Community APP advertisement interaction and instant distribution system and method

Abstract

The invention provides a community APP advertisement interaction and instant delivery system and method. According to the method, community-related advertisements are pushed on an APP chat or a home page, after a user clicks the advertisements, the APP obtains the current position of the user, after the user clicks the advertisements, position obtaining is triggered, an order is distributed to a delivery party, and a delivery route is planned according to the position of the user; and the user position is updated in real time to ensure accurate delivery. Dispatching delivery personnel or equipment to deliver the commodities; user positions are synchronized in real time, and distribution routes are adjusted. According to the invention, the conversion rate of community advertisements is improved, and convenient shopping experience is provided for users. Seamless connection of advertisement and distribution is realized, convenience and commercial value of community service are improved, keyword matching degree is improved, and user click rate is increased compared with that of a traditional mode; and the distribution efficiency is optimized.

Description

Technical Field

[0001] This invention relates to the field of instant delivery technology, and in particular to a community APP advertising interaction and instant delivery method, system, electronic device and computer-readable storage medium. Background Technology

[0002] There are three core pain points in existing community services:

[0003] First, the targeting accuracy of advertisements is insufficient. Traditional community advertisements often adopt a broad-based approach, resulting in a click-through rate of less than 2%, which leads to a waste of marketing costs for businesses.

[0004] Secondly, delivery efficiency is low, with manual dispatch response delays exceeding 3 minutes, and order backlog rates reaching 15% during peak periods.

[0005] Third, the positioning accuracy is insufficient. In indoor scenarios, the error of a single GPS positioning exceeds 50 meters, resulting in a mismatch between the delivery person's and the user's location. Summary of the Invention

[0006] To address the technical problems existing in the prior art, the present invention provides the following technical solution:

[0007] On the one hand, a community APP advertising interaction and instant delivery system is provided, including:

[0008] The front-end system runs on the user's mobile terminal and is used for advertising display, triggering user interaction, real-time location acquisition, and order tracking.

[0009] The backend system adopts a microservice architecture, including user management service, advertising management service, order management service, delivery scheduling service and location service, which are used to process user data, advertising, order generation and scheduling, and real-time location synchronization;

[0010] The delivery system, including delivery robots and / or delivery personnel apps, is used to receive and execute delivery tasks;

[0011] The intelligent module communicates and connects with the backend system to perform user profile building, advertising recommendation, delivery scheduling optimization, and location prediction.

[0012] Preferably, the front-end system includes:

[0013] The ad display module is used to trigger and display ads based on user profile, keyword matching of chat content, or user geolocation.

[0014] The interaction trigger module is used to respond to user actions, generate orders, and call the payment interface;

[0015] The real-time location acquisition module adopts a multi-source positioning fusion scheme and improves positioning accuracy through a Kalman filter algorithm.

[0016] Preferably, the delivery scheduling service in the backend system is configured as follows: dynamic allocation of delivery resources is performed using the reinforcement learning DQN algorithm, and the order execution order of the delivery robots is optimized using a genetic algorithm.

[0017] Preferably, the delivery robot includes:

[0018] The navigation module uses LiDAR, camera, and IMU to achieve SLAM navigation and centimeter-level positioning;

[0019] The path planning module uses the A* algorithm for global path planning and a dynamic window method for local obstacle avoidance.

[0020] The communication module synchronizes order and location data with the backend system via the MQTT protocol.

[0021] Preferably, the intelligent module includes:

[0022] The user profiling model uses a hybrid network of BERT+BiLSTM+Attention to extract interest tags from user behavior, text and attribute data.

[0023] The advertising recommendation model uses the DeepFM model to predict the probability of user clicks.

[0024] The location prediction model uses an LSTM network to predict future locations based on users' historical location sequences.

[0025] On the other hand, a method for community APP advertising interaction and instant delivery is provided, which is implemented based on the above-mentioned system and includes the following steps:

[0026] S1. Push contextualized advertisements to users through the front-end system and obtain user interaction behavior;

[0027] S2. In response to the user's order placement, generate an order and obtain the user's real-time location;

[0028] S3. Through the delivery scheduling service of the backend system, delivery resources are allocated based on the reinforcement learning model, and the delivery route is optimized using the genetic algorithm;

[0029] S4. Send the order information to the delivery system, and the delivery robot or delivery personnel will carry out the delivery task;

[0030] S5. During the delivery process, the location of the user and the delivery terminal are synchronized in real time through location services, and the delivery progress is displayed on the front-end system.

[0031] Preferably, the step of pushing contextualized advertisements includes:

[0032] Push targeted ads on the homepage based on user profiles;

[0033] By analyzing user chat content using NLP, related ads are triggered when the keyword matching degree is ≥80%;

[0034] When a user enters a predefined geofence area, location-related advertisements are triggered.

[0035] Preferably, the step of obtaining the user's real-time location includes:

[0036] Collect multi-source positioning data from GPS, base stations, and WiFi;

[0037] The Kalman filter algorithm is used to fuse multi-source data to improve indoor and outdoor positioning accuracy.

[0038] Preferably, the delivery scheduling step includes:

[0039] Construct a state space that includes delivery resource status, order attributes, and time information;

[0040] The DQN model outputs resource allocation actions to minimize total delivery time and cost.

[0041] Genetic algorithms are used to optimize the order sequence of delivery robots in order to shorten the total delivery distance and time.

[0042] Preferably, it further includes a location prediction step:

[0043] Collect the user's location sequence over the past 5 minutes as input;

[0044] Predict the user's location coordinates for the next minute using a pre-trained LSTM model;

[0045] Delivery routes are dynamically adjusted based on forecast results.

[0046] On the other hand, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions, which, when executed by the processor, implement any one of the methods described above for community APP advertising interaction and instant delivery.

[0047] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the methods described above for community APP advertising interaction and instant delivery.

[0048] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0049] 1. Improved ad accuracy: By using a hybrid model of Word2Vec and DeepFM to achieve scenario-based triggering, keyword matching accuracy is increased to over 80%, and user click-through rate is 3 times higher than traditional methods;

[0050] 2. Delivery efficiency optimization: By adopting genetic algorithm + DQN dynamic scheduling, the order allocation delay is ≤50ms, the delivery time is reduced by 25%, and the peak period order backlog rate is reduced to below 3%;

[0051] 3. Enhanced positioning accuracy: The multi-source fusion Kalman filter algorithm controls the indoor positioning error to 2-5 meters and the outdoor accuracy to 1-3 meters, with a location prediction accuracy of ≥85%. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a schematic diagram of the architecture of a community APP advertising interaction and instant delivery system provided in an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of the component system structure of a delivery robot provided in an embodiment of the present invention;

[0055] Figure 3 This is a flowchart of a community APP advertising interaction and instant delivery method provided by an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0058] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0059] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0060] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0061] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0062] The specific implementation details of this system and method will be described in detail below.

[0063] I. System Overall Architecture Principles

[0064] As attached Figure 1 As shown, this system adopts a four-layer distributed architecture of "front-end-back-delivery end-intelligent module" to realize a closed loop of the entire process of accurate advertising push, instant order generation, intelligent delivery scheduling, and real-time location synchronization.

[0065] (I) Front-end system principles (community APP)

[0066] The front end is the entry point for users to interact with the system, and its core functions include ad display, interaction triggering, location acquisition, and order tracking.

[0067] 1. Hardware Dependency

[0068] User devices: smartphones (supporting GPS / BeiDou, 4G / 5G, WiFi), tablets.

[0069] Sensors: The device has a built-in GPS module, accelerometer, and gyroscope (for positioning assistance).

[0070] 2. Software Module Principles

[0071] (1) Advertising display module

[0072] Contextualized display strategy:

[0073] Homepage push: Based on user profiles (such as "stay-at-home moms" and "fresh food lovers"), display community business advertisements (such as fresh food store promotions, housekeeping services, and community group buying) within a 3-kilometer radius in the "Community Recommendations" section of the homepage, using a hybrid format of "card-style + short video" (the short video is uploaded by the advertiser and is 15-30 seconds long).

[0074] Chat Interface Insertion: Using NLP analysis of user chat content (e.g., "I want to eat hot pot tonight"), a related advertisement (e.g., "Hot pot ingredient store near my neighborhood, ¥10 off for purchases over ¥50") pops up on the right side of the chat window. The trigger condition is "keyword matching degree ≥ 80%" (the keyword database is generated by an offline-trained Word2Vec model. The training data includes community users' historical chat records (de-anonymized), product names, and category tags, with a total sample size of 1 million. The model uses a Skip-gram architecture, with a vector dimension of 200, a sliding window size of 5, 5 negative samples, and 20 rounds of iterative training. Keyword matching degree is calculated using cosine similarity; the advertisement is triggered when the similarity between the user's input text and the advertisement's keyword vector is ≥ 80%).

[0075] Location-based triggering: When a user enters a specific area of ​​the community (such as the entrance of the community or next to a parcel locker), the "Nearby Services" advertisement (such as "Convenience store next to the parcel locker, bottled water 1 yuan / bottle") is triggered through GPS location.

[0076] Technical Implementation:

[0077] The ad content is stored in the backend object storage (such as MinIO), and the frontend obtains multimedia resources through CDN acceleration; the display order is sorted by the ad recommendation model (see the "Intelligent Module" section) and adopts the "Click-through Rate Prediction TOP10" strategy.

[0078] (2) Interaction Trigger Module

[0079] User operation process:

[0080] When a user clicks on an ad card, a "Product Details Page" pops up (containing price, stock, and estimated delivery time).

[0081] When a user clicks "Buy Now," a location permission request is triggered (authorization is required for the first use, and will be automatically granted thereafter).

[0082] The system automatically fills in the "delivery location" (based on the current location), which the user can manually adjust (e.g., "delivered to the unit door").

[0083] User confirms order → Payment interface is called (supports WeChat / Alipay / Community Wallet) → Order generated.

[0084] Technical Implementation:

[0085] The interaction logic is implemented using the Vue.js framework, based on the principle of componentization (such as the advertising card component and the order confirmation component).

[0086] The payment interface uses an aggregated payment SDK (such as Ping++) to ensure fund security;

[0087] Location permission management follows Android / iOS privacy guidelines (e.g., iOS 14+ requires requesting "precise location" permission).

[0088] (3) Real-time location acquisition module

[0089] Positioning technology selection:

[0090] A multi-source positioning fusion solution (GPS + base station + WiFi) is adopted to solve the problem of inaccurate indoor / underground positioning.

[0091] GPS: Outdoor scenes, accuracy 5-10 meters (supports BeiDou satellite);

[0092] Base station positioning: In indoor scenarios, triangulation positioning is achieved using mobile operator base stations, with an accuracy of 50-200 meters.

[0093] WiFi positioning: Matches the MAC addresses of nearby WiFi hotspots (based on Gaode / Baidu WiFi database), with an accuracy of 10-30 meters.

[0094] Data fusion algorithm:

[0095] The Kalman filter is used to fuse multi-source localization data, and the formula is as follows:

[0096] Kalman filtering achieves multi-source data fusion through a two-step prediction-update loop: in the prediction phase, the current state and uncertainties are estimated based on the optimal state at the previous moment; in the update phase, the prediction results are corrected using the current observations, iteratively optimizing the positioning accuracy. In this system, multi-source positioning data are input according to scene priority (GPS is prioritized outdoors, and WiFi + base station is prioritized indoors), and the filtering frequency is set to 1Hz (updated once per second).

[0097] State prediction: ;

[0098] Covariance prediction: ;

[0099] Kalman gain: ;

[0100] Status Update: ;

[0101] Covariance update: ;

[0102] The parameters are defined as follows:

[0103] : Optimal position estimate at time k (including a state vector with three dimensions: longitude, latitude, and velocity);

[0104] A: State transition matrix (used to describe the relationship between states and time, for example, a 2×2 matrix in the form of [[1, Δt], [0, 1]], where Δt represents the sampling time interval in seconds; T represents the matrix transpose).

[0105] B: Control input matrix (There is no active control input in this system, so the matrix value is 0);

[0106] : Control input vector (there is no active control in this system, so the vector value is 0);

[0107] State covariance matrix (diagonal elements represent the variance of longitude, latitude, and velocity estimates, reflecting the degree of uncertainty in the estimates of each state quantity);

[0108] Q: Process noise covariance matrix (diagonal elements represent system model uncertainty; values ​​are approximately 0.1 for GPS scenarios and approximately 0.5 for base station scenarios; units are (m). 2 ) / s 2 );

[0109] H: Observation matrix (maps the state vector to the observation space, for example, a 2×2 matrix in the form of [[1, 0], [0,1]], used to extract position-related state variables);

[0110] : The observation vector at time k (containing GPS, base station, or WiFi positioning results, with each positioning source providing a set of longitude and latitude coordinates);

[0111] R: Observation noise covariance matrix (diagonal elements represent the measurement noise of each location source; GPS value is 0.01, base station value is 0.25, WiFi value is 0.1; unit is m). 2 ).

[0112] : Predicted state estimate at time k (based on the predicted value of the optimal state at the previous time, without incorporating current observations);

[0113] : The prediction covariance matrix at time k (reflecting the degree of uncertainty in the predicted state).

[0114] Kalman gain at time k (the weighting of predicted and observed values ​​determines the degree of influence of observed values ​​on prediction correction);

[0115] I: Identity matrix (its dimension is the same as the state vector, used to maintain the validity of matrix operations);

[0116] By using Kalman filtering, the positioning accuracy after fusion can be improved to 2-5 meters (indoors) and 1-3 meters (outdoors).

[0117] (II) Backend System Principles

[0118] The backend adopts a microservice architecture (Spring Cloud Alibaba), consisting of 6 core microservices deployed on a Kubernetes cluster, supporting elastic scaling.

[0119] 1. List of Microservices

[0120]

[0121] 2. Detailed Principles of Core Services

[0122] (1) User Management Service

[0123] User profile building:

[0124] Collect user behavior data (clicks, browsing, purchases, chat history), attribute data (age, gender, address, occupation), and social data (friend list, group chat interaction), and generate user interest tags (such as "fresh food lover", "pet owner", "frequent delivery user") through deep learning models.

[0125] Model Structure: A hybrid model combining BERT, BiLSTM, and Attention is employed. Training data includes user behavior logs (clicks, purchases, browsing time) from the past 6 months and text data (chat logs, product reviews), totaling 5 million samples. Preprocessing involves word segmentation (Jieba segmentation) and stop word removal for the text data, and normalization of the behavior data (e.g., browsing time normalized to 0-1). Model training utilizes the Adam optimizer with an initial learning rate of 1e-5, a batch size of 32, and 50 iterations. Early stopping (stopping if the validation set loss does not decrease for 5 consecutive iterations) prevents overfitting, ultimately achieving a label prediction accuracy of 89.2% on the test set.

[0126] Tag storage: Use Redis to store user tags (key: user_id, value: {"interest tags": ["fresh food", "pets"], "spending power": "medium"}), valid for 7 days (updated regularly).

[0127] Access control:

[0128] It adopts the RBAC (Role-Based Access Control) model, supporting three roles: advertiser (to view ad performance), administrator (to configure the system), and user (to use the APP). Permission verification is implemented through Spring Security + OAuth2.0.

[0129] (2) Advertising management services

[0130] Advertiser onboarding process:

[0131] Advertiser registration → Submit business license / legal representative's ID card (OCR automatic recognition);

[0132] System review (human + AI: using the TextCNN model to detect illegal content. The training data consists of 100,000 labeled advertising texts (including illegal tags such as false advertising and vulgar content). Preprocessing includes word segmentation and word vector conversion (using pre-trained Word2Vec vectors, dimension 100). The model uses 3 layers of convolutional kernels (sizes 3 / 4 / 5), followed by pooling layers and fully connected layers, outputting binary classification results (compliant / illegal). Training uses the SGD optimizer with a learning rate of 0.01, batch size of 64, and 30 iterations. The test set accuracy reaches 92.3%. If false advertising is detected, the content is automatically rejected.)

[0133] Once approved, advertisers log in to the advertising management backend, upload advertising content (images / videos), and set delivery rules (targeting regions, target users, budget).

[0134] Targeted delivery rules:

[0135] It supports geographic targeting (e.g., "Buildings 1-3 of XX Community"), user tag targeting (e.g., "Fresh food lovers aged 25-35"), and time targeting (e.g., "Weekend 18:00-21:00"). Rules are stored in MongoDB (a document-oriented database suitable for flexible rule structures).

[0136] (3) Order Management Service

[0137] Order generation process:

[0138] User payment successful → Payment platform calls back to order management service → Service generates order (including user_id, product information, delivery location, payment amount) → Order message is sent to RocketMQ (topic: order_created) → Delivery scheduling service consumes message.

[0139] Order state machine:

[0140] Order status is divided into pending delivery → in delivery → completed → cancelled. Status transitions are achieved through status patterns (e.g., "pending delivery" → "in delivery" requires triggering the delivery scheduling service).

[0141] Payment callback handling:

[0142] The principle of idempotency is adopted (using the payment serial number as a unique key) to prevent duplicate callbacks; if the callback fails, it will be retried via a scheduled task (Quartz) (up to 3 times).

[0143] (4) Delivery dispatch service (core)

[0144] Delivery resource pool:

[0145] Managing two types of delivery resources:

[0146] Delivery robots: Autonomous mobile robots deployed within the community (the number is adjusted according to the size of the community, such as 50 robots for a community with 100,000 people).

[0147] Delivery personnel: Part-time / full-time delivery personnel (accessed through the "Delivery Personnel APP").

[0148] Scheduling logic:

[0149] Receive order message (order_created);

[0150] Obtain available resources from the delivery resource pool (Robots: Battery level ≥ 20%; Personnel: Current order count ≤ 5);

[0151] Resource allocation was performed using a reinforcement learning (DQN) algorithm. The training data consisted of 1 million samples from delivery logs over the past three months (including order features, resource status, and delivery results). The model input consisted of order location and real-time resource status (location, load, and battery level), and the output was the Q-value for resource allocation. Training employed an ε-greedy strategy (ε decaying from 0.9 to 0.1), with an experience replay pool of 100,000 records, a batch size of 32, an Adam optimizer, a learning rate of 1e-4, and 500 iterations. The final resource allocation efficiency was improved by 23% (compared to the greedy algorithm).

[0152] The order allocation results are sent to the delivery end (robot / personnel APP).

[0153] Real-time scheduling and adjustment:

[0154] When a user's location changes (e.g., from home to the entrance of the community) or delivery resources encounter obstacles (e.g., the robot is blocked), location data is received in real time through Flink streaming processing, triggering scheduling adjustments (e.g., replanning routes or switching delivery resources).

[0155] (5) Location services

[0156] Real-time location synchronization:

[0157] The user / delivery client sends location data to the location service once per second via the WebSocket protocol (which supports bidirectional communication). The service stores the data in PostgreSQL (a spatial database that supports geographic queries) and synchronizes it to Redis (which caches the current location for order tracking).

[0158] Location data format:

[0159] Location data includes the following fields: device ID (e.g., "user_12345"), type (distinguishing between "user", "robot" or "deliveryman"), latitude (e.g., 30.123456), longitude (e.g., 120.654321), and timestamp (e.g., 1690000000000).

[0160] (III) Delivery System Principles

[0161] The delivery end is divided into two categories: delivery robots and delivery personnel apps, which enable order reception, route execution, and real-time feedback.

[0162] 1. Principles of Delivery Robot Systems

[0163] (1) Hardware components (as shown in the attached document) Figure 2 (As shown)

[0164]

[0165] (2) Software System Principles

[0166] Operating system: Raspberry Pi uses Ubuntu 20.04 LTS (with ROS support), and STM32 uses FreeRTOS (real-time motor control).

[0167] Core modules:

[0168] ①SLAM navigation module:

[0169] The community map is constructed using the GMAPping algorithm (built offline and updated regularly), and real-time positioning uses AMCL (Adaptive Monte Carlo Positioning). Combined with LiDAR point cloud data and IMU data, centimeter-level positioning accuracy is achieved.

[0170] Map storage: Use ROS's .pgm format (grayscale image) to store the map, and use .yaml files to record map parameters (resolution, origin coordinates).

[0171] ② Path planning module:

[0172] Algorithm A is used for global path planning (based on the community map), and Dynamic Window Method (DWA) is used for local obstacle avoidance (handling dynamic obstacles such as pedestrians and vehicles).

[0173] DWA Algorithm: By simulating the robot's trajectory over the next 5 seconds (linear velocity v∈[0,1.5]m / s, angular velocity ω∈[-π / 2, π / 2]rad / s), it calculates the evaluation function (target direction, distance from obstacles, velocity) for each trajectory and selects the optimal trajectory. Evaluation function formula:

[0174] score = α·heading + β·clearance + γ·velocity

[0175] The definitions and technical principles of each parameter are as follows:

[0176] heading: The directional deviation between the trajectory endpoint and the target point (calculated by the arctangent function, the angle between the trajectory tangent and the target direction is normalized to 0-1);

[0177] clearance: the minimum distance from the obstacle on the trajectory (calculated using lidar point cloud data and normalized to 0-1);

[0178] velocity: The average velocity of the trajectory (normalized to 0-1 based on the robot's current velocity and acceleration limits).

[0179] α, β, γ: Weighting coefficients (values ​​range from 0 to 1, the sum of the three is 1, dynamically adjusted: in densely populated pedestrian areas, the value of β is increased to 0.4 to prioritize safety, and in open areas, the value of γ is increased to 0.5 to prioritize efficiency).

[0180] Algorithm implementation flow: 1) Sampling velocity space (linear velocity, angular velocity); 2) Trajectory prediction (based on kinematic model); 3) Evaluation function scoring; 4) Selecting the optimal trajectory for execution. Integrated through the ROS navigation stack, with a real-time requirement of 10Hz (updated every 100ms).

[0181] ③ Obstacle avoidance module:

[0182] The algorithm fuses LiDAR point clouds (to detect distant obstacles, such as vehicles 10 meters away) and camera images (to detect near-field dynamic obstacles, such as pedestrians), uses a YOLOv8 model to detect obstacle types (pedestrians, vehicles, roadblocks), uses the KCF tracking algorithm to track dynamic obstacles (updating their position and velocity), and then inputs the obstacle information into the DWA algorithm to adjust the route.

[0183] ④ Communication module:

[0184] It communicates with the backend delivery scheduling service via the MQTT protocol (using EMQ X broker) to send / receive order information and location data (once per second).

[0185] 2. Principles of the Delivery Personnel App

[0186] Functional modules:

[0187] Order Receiving: Receive delivery tasks in real time (push notifications) and display order details (products, delivery location, user notes);

[0188] Route navigation: Integrates with Amap SDK to display the optimal delivery route (walking / cycling) and supports real-time traffic avoidance (such as congested road sections).

[0189] Location reporting: Automatically reports the current location (once per second), and synchronizes it to the order tracking page on the backend;

[0190] Error handling: Supports reporting errors (such as damaged goods or the user not being home) and triggering backend manual intervention (such as contacting the user for re-delivery).

[0191] Technology stack:

[0192] It uses the Flutter framework (compatible with Android / iOS), the map function uses the Amap SDK, and communication uses WebSocket (synchronized with location services).

[0193] (iv) Intelligent module principle (core optimization component)

[0194] The intelligent module is the "brain" of the system, responsible for user profile building, accurate ad recommendation, delivery scheduling optimization, and location prediction, and adopts an "offline training + real-time inference" mode.

[0195] 1. User profiling model: BERT + BiLSTM + Attention

[0196] Model objective: Extract interest tags from multi-source user data with an accuracy of ≥90%.

[0197] Input data:

[0198] Behavioral data: User click / browse / purchase records (sequence data, such as ["clicked fresh produce ad", "purchased milk", "browsed pet supplies"]);

[0199] Text data: Chat logs (e.g., "I want to make braised pork today");

[0200] Attribute data: age, gender, address (structured data).

[0201] Model structure:

[0202] BERT layer: processes chat log text and outputs semantic vectors (768 dimensions);

[0203] BiLSTM layer: processes behavioral sequence data (such as click records) and outputs bidirectional hidden state (256 dimensions × 2 = 512 dimensions).

[0204] Attention layer: Weights the hidden states of the BiLSTM output (highlighting important actions, such as "buying" being weighted more than "browsing");

[0205] Fusion layer: Fusion of BERT semantic vectors, Attention output, and attribute data (after normalization) (768+512+10=1290 dimensions);

[0206] Output layer: Fully connected layer + Softmax, outputting the probability distribution of interest tags (e.g., "fresh food" 0.85, "pets" 0.70, "housekeeping" 0.20).

[0207] Training and Deployment:

[0208] Offline training: TensorFlow 2.x framework was used, with a dataset of historical data from 100,000 users (labeled with interest tags). Preprocessing stage: Text data was segmented using Jieba, stop words were removed, and then converted to WordPiece encoding (BERT input requirement); behavioral sequence data was sampled using a sliding window method (window size 10) and time-decay weighted (recent behaviors had higher weights); attribute data was One-Hot encoded (e.g., gender) or normalized (e.g., age mapped to 0-1). Model training used the Adam optimizer (initial learning rate 0.001, decaying by 10% every 10 epochs), with the loss function being multi-label cross-entropy, a batch size of 64, and 80 iterations. Hyperparameters were adjusted using 5-fold cross-validation, ultimately achieving a label prediction accuracy of 91.3% and an F1-score of 0.88 on the test set.

[0209] Real-time inference: The model is optimized using TensorRT (FP16 precision quantization, 3x faster inference speed), deployed on a GPU server (e.g., NVIDIA T4), and provided via a RESTful API through TensorFlow Serving. Inference flow: Receive real-time user data → Preprocessing (consistent with the training phase) → Model prediction → Output Top 5 interest tags, with an overall latency ≤100ms. Incremental training is performed daily at 2 AM (updating the day's user data) to ensure tag timeliness.

[0210] 2. Ad Recommendation Model: DeepFM

[0211] Model objective: Predict the probability of users clicking on ads and improve ad click-through rate by ≥30% (compared to random ad delivery).

[0212] Input features:

[0213] User characteristics: interest tags (e.g., "fresh food enthusiast", based on the probability distribution output by the user profile model), spending power (e.g., "medium", obtained by clustering based on historical order amounts), and number of clicks in the last 7 days (continuous feature, normalized to 0-1).

[0214] Advertising features: Category (e.g., "fresh produce", Embedding dimension 64), Price (continuous feature, log normalized), Merchant rating (1-5 points, normalized to 0-1), Distance from user location (meters, segmented features: [0,500) meters, [500,1000) meters, ≥1000 meters);

[0215] Contextual features: current time (hourly segment features: [6-9), [9-12), [12-18), [18-22), [22-6)), weather (classification features: sunny, cloudy, rainy, snowy, embedding dimension 32);

[0216] Model structure:

[0217] DeepFM combines a Wide layer (capturing low-order feature interactions, such as "fresh food enthusiast" × "distance ≤ 500 meters", achieved through cross-product transformation) and a Deep layer (capturing high-order feature interactions, such as "fresh food enthusiast" × "rainy day" × "6-10 pm"). The Wide layer takes the original features and cross features as input and outputs 1D weights; the Deep layer takes the embedding vectors of all features as input (total dimension 64 + 32 + ... = 256), and outputs hidden vectors through a 3-layer fully connected network (256 → 128 → 64); finally, the outputs of the Wide and Deep layers are concatenated, and the click probability is obtained through the Sigmoid activation function.

[0218] Training and Deployment:

[0219] Offline training: PyTorch framework was used, with a dataset of 1 million ad click records (positive samples: clicks; negative samples: no clicks, using a Hard Negative Mining strategy, with a negative-to-positive sample ratio of 3:1). Feature preprocessing: categorical features were mapped through an embedding layer (dimensions 8-64, adaptively based on feature cardinality), and continuous features were normalized using Z-Score. The optimizer was AdamW (learning rate 5e-4, weight decay 1e-5), the loss function was binary cross-entropy, and cosine annealing (T_max=100) was used for learning rate scheduling. Early stopping was used during training (training stopped if the validation set AUC showed no improvement for 5 consecutive rounds). The final test set AUC reached 0.89, and the LogLoss reached 0.32.

[0220] Real-time inference: Deployed on a Flink cluster (using Flink ML), supporting streaming inference (the recommendation list is updated every 10 seconds).

[0221] 3. Delivery scheduling model: Reinforcement learning (DQN)

[0222] Model objective: Optimize delivery resource allocation and reduce total delivery time by ≥20% (compared to greedy algorithms).

[0223] Problem modeling:

[0224] State space: ,in This represents the real-time status of the i-th delivery resource (including location coordinates (x, y), remaining battery power / current number of orders, and load). The attributes of the j-th order to be delivered are (location coordinates (x, y), priority (level 1-5), and estimated delivery time), and t represents the current time (accurate to the minute).

[0225] Action space:

[0226] ,

[0227] in This represents the decision to assign an order to a specific delivery resource (e.g., "order o3 is assigned to robot r2"), where k is the total number of possible assignment combinations (limited by resource load).

[0228] Reward function:

[0229] ,

[0230] Where Δt is the increase in total delivery time (minutes) caused by this allocation, Δd is the increase in total delivery distance (meters), and Δc is the increase in total cost (yuan, including robot energy consumption and personnel commission), α=0.1 (distance weight) and β=0.05 (cost weight), which are determined by grid search method.

[0231] a. Observe the current state (Includes all resources and order information);

[0232] b. Use an ε-greedy strategy to select actions. : Randomly select an action with probability ε (exploration), and select the action with the largest current Q value with probability 1-ε (exploitation). ε is initially 0.9 and decreases by 0.01 every 1000 steps until it reaches 0.1;

[0233] c. Perform the action Rewards based on observation of environmental feedback and new status ;

[0234] d. Empirical samples Store in the experience replay pool (capacity) );

[0235] e. When the number of samples in the replay pool is ≥1000, randomly sample 32 samples (batch size);

[0236] f. Calculate the target Q value: Where γ = 0.9 (discount factor), and Q_target is the target network output;

[0237] g. Calculate the current network Q-value Through the MSE loss function

[0238] Update current network parameters;

[0239] in This indicates the current state predicted by the network. Next action Q value; The target Q-value is represented by N (calculated from the maximum Q-value of the reward and future state); N represents the batch size (32 in this case), which is the number of empirical samples used for a single parameter update; the MSE loss function is used to measure the difference between the predicted Q-value and the target Q-value, and the loss is minimized through backpropagation to update the network parameters.

[0240] h. Copy the current network parameters to the target network every 1000 steps (soft update, τ=0.01);

[0241] Problem Description: Given M orders (O_1, O_2, ..., O_M) and N delivery robots (R_1, R_2, ..., R_N), each robot has a maximum load (e.g., 5 orders), find the optimal order sequence for each robot to minimize the total delivery time.

[0242] Fitness function: Where T is the total delivery time for all orders (minutes), D is the total delivery distance (kilometers), C is the total energy cost (yuan), α=0.2 (distance weight), β=0.1 (cost weight). A larger function value indicates a better delivery plan.

[0243] Training Process: ① Population Initialization: Randomly generate 100 initial delivery schemes (chromosomes), each chromosome being an order allocation sequence (e.g., [O1→R1, O2→R1, O2→R2]), ensuring each robot's load is ≤5 orders; ② Selection Operation: Use tournament selection (tournament size=3) to select the top 30% of chromosomes in fitness from the population; ③ Crossover Operation: Two-point crossover (crossover probability 0.8), randomly select two crossover points to exchange chromosome segments; ④ Mutation Operation: Exchange mutation (mutation probability 0.05), randomly exchange the positions of two orders in the chromosome; ⑤ Elite Retention: Retain the best 5% of chromosomes from each generation to directly enter the next generation; ⑥ Termination Condition: Stop when the fitness improvement is ≤1% after 100 iterations or 10 consecutive generations, and output the optimal delivery scheme;

[0244] Model structure:

[0245] We employ DQN (Deep Q-Network), use CNN to extract state features (such as heatmaps of resource location distribution), use fully connected layers to output Q-values ​​(the expected reward for each action), and select the action with the largest Q-value (assigning orders).

[0246] Training process:

[0247] Initialize the experience replay pool (capacity 10) 8 ,storage sample);

[0248] Initialize the DQN network (the current network has the same structure as the target network: input layer (state dimension 64) → convolutional layer (16 3×3 convolutional kernels) → fully connected layer (128→64) → output layer (Q-value dimension = number of actions), the target network parameters are initially the same as the current network).

[0249] For each episode:

[0250] episode: refers to a complete delivery scheduling cycle (from receiving orders to be assigned to the completion of delivery of all orders);

[0251] a. Observe the current state ( );

[0252] b. Select actions using an ε-greedy strategy (ε is the exploration rate, initially 0.9, which decreases linearly by 0.1 every 1000 steps and eventually stabilizes at 0.1, balancing the exploration of new actions with the utilization of known optimal actions.)

[0253] c. Perform the action ( ), and receive a reward ( ) and the next state ( );

[0254] d. ( Stored in the experience replay pool;

[0255] e. Randomly sample a batch (32 records) from the experience replay pool;

[0256] batch: Batch size, which is the number of empirical samples used for a single model parameter update (32 samples in this case, to balance training efficiency and stability).

[0257] f. Calculate the target Q value: (γ is the discount factor, set to 0.9, representing the current value weight of future rewards; Q_target is the action value output by the target network).

[0258] g. Calculate the current Q value: ;

[0259] h. Update the current network using the MSE loss function: (N is the batch size; Q_current is the predicted Q value of the current network output);

[0260] i. Update the target network parameters every 100 steps (copy the current network parameters).

[0261] deploy:

[0262] The trained DQN model is deployed in the delivery scheduling service to receive order and resource status in real time and output allocation decisions (latency ≤ 50ms).

[0263] 4. Delivery route planning model: Genetic Algorithm (GA)

[0264] Model objective: Optimize the order sequence of delivery robots and reduce the total delivery time by ≥15% (compared to random order).

[0265] Problem Description: Given (M) orders ((O_1, O_2, ..., O_M)) and (N) delivery robots ((R_1, R_2, ..., R_N)), each robot has a maximum load (e.g., 5 orders), find the optimal order sequence for each robot to minimize the total delivery time.

[0266] Technical principle: Based on the theory of biological evolution, the order sequence is iteratively optimized by simulating the processes of natural selection, crossover, and mutation. The population size is set at 100 (to balance diversity and computational efficiency), and 20% of elite individuals are retained in each generation to directly enter the next generation.

[0267] Genetic Algorithm Principle:

[0268] Chromosome encoding: Each chromosome is an integer encoding of the order sequence (e.g., "O3→O1→O5" is represented as [3,1,5]), with a length equal to the robot's load capacity (≤5), ensuring that orders are not duplicated;

[0269] Population initialization: Randomly generate 100 chromosomes (initial population);

[0270] Fitness function: f = 1 / (T + αD + βC), where T is the total delivery time (minutes) for the robot to complete all orders, D is the total travel distance (kilometers), C is the energy cost (yuan), α = 0.2 (distance weight), β = 0.1 (cost weight), and the weight values ​​are determined by the analytic hierarchy process.

[0271] Selection operation: The roulette wheel selection method is used. The probability of an individual being selected is P_i = f_i / Σf_j (f_i is the fitness of the i-th chromosome, and Σf_j is the total fitness of the population). Two parent chromosomes are selected each time, and a total of 50 pairs are selected.

[0272] Crossover operation: Two-point crossover is used. Two crossover points are randomly selected (such as positions 2 and 4), and the middle segments of the parent chromosomes are exchanged. The crossover probability is 0.8 (empirical value, verified by the test set) to ensure that there are no duplicates in the order sequence after crossover.

[0273] Mutation operation: Exchange mutation is used, and two positions in the chromosome are randomly selected to exchange the order numbers. The mutation probability is 0.05 (to avoid destroying good genes, determined through sensitivity analysis).

[0274] Termination condition: 100 iterations or the fitness no longer improves (fitness change ≤1% over 10 consecutive iterations).

[0275] Example:

[0276] Suppose robot R1 has 3 orders (O1: Building 1 of the community, O2: Building 3 of the community, O3: Building 5 of the community), and the initial chromosome is the order sequence O1→O2→O3, with a total delivery time of 20 minutes. After optimization by genetic algorithm, the chromosome is adjusted to O1→O2→O2, and the total delivery time is shortened to 15 minutes (a reduction of 25%).

[0277] Input data: The user's location sequence over the past 5 minutes, including latitude (lat) and longitude (lon) coordinates for 5 time steps (coordinates for minute 1: (lat1, lon1), coordinates for minute 2: (lat2, lon2), up to coordinates for minute 5: (lat5, lon5));

[0278] Output data: The user's location coordinates for the next minute (lat6, lon6), where lat represents latitude, lon represents longitude, and the subscript number represents the time step (1-5 for historical data, 6 for predicted data).

[0279] 5. Location prediction model: LSTM

[0280] Model objective: To predict user location changes during delivery (e.g., from home to the entrance of the community) with an accuracy of ≥85%, and to adjust delivery routes in advance.

[0281] Input data: The user's location sequence over the past 5 minutes, containing geographic coordinates for 5 consecutive time steps. Each coordinate consists of latitude (lat), longitude (lon), specifically (minute 1 (lat1, lon1), minute 2 (lat2, lon2), ..., minute 5 (lat2, lon2)).

[0282] Output data: The user's geographic coordinates for the next minute (lat6, lon6), where lat represents latitude (in degrees), lon represents longitude (in degrees), and the subscript 6 represents the 6th minute of the prediction.

[0283] Technical principle: LSTM (Long Short-Term Memory) solves the gradient vanishing problem of traditional RNNs through gating mechanisms (input gate, forget gate, output gate), effectively capturing the temporal dependencies of location sequences. The model input is a standardized sequence of location coordinates (latitude and longitude are standardized to [-1,1] using Z-score), and the output is the predicted coordinates after inverse standardization.

[0284] Model structure: Input layer (feature dimension 2×5=10) → LSTM layer (64 hidden units, return sequence=False) → Fully connected layer (32 neurons, ReLU activation) → Output layer (2 neurons, linear activation, corresponding to predicting latitude and longitude).

[0285] Training process:

[0286] ① Dataset: Collected user location logs for 3 months, totaling 500,000 samples (each log contains 5 minutes of historical location + 1 minute of real location), divided into training set and validation set in an 8:2 ratio;

[0287] ② Preprocessing: Remove outliers (such as coordinates outside the community area), fill missing values ​​with linear interpolation, and align time series (with a uniform sampling interval of 1 minute).

[0288] ③ Training parameters: Optimizer Adam (learning rate 1e-3, β1=0.9, β2=0.999), loss function MSE (mean squared error), batch size 32, 100 iterations, early stopping method (stop if the validation set loss does not decrease for 5 consecutive iterations).

[0289] ④ Evaluation metrics: Prediction accuracy is measured by mean absolute error (MAE), which requires MAE ≤ 5 meters (the tolerance range for positioning error in community scenarios). The model achieved a final MAE of 3.2 meters on the validation set, meeting the accuracy requirement of ≥ 85%.

[0290] Deployment: The trained LSTM model is exported in ONNX format and deployed on edge computing nodes (such as community servers) to receive user location stream data in real time and output prediction results every minute with a latency of ≤200ms, which is used by the delivery route planning module to dynamically adjust the path.

[0291] As attached Figure 3 As shown, a community app advertising interaction and instant delivery method, implemented based on the aforementioned system, includes the following steps:

[0292] S1. Push contextualized advertisements to users through the front-end system and obtain user interaction behavior;

[0293] S2. In response to the user's order placement, generate an order and obtain the user's real-time location;

[0294] S3. Through the delivery scheduling service of the backend system, delivery resources are allocated based on the reinforcement learning model, and the delivery route is optimized using the genetic algorithm;

[0295] S4. Send the order information to the delivery system, and the delivery robot or delivery personnel will carry out the delivery task;

[0296] S5. During the delivery process, the location of the user and the delivery terminal are synchronized in real time through location services, and the delivery progress is displayed on the front-end system.

[0297] The specific application steps are as follows:

[0298] 1. User interaction: When a user opens the community APP, the system pushes contextualized advertisements based on the user profile; when the user clicks on the advertisement to enter the details page, after confirming the order, the APP automatically obtains the real-time location and submits the order.

[0299] 2. Backend processing: The backend service cluster receives orders, calls the delivery scheduling model (DQN + genetic algorithm) to allocate the optimal delivery resources, and synchronizes order information to the delivery end.

[0300] 3. Delivery execution: Delivery robots / personnel receive orders, the route planning model generates the shortest path, the route is dynamically adjusted in conjunction with the location prediction model, and the status is fed back to the backend after the delivery is completed.

[0301] 4. Real-time synchronization: During the delivery process, the location service updates the location of the user and the delivery terminal every second, and the APP displays order tracking information in real time, allowing users to check the delivery progress at any time.

[0302] This system has the following technical advantages:

[0303] 1. Advertising Effectiveness: Keyword matching accuracy ≥ 80%, user click-through rate 3 times higher than the traditional broad-based approach, and merchant marketing costs reduced by 40%. Assuming the traditional click-through rate is 1.5%, the 3-fold increase results in 4.5%; if the merchant's original marketing cost was 1000 yuan, the 40% reduction results in 600 yuan.

[0304] 2. Delivery Efficiency: Order allocation delay ≤50ms, total delivery time reduced by 25%, peak-hour order backlog rate reduced from 15% to below 3%. For example, if the original delivery time was 20 minutes, it will be reduced to 15 minutes after a 25% reduction; for a peak-hour order backlog of 100 orders, the original backlog of 15 orders will be reduced to ≤3 orders.

[0305] 3. Positioning accuracy: Multi-source fusion Kalman filtering controls indoor positioning error to 2-5 meters and outdoor accuracy to 1-3 meters, with a position prediction accuracy of ≥85%.

[0306] 4. User Experience: Order tracking is real-time at 1Hz, delivery time estimation accuracy is ≥90%, and user satisfaction is improved to 92%. For example, out of 100 delivery time estimations, the accuracy rate is no less than 90 times; out of 100 users, 92 users expressed satisfaction.

[0307] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 4 As shown, optionally, electronic device 410 may include a first processor 2001.

[0308] Optionally, the electronic device 410 may also include a memory 2002 and a transceiver 2003.

[0309] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0310] The following is combined with Figure 4 A detailed description of each component of electronic device 410 is provided below:

[0311] The first processor 2001 is the control center of the electronic device 410. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0312] Optionally, the first processor 2001 can perform various functions of the electronic device 410 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0313] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 4 CPU0 and CPU1 are shown in the diagram.

[0314] In a specific implementation, as one example, the electronic device 410 may also include multiple processors, for example... Figure 4 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0315] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0316] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently and be connected via the interface circuit of the electronic device 410. Figure 4 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0317] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0318] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 4 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0319] Optionally, the transceiver 2003 can be integrated with the first processor 2001, or it can exist independently and be connected via the interface circuit of the electronic device 410. Figure 4 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0320] It should be noted that, Figure 4 The structure of the electronic device 410 shown does not constitute a limitation on the router. Actual knowledge structure identification devices may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0321] Furthermore, the technical effects of the electronic device 410 can be referenced from the technical effects of the community APP advertising interaction and instant delivery method described in the above method embodiments, and will not be repeated here.

[0322] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0323] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0324] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0325] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0326] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0327] It should be understood that, in various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0328] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0329] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, systems, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0330] In the embodiments provided by this invention, it should be understood that the disclosed devices, systems, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0331] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0332] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0333] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0334] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A community APP advertising interaction and instant delivery system, characterized in that, include: The front-end system runs on the user's mobile terminal and is used for advertising display, triggering user interaction, real-time location acquisition, and order tracking. The backend system adopts a microservice architecture, including user management service, advertising management service, order management service, delivery scheduling service and location service, which are used to process user data, advertising, order generation and scheduling, and real-time location synchronization; The delivery system, including delivery robots and / or delivery personnel apps, is used to receive and execute delivery tasks; The intelligent module communicates and connects with the backend system to perform user profile building, advertising recommendation, delivery scheduling optimization, and location prediction.

2. The system according to claim 1, characterized in that, The front-end system includes: The ad display module is used to trigger and display ads based on user profile, keyword matching of chat content, or user geolocation. The interaction trigger module is used to respond to user actions, generate orders, and call the payment interface; The real-time location acquisition module adopts a multi-source positioning fusion scheme and improves positioning accuracy through a Kalman filter algorithm.

3. The system according to claim 1, characterized in that, The delivery scheduling service in the backend system is configured as follows: the reinforcement learning DQN algorithm is used to dynamically allocate delivery resources, and the genetic algorithm is used to optimize the order execution order of the delivery robots.

4. The system according to claim 1, characterized in that, The delivery robot includes: The navigation module uses LiDAR, camera, and IMU to achieve SLAM navigation and centimeter-level positioning; The path planning module uses the A* algorithm for global path planning and a dynamic window method for local obstacle avoidance. The communication module synchronizes order and location data with the backend system via the MQTT protocol.

5. The system according to claim 1, characterized in that, The intelligent module includes: The user profiling model uses a hybrid network of BERT+BiLSTM+Attention to extract interest tags from user behavior, text and attribute data. The advertising recommendation model uses the DeepFM model to predict the probability of user clicks. The location prediction model uses an LSTM network to predict future locations based on users' historical location sequences.

6. A method for community APP advertising interaction and instant delivery, implemented based on the system of any one of claims 1-5, characterized in that, Includes the following steps: The system pushes contextualized advertisements to users through the front-end system and collects user interaction behavior. In response to a user placing an order, an order is generated and the user's real-time location is obtained; The backend system's delivery scheduling service allocates delivery resources based on a reinforcement learning model and optimizes delivery routes using a genetic algorithm. The order information is sent to the delivery system, where delivery robots or delivery personnel carry out the delivery task. During the delivery process, the location of the user and the delivery terminal are synchronized in real time through location services, and the delivery progress is displayed on the front-end system.

7. The method according to claim 6, characterized in that, The steps for pushing contextualized advertisements include: Push targeted ads on the homepage based on user profiles; By analyzing user chat content using NLP, related ads are triggered when the keyword matching degree is ≥80%; When a user enters a predefined geofence area, location-related advertisements are triggered.

8. The method according to claim 6, characterized in that, The steps for obtaining the user's real-time location include: Collect multi-source positioning data from GPS, base stations, and WiFi; The Kalman filter algorithm is used to fuse multi-source data to improve indoor and outdoor positioning accuracy.

9. The method according to claim 6, characterized in that, The delivery scheduling steps include: Construct a state space that includes delivery resource status, order attributes, and time information; The DQN model outputs resource allocation actions to minimize total delivery time and cost. Genetic algorithms are used to optimize the order sequence of delivery robots in order to shorten the total delivery distance and time.

10. The method according to claim 6, characterized in that, It also includes a location prediction step: Collect the user's location sequence over the past 5 minutes as input; Predict the user's location coordinates for the next minute using a pre-trained LSTM model; Delivery routes are dynamically adjusted based on forecast results.

Images