generate operational digital assets; By using the brand headquarters to activate the process in bulk and consumer credentials to trigger the automatic discovery process for independent stores, the system enables the bulk access and continuous expansion of offline stores. The method does not rely on API integration with payment gateway companies, cooperation with independent software vendors, dedicated QR codes for receiving payments, or prior merchant registration.

[0029] By adopting the above technical solution, it can operate solely based on the payment voucher image uploaded by the consumer terminal. It does not require application programming interface (API) integration with payment gateway companies, nor does it rely on cooperation with independent software vendors. Stores do not need to change their dedicated QR codes for receiving payments, nor do they need to complete merchant registration in advance.

[0030] Compared with the prior art, the beneficial effects of the present invention are: this intelligent marketing decision-making system based on payment voucher upload during off-peak hours,

[0031] Through a five-dimensional weighted scoring system that considers off-peak urgency, demand forecasting, free coupon strategy, store priority, and risk assessment, the system dynamically generates the face value, usage period, and validity period of free coupons based on multi-dimensional variables such as real-time seat occupancy rate, weather, traffic, consumer characteristics, and historical data. For the same store, the coupon amount is automatically increased on Fridays and decreased and the validity period is extended on rainy days, enabling the strategy to adapt to changes in conditions.

[0032] Free coupons with limited usage periods are generated specifically for off-peak hours when store seating is less occupied. These coupons are limited to use from 2 PM to 5 PM, guiding consumers away from peak hours and improving the store's operational efficiency throughout the day.

[0033] It provides a batch activation path for brand headquarters. After the brand headquarters completes the signing, it makes a one-time collective decision. The system will mark all its stores as service targets in batches within a preset time. Stores do not need to sign contracts separately, change QR codes, or modify the system. At the same time, it provides a path for consumers to trigger independent stores to automatically discover them by uploading payment vouchers for stores that have not been registered. Consumers can trigger store identification and registration invitations by uploading payment vouchers for stores that have not been registered. Store operators can complete the registration with a single confirmation operation, realizing the rapid expansion of the store network.

[0034] By using reinforcement learning units, the reward value is calculated by weighting the increase in customer traffic, the redemption rate of free coupons, the change in repeat visit rate, and the change in average order value after each strategy execution. The action value function is continuously updated to dynamically balance exploration and utilization. From the initial exploration stage where customer traffic increases by about 10%, it gradually converges to the optimization stage where it steadily increases by about 25%. The effect of the strategy increases with the number of executions, without the need for manual intervention.

[0035] By using a digital twin simulation module, a virtual model is built using historical store operation data before the strategy is officially launched. This model simulates the change rate of customer traffic, redemption rate, change rate of average order value, and repeat visit rate of candidate strategies. The effect score is calculated by weighting the results, and the strategy is only pushed out after reaching a preset threshold. The optimal solution that balances gross profit and customer traffic is selected. The deviation between the actual launch results and the simulation prediction is within an acceptable range, reducing the cost of trial and error. Attached Figure Description

[0036] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0037] Figure 2 This is a flowchart of the multi-level scoring calculation process for the AI ​​decision engine of this invention;

[0038] Figure 3 This is a flowchart illustrating the digital twin simulation decision-making process of the present invention.

[0039] Figure 4 This is a flowchart illustrating the collaborative workflow of federated learning and explainable AI in this invention.

[0040] Figure 5 This is a data flow diagram for Mode A (Headquarters Batch Activation Mode) of the present invention;

[0041] Figure 6 The reinforcement learning loop graph is continuously optimized for this invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Please see Figure 1-6 The present invention provides a technical solution: a marketing intelligent decision-making system based on payment voucher upload during off-peak hours, including a payment voucher processing module, used to acquire payment voucher images uploaded by consumer terminals, complete voucher authenticity verification, and extract structured payment information containing merchant name, payment amount, and payment time from the images;

[0044] The AI ​​decision engine module is used to access data on store status, time attributes, consumer characteristics, operational history, and external environment, and generate marketing decision results for off-peak hours through a multi-dimensional weighted scoring system.

[0045] The dynamic free coupon generation module is used to generate free coupons whose face value, usage period, and validity period can be dynamically adjusted based on the marketing decision results.

[0046] The store activation management module is used to provide at least two store activation paths;

[0047] The self-evolving operations module is used to collect free coupon redemption data and store operation performance data, feed them back to the AI ​​decision engine module to iteratively optimize decision-making strategies, and automatically generate operational digital assets.

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1: Complete Operation Flow of Multi-Level Scoring Calculation for AI Decision Engine

[0050] A certain fast food chain has more than 300 stores nationwide. After the lunch peak, the number of customers drops significantly from 2 pm to 5 pm. The brand headquarters hopes to increase the number of customers visiting the store during this period by a certain percentage through the system, but does not want to rely on high discounts in the long term.

[0051] After the system was connected, it first read the operating records of each store during the target period over the past thirty days. Data from a store located in an office building business district showed that the store's average seat occupancy rate was about 30% during the afternoon hours on weekdays. During the same period, the average occupancy rate of the same brand's stores in the same business district was close to 60%. The store's daily sales during the afternoon hours were about 20% of those during the lunch peak hours. Sales data from the same period in previous years also showed that this month was a slow season.

[0052] The external data collected by the system that day included cloudy weather, a temperature of 28 degrees Celsius, moderate real-time traffic conditions around the store, and no holidays. This data, along with the store's basic score and the current real-time customer flow, was input into the off-peak urgency score submodule. After normalization of each variable, the scores were weighted and summed according to preset weights. The store's off-peak urgency score was 0.82, which is at a relatively high level among all stores.

[0053] The demand forecasting and scoring submodule runs synchronously. Input variables include the store's sales and customer traffic in the same period in the past, the historical correlation coefficient between weather and sales, the historical correlation coefficient between holidays and sales, the redemption data after the recent free coupons were issued, and the recent dynamic index of the business district. After weighted calculation, the demand forecasting score is 0.78, indicating that the predicted demand for this period is insufficient. The system will automatically upgrade the marketing priority of the store.

[0054] The free coupon strategy scoring submodule takes the above two scores as input, combines them with the consumer's price sensitivity tag, the store's recent redemption rate, the current remaining percentage of the marketing budget, and the business district competition index stored in the system, and outputs a free coupon strategy score. The store priority scoring submodule and the risk assessment scoring submodule are calculated simultaneously, and give the store priority score and risk score respectively. The store's historical activity response rate is higher than the brand average, its geographical location is in an office building cluster area, and its risk score is within the safe range.

[0055] After the five scores are generated, the system outputs a comprehensive marketing decision for the store, issuing consumers a free coupon equivalent to 30% of the original payment amount, valid for one day and usable from 2 pm to 5 pm. The decision then enters the digital twin simulation stage for verification. The simulation predicts that customer traffic can be increased by about 40%, the average transaction value can be reduced by about 5 percentage points, and the incremental gross profit is higher than the coupon cost. After the verification is successful, the strategy is officially launched.

[0056] Data feedback after implementation showed that the store's afternoon customer traffic increased by approximately 45%, the average transaction value decreased by approximately 4%, and the redemption rate exceeded 70%. The system recorded the results in the store's strategy performance profile and used them as positive samples to feed back into the scoring model, resulting in minor adjustments to the relevant weighting coefficients. When other stores encounter similar combinations of conditions in the future, the system will refer to this experience to generate strategies.

[0057] Example 2: Specific Implementation of Digital Twin Simulation

[0058] A chain coffee brand has stores in multiple cities. The brand plans to distribute free coupons in the afternoon, but it is unclear what level of coupon amount will balance increasing the number of customers visiting the store and the profit per store. If the amount is too low, consumers may not be willing to visit the store during non-usual consumption hours. If it is too high, the increased gross profit may be offset by the cost of the coupons.

[0059] After the brand sets the target time period and budget limit in the system, the system retrieves the brand's operating data for the afternoon hours of each store over the past three months from the historical database. This includes the distribution of customer traffic every half hour, the average transaction value, the fluctuation range of customer traffic under different weather conditions, and the redemption ratio and redemption time distribution of previous free coupon activities. This data is used to build a digital twin model of the brand's afternoon hours.

[0060] The model fits three core relationships using historical data: the time distribution of consumer arrivals, the redemption probability of free coupons of different denominations, and the correspondence between the denomination of free coupons and changes in the actual average order value.

[0061] The system ran three candidate strategies in the model, setting the free coupon amount to 25%, 35%, and 45% of the payment amount. At 25%, the model predicted an increase in customer traffic of approximately 15%, a decrease in average transaction value of approximately 3%, and an increase in gross profit of approximately 12% after deducting coupon costs. At 35%, customer traffic increased by approximately 30%, the average transaction value decreased by approximately 7%, and gross profit increased by approximately 23%. At 45%, customer traffic increased by approximately 42%, the average transaction value decreased by approximately 15%, and gross profit increased by approximately 27%.

[0062] From a gross profit perspective, a 45% discount is preferable, but considering the impact of discounts on brand image, the brand opted for a 35% discount. After the actual campaign, customer traffic increased by about 32%, and the average order value decreased by about 8 percentage points. The deviation from the model's prediction was within an acceptable range. This actual data was subsequently incorporated into the historical database for calibration of subsequent simulations.

[0063] Example 3: Specific Implementation of Federated Learning and Privacy Protection

[0064] Three different chain restaurant companies have each connected to the system, with categories of rice noodles, noodles and casual meals respectively. The stores are located in different cities, but there is some overlap in consumption scenarios and customer structure. The amount of data of a single brand is limited, making it difficult to independently train a high-accuracy demand prediction model. However, none of them are willing to share their original operating data such as store sales amount and customer flow details.

[0065] The federated learning module is deployed as follows: Each brand deploys a local model training environment on its own server. The original data is always kept on its own server. Within a set period, each brand uses the newly added data to train its local model. After training is completed, the updated values ​​of the model parameters are extracted, encrypted, and uploaded to the central aggregation server.

[0066] After receiving the encrypted parameters from each brand, the aggregation server weights and sums the parameters according to the proportion of local training data for each brand to the total. Brands with larger data volumes contribute more to the global parameters, while brands with smaller data volumes contribute less. The aggregated global model parameters are then distributed back to each brand's local server to update the local model.

[0067] Throughout the entire process, the transaction records, customer flow details, and consumer consumption records of each brand never left the local server. The transmitted content was an encrypted model parameter vector, making it impossible to reverse-engineer the specific operating data of any store or the consumption records of any individual consumer.

[0068] After three months of operation, the global demand forecasting model based on federated learning has improved in accuracy compared to independent training by each brand. It is more stable in judging consumption trends across regions and categories. For example, the demand forecast of noodle brands in southern cities during the winter off-peak season can indirectly benefit from the pattern of rice noodle brands in the same period of previous years. The two have certain similarities in consumption scenarios and seasonal fluctuations, but the brands have never exchanged raw data.

[0069] Example 4: Specific Implementation of Explainable AI

[0070] When reviewing the weekly strategy execution report pushed by the system, the operations staff of a brand headquarters noticed that a store located in an old business district had been allocated a higher amount of free coupons than other stores in the same city for a consecutive week. The operations staff questioned whether there was any basis for continuing to allocate marketing resources to this store, given that the business district's consumer activity has been generally low in recent years.

[0071] When the operations staff clicked on the decision details for the store in the system backend, the system generated the following explanation: The store's average seat occupancy rate after lunch this week was about 35%, far lower than the brand's average of 65%, contributing about 50% to the decision; the demand forecast score showed that the store's afternoon customer traffic in the coming week is expected to be about 20% lower than the average for the same period, contributing about 30%; the risk score was low, and the store's free coupon redemption rate has been consistently in the top 20% of the brand over the past three months, with no abnormal records, contributing about 10%; other factors included the store being included in the brand's recent key activation list, contributing about 10%.

[0072] The explanation also compared the store's strategy effectiveness with the previous week. Last week, when the free coupon quota was lower, afternoon customer traffic increased by about 12%, and the redemption rate was about 65%. After a moderate increase in the quota this week, customer traffic is expected to increase by more than 20%, and the redemption rate is expected to exceed 70%. Historical records show that, under similar seat vacancy rates and demand gaps, the investment corresponding to the current quota can be recovered through incremental gross profit within about two weeks.

[0073] After reading the instructions, the operations staff understood the system's judgment criteria. The main problem with the store was the uneven distribution of customer traffic at different times. Using off-peak free coupons to encourage existing customers to adjust their visit times resulted in a marginal cost that was lower than acquiring new customers. The operations staff did not intervene in the strategy, and the system continued to execute the original plan.

[0074] Example 5: Specific Implementation of Continuous Optimization in Reinforcement Learning

[0075] In the early stages of a store's integration with the system, the system had not yet established an understanding of the store's consumer response patterns and did not directly adopt a fixed strategy, but instead began to try it out from a relatively conservative perspective.

[0076] For the first ten days or so, the system was mainly in the exploratory phase. The amount of free coupons was randomly selected between 15% and 30% of the payment amount, and different combinations of usage time and validity period were made simultaneously. After each combination was launched, the system recorded the changes in customer traffic, redemption time distribution, and customer revisit interval for that combination. During this phase, the effect of increasing customer traffic fluctuated greatly, averaging about 10%, with occasional combinations showing significantly better results.

[0077] Ten days later, the system had accumulated a certain number of samples and began to be utilized. The data showed that the store's consumers responded more actively to the free coupons with a value of about 25%. The redemption rate and the value were not a simple linear relationship. After exceeding 25%, the redemption rate slowed down while the coupon cost continued to increase. The system narrowed its exploration range to the 20% to 30% range and began to identify time patterns. It found that the redemption rate on Friday afternoon was significantly higher than on Monday, and the redemption rate on rainy days was about 10 percentage points lower than on sunny days.

[0078] A month later, the system entered a stable and optimized state. On weekday afternoons, the free coupon quota was stabilized at 25%, and on Fridays it was moderately increased to 30%. On rainy days, it was automatically reduced to 20% and the validity period was extended to offset the impact of reduced consumer willingness to go out. At this time, the store's afternoon customer traffic increased by about 25% compared to before the system was implemented, the average transaction value decreased by about 3 percentage points, and the redemption rate stabilized at over 70%.

[0079] Throughout the process, the system continuously runs Q-learning updates. After each strategy execution, the system calculates the reward value based on the weighted sum of changes in customer traffic, redemption rate, repeat visit rate, and average transaction value, and updates the value estimate of the corresponding state and action. The exploration ratio gradually decreased from about 10% in the initial stage to about 50%, and the strategy gradually converged. Store operators did not manually adjust parameters during the process; the entire process from the initial state to the stable strategy was completed autonomously by the system.

[0080] Example 6: Specific Implementation of Independent Settlement and Real-Time Automatic Collection of Free Coupon Fees

[0081] The face value of the free coupon is independent of the actual payment and settlement system. The system needs to collect a handling fee at the time of issuing each free coupon. It cannot rely on end-of-month reconciliation or post-event collection, otherwise it will create accounts receivable risk.

[0082] The system is configured with three fee collection channels for different store types.

[0083] For stores under signed chain brands, the brand headquarters opens a pre-deposit account and deposits a security deposit at the beginning of the cooperation. The security deposit amount is calculated by the system based on the brand's estimated monthly coupon issuance volume and average coupon amount. When the balance is lower than the preset level, the system sends a replenishment reminder to the headquarters finance department. After the consumer uploads the voucher and passes the review, the system generates a free coupon and deducts a handling fee from the brand headquarters' pre-deposit account according to the preset ratio of the coupon amount. After the deduction is successful, the free coupon is pushed to the consumer. If the pre-deposit account balance is insufficient, the free coupon will not be issued temporarily. The system sends a notification to both the brand and the consumer. At the end of the month, the system automatically generates a statement of account, listing the coupon issuance records and handling fee deduction details for both parties to verify.

[0084] After an independent store is automatically identified and registered by the system through customer-uploaded vouchers, the system opens a virtual account for it. Store operators can top up this account via bank transfer or third-party payment tools. The amount is determined by the store's own business volume, and the system provides a suggested top-up amount based on the store's average daily transaction volume. Each time a free coupon is generated, the system deducts a handling fee from the store's virtual account in real time. After successful deduction, the coupon is issued. If the balance is insufficient, the coupon issuance process is stopped, and the system sends a top-up reminder to the store. The store can view the balance and deduction details on the data dashboard.

[0085] Some stores have activated automatic bank deduction agreements. When registering or signing a contract, stores sign a deduction authorization agreement with the operator and bind their bank accounts. Each time a coupon is issued, the system initiates a deduction request through the bank interface. After receiving a successful deduction confirmation from the bank, the free coupon is issued. This method requires no pre-deposit funds and no manual operation.

[0086] The three methods can run in parallel. The system automatically matches the settlement channel according to the store registration type and contract status. Each transaction fee deduction generates a unique serial number, which supports real-time query and monthly reconciliation.

Claims

1. A smart decision-making system for off-peak marketing based on payment voucher uploads, characterized in that, include: The payment voucher processing module is used to acquire payment voucher images uploaded by consumer terminals, complete voucher authenticity verification, and extract structured payment information containing merchant name, payment amount, and payment time from the images; The AI ​​decision engine module is used to access data on store status, time attributes, consumer characteristics, operational history, and external environment, and generate marketing decision results for off-peak hours through a multi-dimensional weighted scoring system. The dynamic free coupon generation module is used to generate free coupons whose face value, usage period, and validity period can be dynamically adjusted based on the marketing decision results. The store activation management module is used to provide at least two store activation paths; The self-evolving operations module is used to collect free coupon redemption data and store operation performance data, feed them back to the AI ​​decision engine module to iteratively optimize decision-making strategies, and automatically generate operational digital assets.

2. The off-peak marketing intelligent decision-making system based on payment voucher upload as described in claim 1, characterized in that, The payment voucher processing module acquires payment voucher images including real-time payment vouchers and historical payment vouchers. This module uses a generative adversarial network-based algorithm to perform forgery detection on the images and outputs a forgery probability value. ,when It was determined to be counterfeit at the time, among which The preset threshold for determining forgery; The system employs at least one of optical character recognition, visual artificial intelligence, and multimodal artificial intelligence to extract structured payment information from payment voucher images. The structured payment information includes merchant name, payment amount, payment time, and consumption details.

3. The off-peak marketing intelligent decision-making system based on payment voucher upload as described in claim 1, characterized in that, The AI ​​decision engine module's multi-dimensional weighted scoring system includes a falltime urgency scoring submodule, a demand forecasting scoring submodule, a free coupon strategy scoring submodule, a store priority scoring submodule, and a risk assessment scoring submodule. The falltime urgency scoring submodule normalizes seat occupancy rate, weather impact index, traffic impact index, historical sales figures for the same period, holiday coefficient, basic store score, and current time-of-day customer traffic to the [0,1] interval, denoted as... , ,..., Based on the dynamic weight vector =[w1,w2,...,w7] Calculate the idle time urgency score using the following formula: ； The demand forecasting and scoring submodule normalizes the historical sales volume, historical customer traffic, weather-sales correlation coefficient, holiday-sales correlation coefficient, historical free coupon effectiveness index, and business district dynamic index for the same period, and denotes them as y1, y2, ..., y6. The demand forecasting score is then calculated using the following formula: ； The lower the value, the greater the predicted demand gap, and the higher the marketing priority should be. The free coupon strategy scoring submodule will , The following factors were used as inputs: consumer price sensitivity (c1), store historical conversion rate (c2), remaining budget percentage (c3), and business district competition index (c4). The free coupon strategy score was calculated using the following formula: ； in =[ , [c1,c2,c3,c4] For the corresponding weights; The store priority scoring submodule normalizes the off-peak urgency level u1, store historical activity response rate u2, store location value score u3, brand strategy consistency u4, and store long-term growth potential value u5, and calculates the store priority score using the following formula: ; The risk assessment scoring submodule normalizes the voucher forgery detection score r1, the number of historical anomalies in stores and consumers r2, the deviation of free coupon amount r3, and the abnormal usage pattern detection value r4, and then calculates the risk assessment score using the following formula: ； when When the risk threshold τ is exceeded, the system will automatically reduce the amount of free coupons or suspend their distribution.

4. A smart decision-making system for off-peak marketing based on payment voucher upload as described in claim 1 or 3, characterized in that, The dynamic free coupon generation module generates the free coupon in at least one of the following ways: Face value Based on the payment amount in the payment information Multiply by dynamic scaling factor Determined, that is ,in Calculate using the following formula: ； As the benchmark scaling factor, For consumers' price sensitivity, The expected conversion rate for the time period. The remaining percentage of the budget, to For coefficients; Use time periods through time period decision functions Output, where For real-time passenger flow data, function g calculates the initial off-peak time period based on the real-time passenger flow. , Adaptive adjustment is performed using the following formula: ； Adjust the step size for the time period; The validity period d is dynamically set according to the consumer return probability p_return, the average historical response period μ of the consumer, and the strategy objective parameter λ, using the following formula: ； in Based on the number of valid days, This is the attenuation coefficient.

5. The off-peak marketing intelligent decision-making system based on payment voucher upload according to claim 1, characterized in that, The store activation management module includes a first activation path and a second activation path; Under the first activation path, after the brand headquarters completes the signing, it makes a one-time collective decision. The system marks all stores under the brand as service targets in batches within a preset time. The marked stores do not need to change the payment QR code, modify the existing system, or conduct additional training. Under the second activation path, when the store corresponding to the payment voucher uploaded by the consumer is not registered in the system, the system automatically identifies the store information and sends a registration invitation to the store operator. The store operator completes the registration with a single confirmation operation, and all consumer incentive requests initiated before registration are processed normally as temporary requests.

6. The off-peak marketing intelligent decision-making system based on payment voucher upload according to claim 1, characterized in that, It also includes a digital twin simulation module, used to retrieve historical operational data of target stores before the official launch of marketing strategies, and construct a probabilistic model that includes the consumer arrival process, redemption probability distribution, and average transaction value variation patterns as a virtual digital twin model of the store; candidate strategies are run in this virtual model, and the customer traffic change rate is output. Free coupon redemption rate Average order value change rate and follow-up rate At least one of the indicators; weight the indicators according to the following formula to obtain the effect score. : ； in to This is the weighting coefficient, which can take positive or negative values; when At that time, the candidate strategy is pushed to the actual store for execution; like If the conditions for deployment are met, the strategy parameters will be automatically adjusted and the simulation will be rerun until the maximum number of simulations is reached. .

7. The off-peak marketing intelligent decision-making system based on payment voucher upload according to claim 1, characterized in that, It also includes a federated learning module for scenarios involving collaborative operations across multiple brands, where each brand's original operational data is stored independently on a local server without centralized transmission; Each brand trains a local model using its own data and updates the local model parameters using a stochastic gradient descent algorithm. Only the updated The encrypted data is uploaded to the central aggregator; the central aggregator uses a federated averaging algorithm to calculate the global model parameters using the following formula: ； in For the first Local data volume for each brand , Total data for all brands To determine the number of participating brands; to obtain Then it is distributed to each brand's local machine to update the local model, so as to achieve cross-brand model collaborative optimization without the original data leaving the local machine.

8. The off-peak marketing intelligent decision-making system based on payment voucher upload according to claim 1, characterized in that, It also includes an explainable artificial intelligence module, which automatically generates decision explanation text for each marketing decision output by the AI ​​decision engine module by calculating the contribution value of each input feature to the decision result; For features , Calculate using the following formula: ； in For the set of all features, For features not included Feature subset, For use only The output value of the model when the feature is in the middle; Reflecting characteristics Based on the marginal contribution of the decision-making outcome, a decision explanation text is generated, which includes the ranking of core influencing factors, the contribution ratio of each score, and a comparative analysis with the historical strategies of the same store.

9. The off-peak marketing intelligent decision-making system based on payment voucher upload according to claim 1, characterized in that, The self-evolving operation module has a built-in reinforcement learning unit, which will learn the current state of the store. Defined as a state space, it contains optional marketing strategies. Defined as the action space, it represents the monitored increase in customer traffic after each strategy execution. Free coupon redemption rate Changes in follow-up visit rate and changes in average order value Calculate the reward value using the following formula: : ; in to The weights are coefficients, which can be positive or negative; the action value function is updated using the Q-learning algorithm as follows: ； In the formula For learning rate, As a discount factor, To perform the action The next state after that; And adopt - Greedy strategy selects actions based on probability. Randomly explore new strategies, using probability Select the current optimal strategy, and gradually reduce its effectiveness as the number of executions increases. Values ​​should be enhanced to continuously improve the accuracy of decision-making; The self-evolving operation module automatically generates operational digital assets including at least one of the following: regional consumption heat map, time-period consumption characteristic map, brand competitiveness index, consumer migration path map, marketing strategy effect library, industry benchmark dataset, and AI-generated business suggestion report. All types of digital assets are automatically updated and archived, and support independent commercial authorization. It also includes a real-time settlement module for handling fees, which supports at least one of three settlement methods: deduction from the brand headquarters' pre-deposited account, deduction from the virtual account of independent stores, and automatic deduction from bank accounts; when the dynamic free coupon generation module generates a free coupon, a settlement operation is triggered simultaneously, and the amount of the free coupon is deducted from the corresponding account. preset ratio Real-time deduction of handling fees : ； The free coupon will only be sent to the consumer's terminal after the deduction is successfully confirmed.

10. The off-peak marketing intelligent decision-making method based on payment voucher upload according to claim 1, applied to an off-peak marketing intelligent decision-making system based on payment voucher upload, characterized in that, Includes the following steps: Collect data on store status, time attributes, consumer characteristics, product information, external environment, historical operations, and brand policies; input the collected data into a multi-dimensional weighted scoring model, normalize each variable, and then sum the weighted values ​​using a weighted vector to generate candidate marketing strategies for off-peak periods; The candidate marketing strategies are simulated in a virtual digital twin model of the store built based on historical data, and the effectiveness score is calculated. : ； when Then it enters the execution phase; Obtain the payment voucher image uploaded by the consumer, verify its authenticity using a generative adversarial network, and output a forgery probability value. and with threshold The system compares and extracts merchant name, payment amount, and payment time; matches corresponding stores and confirms store activation status; based on the candidate marketing strategies and payment information, it uses a formula... Free coupons with dynamically adjustable face value are issued to consumers, among which... A dynamic proportional coefficient related to off-peak urgency score, consumer price sensitivity, expected conversion rate during time period, and remaining budget percentage; Collect data on free coupon redemption and store operation performance, and apply the formula. Calculate reward values, use Q-learning algorithms to optimize the parameters of the decision-making model, and automatically generate operational digital assets; By using the brand headquarters' batch activation path and consumer credentials to trigger the automatic discovery path of independent stores, the system enables the batch access and continuous expansion of offline stores. The method does not rely on API integration with payment gateway companies, cooperation with independent software vendors, dedicated QR codes for receiving payments, or prior merchant registration.

Images