Service area intelligent marketing and consumption guiding system based on cross-scene data fusion

Abstract

The invention relates to a service area intelligent marketing and consumption guiding system based on cross-scene data fusion. The system comprises a semantic understanding module, a dynamic knowledge management module, an illusion suppression module, a security adaptation module, an intelligent retrieval interaction module and a reasoning and evaluation module. The semantic understanding module realizes accurate intention recognition through a civil aviation field exclusive semantic analysis network and a dialogue state tracking model; the dynamic knowledge management module constructs a layered knowledge base and realizes knowledge increment updating through a real-time synchronous interface; the illusion suppression module adopts a triple domain fact verification mechanism to filter error information; the security adaptation module guarantees data security through local private deployment and data desensitization; the intelligent retrieval interaction module realizes accurate recall based on a multi-dimensional feature aggregation algorithm; and the reasoning and evaluation module is embedded into an exclusive reasoning engine and monitors the effect through an exclusive index system. The invention aims to solve the problems that the intelligent security and protection system in the industrial factory area is difficult to adapt to multi-brand / multi-protocol equipment, algorithm training depends on mass data, the period is long, the operation efficiency of a single-equipment algorithm is low, and data leakage risks exist.

Description

Service Area Intelligent Marketing and Consumer Guidance System Based on Cross-Scenario Data Fusion Technical Field

[0001] This invention belongs to the field of computer information retrieval and natural language processing technology, and relates to a service area intelligent marketing and consumer guidance system based on cross-scenario data fusion. Background Technology

[0002] With the continuous improvement of the highway network and the sustained growth of motor vehicle ownership, service areas, as important supporting facilities of transportation hubs, have transformed from a single "rest and supply function" to a "diversified service complex," encompassing multiple business scenarios such as refueling, catering, retail, and tourism information inquiry. According to data from the Ministry of Transport, in 2024, highway service areas nationwide served an average of over 50 million vehicles per day, with users showing a significant increase in demand for personalized and real-time service. Against this backdrop, the core requirement for service area operation and management has shifted from "basic resource guarantee" to "precise service guided by user needs." Data, as a key carrier connecting users and services, directly determines the operational efficiency and user experience of service areas through its integration and utilization capabilities. This situation is driving the industry to explore technologies such as "cross-scenario data value mining" and "intelligent marketing implementation."

[0003] However, in the actual operation and service provision process, the service areas still face multi-dimensional technical bottlenecks, which seriously restrict the realization of precise marketing and consumer guidance. First, the data dimensions are narrow and cross-scenario correlations are weak. Existing data collection is mostly limited to single business formats, such as refueling records at gas stations and POS data at restaurants. There is a lack of integration of user behavior data across the entire "refueling-dining-shopping" chain. Furthermore, the data formats are heterogeneous across different scenarios, such as vehicle recognition data in image format and consumption data in structured tables, making it difficult to form a complete view of user behavior. Second, there is a serious disconnect between customer profiling and marketing execution. Most service areas only build shallow profiles through simple statistics, such as age distribution and spending range, which cannot extract deep consumer preferences, such as sensitivity to promotional activities and cross-business consumption correlations. Moreover, the profile results are only used for static analysis and are not linked to real-time marketing actions. Finally, there is a lack of real-time feedback and optimization mechanisms. Marketing plans are mostly manually preset, such as fixed holiday discounts, and cannot be dynamically adjusted according to real-time traffic flow and user interaction behavior. This results in insufficient marketing personalization, such as pushing refueling coupons to non-fuel users, slow response times, and failure to guide user flow in a timely manner during peak hours, ultimately leading to low operational efficiency and poor user experience.

[0004] To address these issues, the industry has conducted relevant technological research, resulting in two main approaches. One focuses on building customer profiles, such as the Chinese invention patent CN119398831A, titled "Method and System for Customer Profile Building in Service Areas Based on Deep Clustering." Its technical principle involves acquiring vehicle and basic consumption data from service areas, preprocessing and feature processing, extracting deep features using a Transformer model, and combining this with K-means clustering to classify customer groups and ultimately build customer profiles. The advantage of this method is that it improves the accuracy of customer classification through deep clustering, more accurately identifying different customer group characteristics compared to traditional statistical methods. However, its disadvantages include limited data sources (vehicle and basic consumption data), failing to cover multi-business and cross-scenario data, and only outputting static profiles without marketing execution and feedback mechanisms, thus failing to achieve a closed loop from profile to service. Another type focuses on multi-source data processing and industry applications, such as the Chinese invention patent CN114881678A, "A High-Precision Customer Profiling Method for Logistics Enterprises Based on Big Data Technology." Its technical principle involves collecting logistics user data through databases and multi-source heterogeneous web crawlers, preprocessing it, classifying it using density clustering algorithms, and then combining this with optimized convolutional neural networks to classify user packages, ultimately generating customer profiles. The advantage of this method lies in its broad data collection dimensions and its integration with business scenarios; however, its disadvantages include being designed specifically for logistics enterprise customers and not adapting to the multi-business (gasoline, catering, retail) characteristics of service areas, and it does not involve real-time marketing guidance and dynamic optimization, failing to meet the "high mobility and strong real-time" operational needs of service areas.

[0005] A comprehensive analysis of existing technologies reveals that while CN119398831A addresses the accuracy issue of customer segmentation in service areas, its data dimensions and application implementation capabilities are insufficient. While CN114881678A expands data sources, its scenario adaptability and real-time performance are lacking, failing to fundamentally solve the problems of insufficient personalized marketing, slow response, and low operational efficiency in service areas. Therefore, there is an urgent need to build an intelligent marketing and consumer guidance system that covers the entire process of "data collection-processing-analysis-execution-feedback" and is adaptable to multiple business scenarios in service areas. This system should achieve an end-to-end solution, from accurate user profiling to real-time marketing implementation, through deep cross-scenario data fusion and intelligent algorithm-driven approaches. Summary of the Invention

[0006] This invention provides a service area intelligent marketing and consumption guidance system based on cross-scenario data fusion, which solves the technical problems in the prior art, such as narrow service area data dimensions, weak cross-scenario correlation, disconnect between customer profiles and marketing execution and lack of real-time feedback optimization, resulting in insufficient marketing personalization, slow response and low operational efficiency.

[0007] To address the aforementioned issues, the technical solution adopted in this invention is: a service area intelligent marketing and consumption guidance system based on cross-scenario data fusion, comprising a data processing layer, an intelligent analysis layer, a marketing execution layer, and a user interaction layer. This system achieves personalized marketing and real-time consumption guidance through cross-scenario multi-source data fusion and intelligent algorithm-driven implementation. Specifically, it includes: a data acquisition layer: deploying IoT devices and system interfaces across multiple service area scenarios to collect user consumption data, user attribute data, and scenario environment data in real time, and sending them to the data processing layer via an encrypted transmission protocol; a data processing layer: using a distributed database to store multi-source data, and after data cleaning and standardization, achieving cross-scenario data association through a data fusion engine; an intelligent analysis layer: based on the fused dataset, constructing user profiles and extracting consumption preferences through machine learning models, predicting pedestrian traffic and consumption hotspots using time-series analysis models, and discovering cross-industry consumption patterns using association rule mining algorithms; a marketing execution layer: automatically generating personalized marketing plans based on the intelligent analysis results, pushing customized offers and real-time consumption guidance information through multiple channels, and automating the configuration and execution of promotional activities; and a user interaction layer: providing multi-terminal interaction entry points, recording user interaction behavior, and feeding it back to the intelligent analysis layer for model optimization.

[0008] The principle and advantages of this solution are as follows: The principle of this solution is to build a closed-loop system with a full chain of "data collection-processing-analysis-execution-feedback", with cross-scenario data fusion as the core support, and to achieve personalized marketing and real-time consumption guidance through layered collaboration. The data acquisition layer, relying on the multi-business scenarios of the service area, deploys IoT devices and system interfaces to comprehensively capture user consumption data, attribute data, and scene environment data. Encrypted transmission ensures secure data transfer to the data processing layer. The data processing layer uses a distributed database to store multi-source heterogeneous data. After cleaning and removing invalid information and standardizing the data format, a data fusion engine establishes cross-scene data associations, forming a complete dataset of user behavior and scene status. The intelligent analysis layer, based on the fused data, uses machine learning models to mine user consumption preferences, time-series analysis models to predict foot traffic and consumption hotspots, and association rule mining algorithms to discover cross-business consumption patterns, providing accurate basis for marketing decisions. The marketing execution layer automatically generates personalized plans based on the analysis results, pushing customized offers and real-time guidance information through multiple channels to automate promotional activities. The user interaction layer provides multi-terminal interaction entry points, synchronously recording user interaction behavior and feeding it back to the intelligent analysis layer to continuously optimize model performance and ensure the system adapts to dynamically changing service scenarios and user needs.

[0009] Compared to existing technologies, this solution comprehensively collects three types of data—consumption, attributes, and environment—through multi-scenario IoT devices. Then, a data fusion engine enables cross-industry data correlation, forming a complete view of the user's end-to-end behavior. Existing technologies can only acquire data from a single industry, such as collecting only refueling records or catering consumption data. This solution, however, can integrate and correlate user refueling behavior, catering consumption, and retail shopping data. For example, it can identify the user behavior pattern of "purchasing snacks within 30 minutes of refueling," providing a data foundation for precision marketing. Secondly, it breaks down the barrier between customer profiling and marketing execution. Existing technologies can only output static customer profiles for analysis, while this solution, through an intelligent analysis layer and marketing... The coordinated action at the sales execution layer directly translates user preferences into real-time marketing actions. For example, it automatically pushes restaurant discount coupons to users with "frequent refueling and fast food preferences." Compared to the generalized marketing of existing technologies, such as pushing uniform coupons to all users, the marketing accuracy is significantly improved. This solution uses behavioral data recorded at the user interaction layer to optimize the model in reverse. For example, it adjusts marketing plans based on user coupon redemption rates, enabling the system to continuously adapt to changes in user needs. Existing marketing plans are mostly manually preset and fixed, while this solution can achieve dynamic iteration of marketing strategies. For example, during peak hours, it can adjust the number of coupons issued in real time based on traffic flow to guide user flow, greatly improving operational efficiency.

[0010] Furthermore, the IoT devices in the data acquisition layer include smart fuel dispenser data acquisition modules at gas stations, smart POS machines in restaurants, QR code payment terminals in retail stores, user interaction kiosks at tourist information points, and vehicle recognition devices in parking lots. User consumption data includes consumption amount, product SKU code, payment channel, and consumption duration. User attribute data includes license plate number, vehicle type, member ID, historical consumption frequency, and consumption amount threshold. Scene environment data includes real-time pedestrian flow, queue length, regional temperature, and light intensity in each scene. By deploying targeted IoT devices in core business areas such as gas stations and restaurants, the system can cover the entire user experience, from vehicle recognition upon entering the service area to refueling, dining, shopping, and information access to tourist information. The entire inquiry process ensures that no key user touchpoint data is missed. Specifying concrete data items such as consumption amount, license plate number, and real-time pedestrian flow guarantees the standardization and targeting of data collection, avoiding redundant and invalid data. Furthermore, by complementing consumption data, user attribute data, and scenario environment data, a multi-dimensional and comprehensive view of users and scenarios can be constructed. For example, combining license plate number, fuel consumption amount, and restaurant queue length can accurately identify high-spending users traveling long distances and provide guidance on restaurants that do not require queuing. Compared to existing technologies that only collect single types of data, this approach more accurately captures user needs and scenario states, providing reliable support for subsequent data fusion, intelligent analysis, and personalized marketing plan generation, significantly improving the scientific nature and effectiveness of marketing decisions.

[0011] Furthermore, the data fusion engine of the data processing layer employs a weighted association algorithm to achieve cross-scenario data association, specifically including: S301: Assigning a unique identifier ID to each user, the unique identifier ID is generated through a license plate number hash operation or a member ID association, and the hash operation formula is: Unique Identifier ID = SHA-256. The service area code is a fixed 6-digit number, and the timestamp salt value is a millisecond-level timestamp of the data collection time. Used to map hash results to 12-bit unique identifiers; S302: Defines cross-scenario data association weights and user attribute data association weights. Consumer behavior similarity weight Time-related weights The formula for calculating the total correlation degree is: in, The user attribute matching score ranges from 0 to 1, with 1 representing a perfect match. For similarity of consumer product categories, The decay coefficient is the time interval between two consumptions. t is the time interval between two consumptions, and T is the decay threshold; S303: When the total correlation R ≥ 0.7, it is determined to be cross-scenario data of the same user and associated with the same unique identifier ID; when When this occurs, a secondary verification mechanism is triggered, supplementing verification by comparing the last four digits of the user's payment channel number; when When user data is identified as different from other user data, it is stored separately. A 12-bit unique identifier ID is generated using SHA-256 hashing combined with service area coding and millisecond-level timestamp salt. This ensures both the uniqueness and security of user identification while also accommodating the identity association needs of non-member and member users, solving the problem of unifying the identification of different user identity data. By setting three weights—user attributes, consumption behavior similarity, and time correlation—and constructing a formula for calculating the overall correlation, the correlation strength of data across different dimensions can be quantified. Furthermore, the time interval decay coefficient reflects the objective law that shorter intervals between consumption behaviors indicate stronger correlation, avoiding the bias caused by a single-dimensional correlation. This approach avoids misjudgments. By setting hierarchical judgment rules—direct association for R≥0.7, secondary verification for 0.4≤R<0.7, and separate storage for R<0.4—it can quickly and accurately associate high-confidence cross-scenario data, reduce the false association rate of medium-confidence data through secondary verification, and avoid interference from low-confidence data. Compared with existing technologies' simple feature matching or unweighted association methods, this significantly improves the accuracy and reliability of cross-scenario data association, ensuring that the fused data can truly reflect the user's full-link behavior. This provides high-quality data support for the subsequent intelligent analysis layer to extract consumer preferences and discover cross-industry patterns, thereby improving the accuracy of personalized marketing solutions.

[0012] Furthermore, the association rule mining algorithm of the intelligent analysis layer adopts the improved Apriori algorithm, introducing a time decay factor and scene weight coefficient to optimize frequent itemset mining, specifically including: S401: Define cross-business consumption itemsets ,in For consumption behavior in a single business format, for each consumption item Assign scene weights ,Gas station Restaurants retail stores Tourist Information Points S402: Introduce a time decay factor β when calculating itemset support. Where λ is the attenuation coefficient, by default. , The improved itemset support formula is: (The interval between the time of the consumption behavior and the current time is given.) in, The indicator function takes a value of 1 if the condition occurs and 0 otherwise, where N is the total number of users within the statistical period; S403: Set the minimum support threshold to 0.05 and the minimum confidence threshold to 0.3, and mine cross-business consumption association rules X→Y that meet the conditions. The confidence calculation formula is as follows: Simultaneously calculate the rule lift (Lift(X→Y)) using the following formula: when When the value is positively correlated with X, the greater the boost, the stronger the correlation; when... If X and Y are not significantly correlated or negatively correlated, the rule is removed. S404: The top 10 association rules are sorted in descending order of confidence level, and the support, improvement, and applicable time period of each rule are labeled. This serves as a basis for cross-industry consumption guidance. By assigning differentiated scenario weights to different business formats such as gas stations and restaurants, the association value of core service scenarios is highlighted, while avoiding interference from secondary scenario data on rule mining, ensuring that the mining results align with the core needs of service area operations. Introducing a time decay factor effectively reduces the impact of outdated historical consumption behavior, making the association rules more aligned with real-time consumption trends. For example, it prioritizes capturing high-frequency behavior such as purchasing snacks after refueling within the past hour, rather than relying on a few... By using historical data from the past few days and setting minimum support, confidence thresholds, and lift criteria, this algorithm can accurately filter out invalid associations and retain rules with practical guiding significance. Combined with confidence ranking, it outputs the Top 10 rules and marks the applicable time periods, allowing operators to clearly grasp high-value cross-industry associations, such as "42% of gas station users have a chance of dining out within 30 minutes." Compared to the traditional Apriori algorithm, which only mechanically mines frequent itemsets and lacks scenario adaptability and real-time performance, the rules mined by this algorithm are more targeted, timely, and actionable. They can be directly transformed into consumer guidance actions, such as pushing dining coupons to gas station users, significantly improving cross-scenario consumption conversion rates and marketing resource utilization efficiency.

[0013] Furthermore, the time-series analysis model of the intelligent analysis layer uses a bidirectional LSTM (Bi-LSTM) model to predict pedestrian flow at different times. The model input features include historical pedestrian flow for the same period, date type, time period code, weather data, and the cumulative pedestrian flow for the previous three time periods. The output is a predicted pedestrian flow value every 15 minutes for the next 1-3 hours. An adaptive learning rate optimization algorithm is used during model training, with a learning rate of... ,in , , , To train the number of iterations, multi-dimensional input features such as historical pedestrian traffic, date type, and weather data are selected. This fully considers the periodicity of pedestrian traffic changes, such as daily peak hours, differences between weekends and weekdays, the impact of cumulative pedestrian traffic in the top three time periods, and external interference factors like weather. It also leverages the Bi-LSTM model to capture the forward and backward dependencies of time-series data, providing a more comprehensive fit to pedestrian traffic patterns compared to traditional time-series models. Furthermore, the adaptive learning rate algorithm dynamically adjusts the training step size, ensuring rapid convergence in the early stages of model training while avoiding parameter degradation caused by a fixed learning rate in later stages. The oscillation significantly improves model training efficiency and prediction accuracy, ensuring that the output of pedestrian flow predictions every 15 minutes for the next 1-3 hours is more in line with the actual scenario. Compared with the simple statistical or static model prediction of pedestrian flow in existing technologies, this model can accurately predict the distribution of pedestrian flow during peak hours and in popular areas. For example, knowing in advance that the catering area will see a large number of customers during the lunch peak allows the service area to adjust its marketing plan in advance, such as increasing the number of coupons issued, optimizing resource allocation, and opening more checkout lanes, effectively alleviating congestion, improving user experience, and avoiding waste of marketing resources, thus significantly improving the scientific nature and timeliness of operational decisions.

[0014] Furthermore, the personalized marketing plan generation in the marketing execution layer adopts a dynamic discount calculation model, the formula of which is: Where D represents the final discount level, ranging from 0.1 to 0.8. The basic discount level is given by S, which is the user's consumption sensitivity coefficient, ranging from 0 to 1. C represents the product gross profit margin, and T is the real-time scene congestion coefficient, where T = current foot traffic / maximum scene capacity. , , , This is the adjustment coefficient; when When a congestion mitigation mechanism is triggered, an additional discount of 0.1 is added to guide users to lower-traffic scenarios. This is achieved by incorporating user consumption sensitivity coefficients, different user responses to promotions, product gross profit margins, and real-time scenario congestion coefficients, while also adapting to the current service capacity. A fixed adjustment coefficient quantifies the weight of each factor, ensuring the final discount aligns with individual user characteristics (e.g., offering higher discounts to highly sensitive users) while also matching product profit margins with the real-time scenario status. Furthermore, the design of adding an extra discount when the scenario congestion coefficient T > 1.2 not only provides precise profit sharing... By guiding users to lower-traffic areas, congestion pressure is alleviated and user experience is improved. This prevents users from abandoning their purchases due to long queues and avoids the waste of resources or insufficient traffic caused by the one-size-fits-all approach of traditional fixed discount schemes. For example, while pushing higher discounts to highly sensitive users, congestion mitigation mechanisms can guide them to less busy dining areas, increasing user redemption rates and balancing service load across different areas. Compared to existing marketing models that lack dynamic adjustments, this significantly improves the utilization efficiency of discount resources and the practical value of marketing solutions, achieving a win-win situation for user needs, business revenue, and scenario operations.

[0015] Furthermore, the data cleaning algorithm of the data processing layer includes: S701: using a density-based Local Outlier Factor (LOF) algorithm to identify abnormal payment data, setting the LOF threshold to 2.5, and determining that data points with an LOF value greater than 2.5 are abnormal data and removed; S702: processing duplicate transaction records using a sliding window deduplication method, with the window size set to 5 minutes, retaining the first occurrence of transaction records for the same product and amount from the same user within the window; S703: filling missing data using an interpolation method based on K-nearest neighbors (K=5), and determining the similarity based on user attributes and consumption behavior. The algorithm calculates missing values ​​using similarity weighting and identifies and removes abnormal payment data by setting a threshold of 2.5 using the LOF algorithm. This effectively filters out invalid information such as malicious transactions and system errors, preventing abnormal data from interfering with the analysis results. A 5-minute sliding window deduplication method is used to handle duplicate transaction records, accurately removing erroneous operations or duplicate system records by the same user within a short period, ensuring the uniqueness of transaction data. Furthermore, a K=5 nearest neighbor interpolation method is used to fill in missing data. By weighting the calculation based on the similarity of user attributes and consumption behavior, the algorithm can restore the true characteristics of missing data to the greatest extent, avoiding biased user profiles or analytical deviations caused by missing data. Compared to existing technologies that simply remove outliers and missing values, this algorithm not only achieves data integrity and accuracy but also ensures data completeness and accuracy. This leads to more accurate cross-scenario data association, more comprehensive user profile construction, and more scientific marketing plan generation, improving the overall system's operational efficiency and decision-making reliability from the source.

[0016] Furthermore, the user profile construction of the intelligent analysis layer adopts a multi-level feature fusion model, including: Level 1 features: basic attribute features, which are independently encoded and transformed; Level 2 features: consumption behavior features, which are Z-score standardized, with the formula as follows: Where x is the original feature value, μ is the feature mean, and σ is the feature standard deviation; Third-level feature: preference feature, the preference weight is calculated using the TF-IDF algorithm, the formula is: Among them, TF ij Let DF be the percentage of user i's purchase frequency in category j, N be the total number of product categories, and DF be the percentage of user i's purchase frequency in category j. j The model calculates the number of users purchasing category j. It weights and fuses the three-level features and inputs them into a decision tree model, outputting user profile labels. By dividing basic attribute features, consumption behavior features, and preference features into three levels and employing targeted encoding and calculation methods, it ensures effective representation of different types of features. One-hot encoding preserves the category information of basic attributes, Z-score standardization eliminates the dimensional differences of consumption behavior features, and the TF-IDF algorithm accurately quantifies the intensity of user preferences for different product categories. Furthermore, weighted fusion integrates multi-dimensional features, avoiding the limitations of a single feature dimension. The fused features are input into the decision tree model to output profile labels, accurately capturing core user characteristics, such as "frequent refueling + fast food preference" and "long-distance rest + snack purchase preference." Compared to existing technologies that only build profiles based on shallow features or single dimensions, this model generates profiles that better match users' real needs and behavioral patterns. This allows subsequent marketing plans to accurately match the preferences of different user groups, such as pushing retail discount coupons to long-distance travelers who prefer snacks, significantly improving marketing response rates and user satisfaction. It also provides more targeted decision support for cross-industry consumption guidance.

[0017] Furthermore, the real-time consumption guidance information generation of the marketing execution layer adopts a scenario matching degree calculation model, specifically including: S901: Calculate the matching degree M between the user's current scenario and the target scenario, the formula is: Where P represents the user's preference for products in the target scene, Q represents the inverse value of the real-time congestion level in the target scene, and R represents the distance attenuation coefficient from the user's current location to the target scene. , , S902: Filter target scenarios with M≥0.6, and generate structured guidance information based on the current promotional activities in that scenario. The guidance information includes the scenario name, real-time status, navigation path, and estimated arrival time. S903: Adjust the information display format according to the user's current terminal type. Add a one-click navigation function to the mini-program, and highlight distance and queue time on the smart navigation screen. By comprehensively considering the user's preference for products in the target scenario, the real-time congestion inverse value of the target scenario, and the distance attenuation coefficient from the user's current location to the target scenario, and assigning appropriate weights to calculate the matching degree, the most suitable target scenario for the user can be accurately selected, avoiding ineffective guidance. Filtering target scenarios with M≥0.6 and generating structured guidance content containing scenario name, real-time status, etc., allows users to quickly grasp key information. Then, adjust the display format according to the terminal characteristics of the mini-program and smart navigation screen to further improve the practicality and ease of operation of the guidance. For example, the one-click navigation function on the mini-program reduces the user's path-finding cost, and the smart navigation screen highlights distance and queue time to facilitate quick decision-making by on-site users. Compared to existing technologies that lack targeted and generalized guidance, the guidance information generated by this model is more aligned with individual user preferences, real-time scene status, and usage scenarios. This not only effectively improves the user's cross-scenario consumption conversion rate but also balances the service load of different areas by guiding users to less crowded scenes, reducing user waiting time and achieving a dual improvement in user experience and service area operational efficiency.

[0018] Furthermore, the feedback data optimization model of the user interaction layer adopts a closed-loop iterative mechanism, specifically including: S1001: defining marketing effectiveness evaluation indicators, including discount redemption rate η, user dwell time increase rate ΔT, and cross-scenario consumption conversion rate θ; S1002: establishing the correlation function between indicators and model parameters, the formula is: in, Adjust the model parameters by a certain amount. , , For weighting coefficients; S1003: Marketing performance indicators are calculated every 24 hours, when... At that time, the parameters of the machine learning model in the intelligent analysis layer are dynamically adjusted, with an adjustment range of [missing value]. By clearly defining three core evaluation indicators—discount redemption rate, user dwell time increase rate, and cross-scenario consumption conversion rate—marketing effectiveness and user feedback can be comprehensively quantified. Establishing a correlation function between indicators and model parameters and assigning appropriate weights allows for precise calculation of model adjustment ranges, ensuring clear data support for parameter optimization. Indicators are statistically analyzed every 24 hours and... The intelligent analysis layer model parameters are dynamically adjusted in real time, ensuring timely optimization while avoiding blind modifications through quantitative adjustments, allowing the model to continuously iterate towards a direction that better aligns with actual operational needs. Compared to existing technologies that lack feedback optimization mechanisms and have fixed model parameters, this mechanism allows the system to dynamically adjust based on marketing results, such as optimizing user profile feature weights based on redemption rate increases, continuously improving marketing accuracy, user response rate, and cross-scenario conversion effectiveness. Attached Figure Description

[0019] Figure 1 is a flowchart of the present invention; Specific Implementation Example 1, as shown in Figure 1, describes a service area intelligent marketing and consumption guidance system based on cross-scenario data fusion. This system includes a data processing layer, an intelligent analysis layer, a marketing execution layer, and a user interaction layer. The system achieves personalized marketing and real-time consumption guidance through cross-scenario multi-source data fusion and intelligent algorithm-driven implementation. Specifically, it includes: a data acquisition layer: deploying IoT devices and system interfaces across multiple service area scenarios to collect user consumption data, user attribute data, and scenario environment data in real time, and sending them to the data processing layer via an encrypted transmission protocol; a data processing layer: using a distributed database to store multi-source data, performing data cleaning and standardization, and then using a data fusion engine to achieve cross-scenario data association; an intelligent analysis layer: based on the fused dataset, constructing user profiles and extracting consumption preferences through machine learning models, predicting pedestrian traffic and consumption hotspots using time-series analysis models, and discovering cross-industry consumption patterns using association rule mining algorithms; a marketing execution layer: automatically generating personalized marketing plans based on the intelligent analysis results, pushing customized discounts and real-time consumption guidance information through multiple channels, and automating the configuration and execution of promotional activities; and a user interaction layer: providing multi-terminal interaction entry points, recording user interaction behavior, and feeding it back to the intelligent analysis layer for model optimization.

[0020] The principle behind this solution is to build a closed-loop system that integrates "data collection, processing, analysis, execution, and feedback," with cross-scenario data fusion as the core support, and to achieve personalized marketing and real-time consumption guidance through layered collaboration. The data acquisition layer, relying on the multi-business scenarios of the service area, deploys IoT devices and system interfaces to comprehensively capture user consumption data, attribute data, and scene environment data. Encrypted transmission ensures secure data transfer to the data processing layer. The data processing layer uses a distributed database to store multi-source heterogeneous data. After cleaning and removing invalid information and standardizing the data format, a data fusion engine establishes cross-scene data associations, forming a complete dataset of user behavior and scene status. The intelligent analysis layer, based on the fused data, uses machine learning models to mine user consumption preferences, time-series analysis models to predict foot traffic and consumption hotspots, and association rule mining algorithms to discover cross-business consumption patterns, providing accurate basis for marketing decisions. The marketing execution layer automatically generates personalized plans based on the analysis results, pushing customized offers and real-time guidance information through multiple channels to automate promotional activities. The user interaction layer provides multi-terminal interaction entry points, synchronously recording user interaction behavior and feeding it back to the intelligent analysis layer to continuously optimize model performance and ensure the system adapts to dynamically changing service scenarios and user needs.

[0021] Compared to existing technologies, this solution comprehensively collects three types of data—consumption, attributes, and environment—through multi-scenario IoT devices. Then, a data fusion engine enables cross-industry data correlation, forming a complete view of the user's end-to-end behavior. Existing technologies can only acquire data from a single industry, such as collecting only refueling records or catering consumption data. This solution, however, can integrate and correlate user refueling behavior, catering consumption, and retail shopping data. For example, it can identify the user behavior pattern of "purchasing snacks within 30 minutes of refueling," providing a data foundation for precision marketing. Secondly, it breaks down the barrier between customer profiling and marketing execution. Existing technologies can only output static customer profiles for analysis, while this solution, through an intelligent analysis layer and marketing... The coordinated action at the sales execution layer directly translates user preferences into real-time marketing actions. For example, it automatically pushes restaurant discount coupons to users with "frequent refueling and fast food preferences." Compared to the generalized marketing of existing technologies, such as pushing uniform coupons to all users, the marketing accuracy is significantly improved. This solution uses behavioral data recorded at the user interaction layer to optimize the model in reverse. For example, it adjusts marketing plans based on user coupon redemption rates, enabling the system to continuously adapt to changes in user needs. Existing marketing plans are mostly manually preset and fixed, while this solution can achieve dynamic iteration of marketing strategies. For example, during peak hours, it can adjust the number of coupons issued in real time based on traffic flow to guide user flow, greatly improving operational efficiency.

[0022] The IoT devices in the data acquisition layer include smart fuel dispenser data acquisition modules at gas stations, smart POS machines in restaurants, QR code payment terminals in retail stores, user interaction kiosks at tourist information points, and vehicle recognition devices in parking lots. User consumption data includes consumption amount, product SKU code, payment channel, and consumption duration. User attribute data includes license plate number, vehicle type, member ID, historical consumption frequency, and consumption amount threshold. Scene environment data includes real-time pedestrian flow, queue length, regional temperature, and light intensity for each scene. By deploying targeted IoT devices in core business areas such as gas stations and restaurants, the system can cover the entire user experience, from vehicle recognition upon entering the service area to refueling, dining, shopping, and information access to tourist information. The entire process ensures that no key user touchpoint data is missed. By specifying specific data items such as consumption amount, license plate number, and real-time pedestrian flow, it not only guarantees the standardization and targeting of data collection and avoids invalid data redundancy, but also builds a multi-dimensional and three-dimensional view of users and scenarios by complementing consumption data, user attribute data, and scenario environment data. For example, by combining license plate number, fuel consumption amount, and queue time in the catering area, it is possible to accurately identify high-spending users traveling long distances and push guidance to catering services that do not require queuing. Compared with existing technologies that only collect single types of data, it can more accurately capture user needs and scenario status, providing reliable support for subsequent data fusion, intelligent analysis, and personalized marketing plan generation, and significantly improving the scientific nature and execution effect of marketing decisions.

[0023] The data fusion engine in the data processing layer uses a weighted association algorithm to achieve cross-scenario data association, specifically including: S301: Assigning a unique identifier ID to each user. The unique identifier ID is generated through a license plate number hash operation or association with a member ID. The hash operation formula is: Unique Identifier ID = SHA-256 The service area code is a fixed 6-digit number, and the timestamp salt value is a millisecond-level timestamp of the data collection time. Used to map hash results to 12-bit unique identifiers; S302: Defines cross-scenario data association weights and user attribute data association weights. Consumer behavior similarity weight Time-related weights The formula for calculating the total correlation degree is: in, The user attribute matching score ranges from 0 to 1, with 1 representing a perfect match. For similarity of consumer product categories, The decay coefficient is the time interval between two consumptions. t is the time interval between two consumptions, and T is the decay threshold; S303: When the total correlation R ≥ 0.7, it is determined to be cross-scenario data of the same user and associated with the same unique identifier ID; when When this occurs, a secondary verification mechanism is triggered, supplementing verification by comparing the last four digits of the user's payment channel number; when When user data is identified as different from other user data, it is stored separately. A 12-bit unique identifier ID is generated using SHA-256 hashing combined with service area coding and millisecond-level timestamp salt. This ensures both the uniqueness and security of user identification while also accommodating the identity association needs of non-member and member users, solving the problem of unifying the identification of different user identity data. By setting three weights—user attributes, consumption behavior similarity, and time correlation—and constructing a formula for calculating the overall correlation, the correlation strength of data across different dimensions can be quantified. Furthermore, the time interval decay coefficient reflects the objective law that shorter intervals between consumption behaviors indicate stronger correlation, avoiding the bias caused by a single-dimensional correlation. This approach avoids misjudgments. By setting hierarchical judgment rules—direct association for R≥0.7, secondary verification for 0.4≤R<0.7, and separate storage for R<0.4—it can quickly and accurately associate high-confidence cross-scenario data, reduce the false association rate of medium-confidence data through secondary verification, and avoid interference from low-confidence data. Compared with existing technologies' simple feature matching or unweighted association methods, this significantly improves the accuracy and reliability of cross-scenario data association, ensuring that the fused data can truly reflect the user's full-link behavior. This provides high-quality data support for the subsequent intelligent analysis layer to extract consumer preferences and discover cross-industry patterns, thereby improving the accuracy of personalized marketing solutions.

[0024] The association rule mining algorithm of the intelligent analysis layer adopts the improved Apriori algorithm, which introduces a time decay factor and a scene weight coefficient to optimize frequent itemset mining. Specifically, it includes: S401: Defining cross-business consumption itemsets. ,in For consumption behavior in a single business format, for each consumption item Assign scene weights ,Gas station Restaurants retail stores Tourist Information Points S402: Introduce a time decay factor β when calculating itemset support. Where λ is the attenuation coefficient, by default. , The improved itemset support formula is: (The interval between the time of the consumption behavior and the current time is given.) in, The indicator function takes a value of 1 if the condition occurs and 0 otherwise, where N is the total number of users within the statistical period; S403: Set the minimum support threshold to 0.05 and the minimum confidence threshold to 0.3, and mine cross-business consumption association rules X→Y that meet the conditions. The confidence calculation formula is as follows: Simultaneously calculate the rule lift (Lift(X→Y)) using the following formula: when When the value is positively correlated with X, the greater the boost, the stronger the correlation; when... If X and Y are not significantly correlated or negatively correlated, the rule is removed. S404: The top 10 association rules are sorted in descending order of confidence level, and the support, improvement, and applicable time period of each rule are labeled. This serves as a basis for cross-industry consumption guidance. By assigning differentiated scenario weights to different business formats such as gas stations and restaurants, the association value of core service scenarios is highlighted, while avoiding interference from secondary scenario data on rule mining, ensuring that the mining results align with the core needs of service area operations. Introducing a time decay factor effectively reduces the impact of outdated historical consumption behavior, making the association rules more aligned with real-time consumption trends. For example, it prioritizes capturing high-frequency behavior such as purchasing snacks after refueling within the past hour, rather than relying on a few... By using historical data from the past few days and setting minimum support, confidence thresholds, and lift criteria, this algorithm can accurately filter out invalid associations and retain rules with practical guiding significance. Combined with confidence ranking, it outputs the Top 10 rules and marks the applicable time periods, allowing operators to clearly grasp high-value cross-industry associations, such as "42% of gas station users have a chance of dining out within 30 minutes." Compared to the traditional Apriori algorithm, which only mechanically mines frequent itemsets and lacks scenario adaptability and real-time performance, the rules mined by this algorithm are more targeted, timely, and actionable. They can be directly transformed into consumer guidance actions, such as pushing dining coupons to gas station users, significantly improving cross-scenario consumption conversion rates and marketing resource utilization efficiency.

[0025] The time-series analysis model of the intelligent analysis layer uses a bidirectional LSTM (Bi-LSTM) model to predict pedestrian flow at different times. The model input features include historical pedestrian flow for the same period, date type, time period code, weather data, and the cumulative pedestrian flow for the previous three time periods. The output is a predicted pedestrian flow value every 15 minutes for the next 1-3 hours. An adaptive learning rate optimization algorithm is used during model training. ,in , , , To train the number of iterations, multi-dimensional input features such as historical pedestrian traffic, date type, and weather data are selected. This fully considers the periodicity of pedestrian traffic changes, such as daily peak hours, differences between weekends and weekdays, the impact of cumulative pedestrian traffic in the top three time periods, and external interference factors like weather. It also leverages the Bi-LSTM model to capture the forward and backward dependencies of time-series data, providing a more comprehensive fit to pedestrian traffic patterns compared to traditional time-series models. Furthermore, the adaptive learning rate algorithm dynamically adjusts the training step size, ensuring rapid convergence in the early stages of model training while avoiding parameter degradation caused by a fixed learning rate in later stages. The oscillation significantly improves model training efficiency and prediction accuracy, ensuring that the output of pedestrian flow predictions every 15 minutes for the next 1-3 hours is more in line with the actual scenario. Compared with the simple statistical or static model prediction of pedestrian flow in existing technologies, this model can accurately predict the distribution of pedestrian flow during peak hours and in popular areas. For example, knowing in advance that the catering area will see a large number of customers during the lunch peak allows the service area to adjust its marketing plan in advance, such as increasing the number of coupons issued, optimizing resource allocation, and opening more checkout lanes, effectively alleviating congestion, improving user experience, and avoiding waste of marketing resources, thus significantly improving the scientific nature and timeliness of operational decisions.

[0026] The personalized marketing plan generation in the marketing execution layer uses a dynamic discount calculation model, the formula of which is: Where D represents the final discount level, ranging from 0.1 to 0.8. The basic discount level is given by S, which is the user's consumption sensitivity coefficient, ranging from 0 to 1. C represents the product gross profit margin, and T is the real-time scene congestion coefficient, where T = current foot traffic / maximum scene capacity. , , , This is the adjustment coefficient; when When a congestion mitigation mechanism is triggered, an additional discount of 0.1 is added to guide users to lower-traffic scenarios. This is achieved by incorporating user consumption sensitivity coefficients, different user responses to promotions, product gross profit margins, and real-time scenario congestion coefficients, while also adapting to the current service capacity. A fixed adjustment coefficient quantifies the weight of each factor, ensuring the final discount aligns with individual user characteristics (e.g., offering higher discounts to highly sensitive users) while also matching product profit margins with the real-time scenario status. Furthermore, the design of adding an extra discount when the scenario congestion coefficient T > 1.2 not only provides precise profit sharing... By guiding users to lower-traffic areas, congestion pressure is alleviated and user experience is improved. This prevents users from abandoning their purchases due to long queues and avoids the waste of resources or insufficient traffic caused by the one-size-fits-all approach of traditional fixed discount schemes. For example, while pushing higher discounts to highly sensitive users, congestion mitigation mechanisms can guide them to less busy dining areas, increasing user redemption rates and balancing service load across different areas. Compared to existing marketing models that lack dynamic adjustments, this significantly improves the utilization efficiency of discount resources and the practical value of marketing solutions, achieving a win-win situation for user needs, business revenue, and scenario operations.

[0027] The data cleaning algorithm of the data processing layer includes: S701: Identifying abnormal payment data using a density-based Local Outlier Factor (LOF) algorithm, setting the LOF threshold to 2.5. When the LOF value of a data point is greater than 2.5, it is determined to be abnormal data and removed; S702: Processing duplicate transaction records using a sliding window deduplication method, with the window size set to 5 minutes. For the same user, the first transaction record of the same product and the same amount within the window is retained; S703: Filling missing data using an interpolation method based on K-nearest neighbors (K=5), based on user attribute similarity and consumption behavior similarity. The algorithm employs a weighted calculation method to identify and remove missing values, using the LOF algorithm with a threshold of 2.5 to detect and eliminate abnormal payment data. This effectively filters out invalid information such as malicious transactions and system errors, preventing abnormal data from interfering with analysis results. A 5-minute sliding window deduplication method is used to process duplicate transaction records, accurately eliminating erroneous operations or duplicate system records by the same user within a short period, ensuring the uniqueness of transaction data. Furthermore, a K=5 nearest neighbor interpolation method is used to fill in missing data. By weighting the calculation based on the similarity between user attributes and consumption behavior, the algorithm can restore the true characteristics of missing data to the greatest extent possible, avoiding biased user profiles or analytical deviations caused by missing data. Compared to existing technologies that simply remove outliers and missing values, this algorithm not only achieves data integrity and accuracy but also ensures data completeness and accuracy. This leads to more accurate cross-scenario data association, more comprehensive user profile construction, and more scientific marketing plan generation, improving the overall system's operational efficiency and decision-making reliability from the source.

[0028] The user profile construction of the intelligent analysis layer adopts a multi-level feature fusion model, including: Level 1 features: basic attribute features, which are independently encoded and transformed; Level 2 features: consumption behavior features, which are Z-score standardized, with the formula as follows: Where x is the original feature value, μ is the feature mean, and σ is the feature standard deviation; Third-level feature: preference feature, the preference weight is calculated using the TF-IDF algorithm, the formula is: Among them, TF ij Let DF be the percentage of user i's purchase frequency in category j, N be the total number of product categories, and DF be the percentage of user i's purchase frequency in category j. j The model calculates the number of users purchasing category j. It weights and fuses the three-level features and inputs them into a decision tree model, outputting user profile labels. By dividing basic attribute features, consumption behavior features, and preference features into three levels and employing targeted encoding and calculation methods, it ensures effective representation of different types of features. One-hot encoding preserves the category information of basic attributes, Z-score standardization eliminates the dimensional differences of consumption behavior features, and the TF-IDF algorithm accurately quantifies the intensity of user preferences for different product categories. Furthermore, weighted fusion integrates multi-dimensional features, avoiding the limitations of a single feature dimension. The fused features are input into the decision tree model to output profile labels, accurately capturing core user characteristics, such as "frequent refueling + fast food preference" and "long-distance rest + snack purchase preference." Compared to existing technologies that only build profiles based on shallow features or single dimensions, this model generates profiles that better match users' real needs and behavioral patterns. This allows subsequent marketing plans to accurately match the preferences of different user groups, such as pushing retail discount coupons to long-distance travelers who prefer snacks, significantly improving marketing response rates and user satisfaction. It also provides more targeted decision support for cross-industry consumption guidance.

[0029] The real-time consumption guidance information generation in the marketing execution layer adopts a scenario matching degree calculation model, specifically including: S901: Calculate the matching degree M between the user's current scenario and the target scenario, the formula is: Where P represents the user's preference for products in the target scene, Q represents the inverse value of the real-time congestion level in the target scene, and R represents the distance attenuation coefficient from the user's current location to the target scene. , , S902: Filter target scenarios with M≥0.6, and generate structured guidance information based on the current promotional activities in that scenario. The guidance information includes the scenario name, real-time status, navigation path, and estimated arrival time. S903: Adjust the information display format according to the user's current terminal type. Add a one-click navigation function to the mini-program, and highlight distance and queue time on the smart navigation screen. By comprehensively considering the user's preference for products in the target scenario, the real-time congestion inverse value of the target scenario, and the distance attenuation coefficient from the user's current location to the target scenario, and assigning appropriate weights to calculate the matching degree, the most suitable target scenario for the user can be accurately selected, avoiding ineffective guidance. Filtering target scenarios with M≥0.6 and generating structured guidance content containing scenario name, real-time status, etc., allows users to quickly grasp key information. Then, adjust the display format according to the terminal characteristics of the mini-program and smart navigation screen to further improve the practicality and ease of operation of the guidance. For example, the one-click navigation function on the mini-program reduces the user's path-finding cost, and the smart navigation screen highlights distance and queue time to facilitate quick decision-making by on-site users. Compared to existing technologies that lack targeted and generalized guidance, the guidance information generated by this model is more aligned with individual user preferences, real-time scene status, and usage scenarios. This not only effectively improves the user's cross-scenario consumption conversion rate but also balances the service load of different areas by guiding users to less crowded scenes, reducing user waiting time and achieving a dual improvement in user experience and service area operational efficiency.

[0030] The feedback data optimization model of the user interaction layer adopts a closed-loop iterative mechanism, specifically including: S1001: Defining marketing effectiveness evaluation indicators, including discount redemption rate η, user dwell time increase rate ΔT, and cross-scenario consumption conversion rate θ; S1002: Establishing the correlation function between indicators and model parameters, the formula is: in, Adjust the model parameters by a certain amount. , , For weighting coefficients; S1003: Marketing performance indicators are calculated every 24 hours, when... At that time, the parameters of the machine learning model in the intelligent analysis layer are dynamically adjusted, with an adjustment range of [missing value]. By clearly defining three core evaluation indicators—discount redemption rate, user dwell time increase rate, and cross-scenario consumption conversion rate—marketing effectiveness and user feedback can be comprehensively quantified. Establishing a correlation function between indicators and model parameters and assigning appropriate weights allows for precise calculation of model adjustment ranges, ensuring clear data support for parameter optimization. Indicators are statistically analyzed every 24 hours and... The intelligent analysis layer model parameters are dynamically adjusted in real time, ensuring timely optimization while avoiding blind modifications through quantitative adjustments, allowing the model to continuously iterate towards a direction that better aligns with actual operational needs. Compared to existing technologies that lack feedback optimization mechanisms and have fixed model parameters, this mechanism allows the system to dynamically adjust based on marketing results, such as optimizing user profile feature weights based on redemption rate increases, continuously improving marketing accuracy, user response rate, and cross-scenario conversion effectiveness.

[0031] The above are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the description of specific embodiments in the specification can be used to interpret the content of the claims.

Claims

1. A service area intelligent marketing and consumption guidance system based on cross-scenario data fusion, characterized in that, The system comprises a data processing layer, an intelligent analysis layer, a marketing execution layer, and a user interaction layer. It achieves personalized marketing and real-time consumption guidance through cross-scenario multi-source data fusion and intelligent algorithm-driven approaches. Specifically, it includes: a data acquisition layer: deploying IoT devices and system interfaces across multiple service areas to collect user consumption data, user attribute data, and scene environment data in real time, and sending them to the data processing layer via encrypted transmission protocols; a data processing layer: using a distributed database to store multi-source data, and after data cleaning and standardization, achieving cross-scenario data association through a data fusion engine; an intelligent analysis layer: based on the fused dataset, constructing user profiles and extracting consumption preferences through machine learning models, predicting pedestrian traffic and consumption hotspots using time-series analysis models, and discovering cross-industry consumption patterns using association rule mining algorithms; a marketing execution layer: automatically generating personalized marketing plans based on the intelligent analysis results, pushing customized offers and real-time consumption guidance information through multiple channels, and automating the configuration and execution of promotional activities; and a user interaction layer: providing multi-terminal interaction entry points, recording user interaction behavior, and feeding it back to the intelligent analysis layer for model optimization.

2. The service area intelligent marketing and consumption guidance system based on cross-scenario data fusion as described in claim 1, characterized in that, The IoT devices in the data acquisition layer include smart fuel dispenser data acquisition modules at gas stations, smart POS machines at restaurants, QR code payment terminals at retail stores, user interaction kiosks at tourist information points, and vehicle recognition devices in parking lots. The user consumption data includes consumption amount, product SKU code, payment channel, and consumption duration. The user attribute data includes license plate number, vehicle type, member ID, historical consumption frequency, and consumption amount threshold. The scene environment data includes real-time pedestrian flow, queuing time, regional temperature, and light intensity in each scene.

3. The service area intelligent marketing and consumption guidance system based on cross-scenario data fusion according to claim 1, characterized in that, The data fusion engine in the data processing layer uses a weighted association algorithm to achieve cross-scenario data association, specifically including: S301: Assigning a unique identifier ID to each user. The unique identifier ID is generated through a license plate number hash operation or association with a member ID. The hash operation formula is: Unique Identifier ID = SHA-256 The service area code is a fixed 6-digit number, and the timestamp salt value is a millisecond-level timestamp of the data collection time. Used to map hash results to 12-bit unique identifiers; S302: Defines cross-scenario data association weights and user attribute data association weights. Consumer behavior similarity weight Time-related weights The formula for calculating the total correlation degree is: in, The user attribute matching score ranges from 0 to 1, with 1 representing a perfect match. For similarity of consumer product categories, The decay coefficient is the time interval between two consumptions. t is the time interval between two consumptions, and T is the decay threshold; S303: When the total correlation R ≥ 0.7, it is determined to be cross-scenario data of the same user and associated with the same unique identifier ID; when When this occurs, a secondary verification mechanism is triggered, supplementing verification by comparing the last four digits of the user's payment channel number; when When these are identified as different user data, they are stored separately.

4. The service area intelligent marketing and consumption guidance system based on cross-scenario data fusion according to claim 1, characterized in that, The association rule mining algorithm of the intelligent analysis layer adopts the improved Apriori algorithm, which introduces a time decay factor and a scene weight coefficient to optimize frequent itemset mining. Specifically, it includes: S401: Defining cross-business consumption itemsets. ,in For consumption behavior in a single business format, for each consumption item Assign scene weights ,Gas station Restaurants retail stores Tourist Information Points S402: Introduce a time decay factor β when calculating itemset support. Where λ is the attenuation coefficient, by default. ， The improved itemset support formula is: (The interval between the time of the consumption behavior and the current time is given.) in, The indicator function takes a value of 1 if the condition occurs and 0 otherwise, where N is the total number of users within the statistical period; S403: Set the minimum support threshold to 0.05 and the minimum confidence threshold to 0.3, and mine cross-business consumption association rules X→Y that meet the conditions. The confidence calculation formula is as follows: Simultaneously calculate the rule lift (Lift(X→Y)) using the following formula: when When the value is positively correlated with X, the greater the boost, the stronger the correlation; when... If X and Y are not significantly correlated or negatively correlated, the rule is removed; S404: Sort by confidence level in descending order, output the top 10 association rules, and mark the support, improvement and applicable time period of each rule as the basis for cross-industry consumption guidance.

5. The service area intelligent marketing and consumption guidance system based on cross-scenario data fusion according to claim 1, characterized in that, The time-series analysis model of the intelligent analysis layer uses a bidirectional LSTM (Bi-LSTM) model to predict pedestrian flow at different times. The model input features include historical pedestrian flow for the same period, date type, time period code, weather data, and the cumulative pedestrian flow for the previous three time periods. The output is a predicted pedestrian flow value every 15 minutes for the next 1-3 hours. An adaptive learning rate optimization algorithm is used during model training. ,in ， ， t is the number of training iterations.

6. The service area intelligent marketing and consumption guidance system based on cross-scenario data fusion according to claim 1, characterized in that, The personalized marketing plan generation in the marketing execution layer uses a dynamic discount calculation model, the formula of which is: Where D represents the final discount level, ranging from 0.1 to 0.

8. The basic discount level is given by S, which is the user's consumption sensitivity coefficient, ranging from 0 to 1. C represents the product gross profit margin, and T is the real-time scene congestion coefficient, where T = current foot traffic / maximum scene capacity. ， ， ， This is the adjustment coefficient; when When the congestion relief mechanism is triggered, an additional discount of 0.1 is added to guide users to low-traffic scenarios.

7. The service area intelligent marketing and consumption guidance system based on cross-scenario data fusion according to claim 1, characterized in that, The data cleaning algorithm of the data processing layer includes: S701: using the density-based Local Outlier Factor (LOF) algorithm to identify abnormal payment data, setting the LOF threshold to 2.5, and judging data points as abnormal data and removing them when the LOF value is greater than 2.5; S702: processing duplicate transaction records using the sliding window deduplication method, with the window size set to 5 minutes, retaining the first occurrence of transaction records of the same product and the same amount for the same user within the window; S703: filling missing data using the K-nearest neighbor (K=5) based interpolation method, and calculating missing values ​​based on the weighted similarity of user attributes and consumption behavior.

8. The service area intelligent marketing and consumption guidance system based on cross-scenario data fusion according to claim 1, characterized in that, The user profile construction of the intelligent analysis layer adopts a multi-level feature fusion model, including: Level 1 features: basic attribute features, which are independently encoded and transformed; Level 2 features: consumption behavior features, which are Z-score standardized, with the formula as follows: Where x is the original feature value, μ is the feature mean, and σ is the feature standard deviation; Third-level feature: preference feature, the preference weight is calculated using the TF-IDF algorithm, the formula is: Among them, TF ij Let DF be the percentage of user i's purchase frequency in category j, N be the total number of product categories, and DF be the percentage of user i's purchase frequency in category j. j The number of users who purchase category j; the weighted fusion of the three-level features is input into the decision tree model, and the user profile labels are output.

9. The service area intelligent marketing and consumption guidance system based on cross-scenario data fusion according to claim 1, characterized in that, The real-time consumption guidance information generation in the marketing execution layer adopts a scenario matching degree calculation model, specifically including: S901: Calculate the matching degree M between the user's current scenario and the target scenario, the formula is: Where P represents the user's preference for products in the target scene, Q represents the inverse value of the real-time congestion level in the target scene, and R represents the distance attenuation coefficient from the user's current location to the target scene. ， ， S902: Filter target scenarios with M≥0.6, combine the current promotional activities of the scenario, and generate structured guidance information. The guidance information includes scenario name, real-time status, navigation path and estimated arrival time. S903: Adjust the information display format according to the user's current terminal type. Add a one-click navigation function on the mini-program terminal and highlight the distance and queuing time on the smart guide screen terminal.

10. The service area intelligent marketing and consumption guidance system based on cross-scenario data fusion according to claim 1, characterized in that, The feedback data optimization model of the user interaction layer adopts a closed-loop iterative mechanism, specifically including: S1001: Defining marketing effectiveness evaluation indicators, including discount redemption rate η, user dwell time increase rate ΔT, and cross-scenario consumption conversion rate θ; S1002: Establishing a correlation function between indicators and model parameters, the formula is: in, Adjust the model parameters by a certain amount. ， ， For weighting coefficients; S1003: Marketing performance indicators are calculated every 24 hours, when... At that time, the parameters of the machine learning model in the intelligent analysis layer are dynamically adjusted, with an adjustment range of [missing value]. 。

Images