An AI-based dynamic interaction method and system for a cycling map

Abstract

The application relates to the technical field of cycling navigation, and discloses an AI-based cycling map dynamic interaction method and system. The method comprises the following steps: a multi-source data sensing terminal collects multiple types of data such as a user terminal and an environment monitoring device, extracts key parameters to generate standardized data, pre-processes emergency data and gives early warnings, and synchronizes privacy data to a cloud database after local encryption training; multi-dimensional user portraits are constructed based on the standardized data, and real-time cycling demand is predicted, a solution strategy is generated when conflicts occur, and the solution strategy is executed by edge nodes in stages; a target edge node is selected, and multi-modal interaction data cache is dynamically configured; user multi-modal instructions are analyzed, health path linkage matching features are generated, personalized paths and hierarchical warning information are determined and output, user feedback is collected, three types of models are updated through an incremental training mechanism, and dynamic interaction closed-loop iterative optimization is realized. The application can improve the adaptability of existing cycling maps to personalized services and user individual and scene demands.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cycling navigation, in particular to an AI-based cycling map dynamic interaction method and system. BACKGROUND

[0002] With the popularization of the concept of green travel, cycling has become an important way for urban short-distance transportation and leisure fitness. As a core auxiliary tool, cycling maps directly affect user safety and experience. Current mainstream cycling maps can achieve basic navigation, but personalized services have shortcomings. They mostly use unified path planning and service logic, without fully considering user fitness levels, riding preferences, equipment performance differences, and team cycling collaboration needs, resulting in low matching degree between recommended content and actual needs. Moreover, for emergency data such as road construction and user physiological abnormalities, they mostly rely on cloud-based centralized processing, lack immediate preprocessing mechanisms, have delayed responses and untimely warnings, and are prone to safety hazards.

[0003] In terms of interactive experience and technology iteration, the multi-modal interaction of existing cycling maps lacks scene-based dynamic adaptation, mostly in fixed voice or visual form, without adjustments according to scene characteristics such as rainy days and high-speed cycling, which may distract users and affect safety. Moreover, they lack a perfect user feedback loop mechanism, cannot convert user satisfaction feedback and cycling behavior data into model optimization power, and lead to difficulty in continuously improving model performance, decreased service accuracy after long-term use, and restricted intelligent development of cycling maps.

[0004] As can be seen from the above, how to improve the adaptation of existing cycling maps' personalized services to individual users and scene needs remains to be solved. SUMMARY

[0005] To improve the adaptation of existing cycling maps' personalized services to individual users and scene needs, the present application provides an AI-based cycling map dynamic interaction method and system.

[0006] In the first aspect, the present application provides an AI-based cycling map dynamic interaction method, which adopts the following technical solution:

[0007] An AI-based cycling map dynamic interaction method, comprising:

[0008] At the multi-source data sensing end, multiple types of data from user terminals, environmental monitoring equipment, road network service platforms, and cycling equipment sensors are collected in real time. User cycling behavior data, physiological state data, regional real-time environmental data, cycling service resource data, and equipment performance parameters are extracted and unified standardized data is generated. Before standardization, emergency data is preprocessed locally at the edge node and an immediate warning is triggered. Privacy data is encrypted and trained locally based on a federated learning mechanism. Then, the standardized data is transmitted to the cloud database through a distributed consistency synchronization mechanism.

[0009] Based on standardized user basic information, historical cycling characteristics, equipment performance parameters, and team member physical fitness data, a multi-dimensional accurate user profile is constructed to predict real-time cycling needs. When there is a mismatch between profile tags and road difficulty or equipment performance, a conflict resolution strategy is dynamically generated. The conflict resolution strategy is executed by a local preprocessing module deployed on edge nodes. The local preprocessing module performs millisecond-level fast response to sudden data and performs batch clustering processing on periodic data.

[0010] After the conflict is resolved, in the hybrid computing architecture where edge nodes and central servers collaborate, the target edge node is selected by combining user geographic location information, real-time physical condition, and equipment adaptation scenario, and a multimodal interactive data cache set is dynamically configured.

[0011] The system acquires user multimodal interaction commands and parses cycling intent information. It combines cached paths and health data to generate multidimensional matching features for health path linkage. It uses a time-series deep learning model to determine personalized cycling paths and graded risk warning information. Simultaneously, it generates and outputs cross-device multimodal interaction content and acquires user feedback information. Based on the feedback information and an incremental training mechanism, it updates the three types of models: user profile, health path linkage, and multimodal interaction, to achieve closed-loop iterative optimization of the dynamic interaction capabilities of the cycling map.

[0012] Optionally, the method further includes performing local lightweight preprocessing on emergency data at edge nodes before standardization and triggering immediate alerts, while simultaneously implementing local encrypted training on privacy data based on a federated learning mechanism:

[0013] The collected data on component wear and tear, historical fault records, and real-time operating status of cycling equipment are integrated into a multi-source data acquisition system. An equipment health assessment model is built at the edge node. The fault risk level of key components of the equipment is calculated through the equipment health assessment model. If the fault risk level reaches the preset fault risk threshold, an immediate warning is triggered, and the equipment fault prediction result is added to the additional field of the standardized data.

[0014] The privacy data is encrypted in layers. The user's core physiological data is classified as Level 1 privacy data, which is used for model training only on the terminal side and no raw data is transmitted. The user's cycling preference data is classified as Level 2 privacy data, which is used for transmitting only the de-identified feature parameters to the cloud database. The publicly available cycling route data is classified as Level 3 data, which is used for the global training of the cloud model.

[0015] After completing emergency data preprocessing, the edge node generates a preprocessing log, which includes data type, processing time, and early warning triggering basis. The log is synchronized to the cloud database for subsequent verification of the effectiveness of the edge preprocessing logic and optimization of preprocessing thresholds.

[0016] Optionally, in the process of dynamically generating conflict resolution strategies, the method further includes:

[0017] Construct team members' complementary physical fitness characteristics and historical collaboration characteristics. Complementary physical fitness characteristics include the matching relationship of advantages in climbing, flat roads, and endurance. Historical collaboration characteristics cover team tacit understanding and emergency response speed.

[0018] To address conflicts between individual and road segment difficulty, historical data of users on road segments of similar difficulty are used to generate an adaptation solution. To address conflicts between team and equipment performance, equipment usage priority is allocated based on complementary physical abilities.

[0019] The team's overall profile is updated according to a preset duration. If a member's physical fitness drops to the physical fitness threshold, the profile label of the corresponding member is adjusted and the cycling needs are re-predicted and the conflict resolution strategy is updated.

[0020] Optionally, the method further includes the following steps in dynamically configuring the multimodal interaction data cache set:

[0021] When selecting target edge nodes, two filtering dimensions are added: the service load status of the node and the data transmission latency. Edge nodes with service load below the preset service load threshold and the shortest data transmission latency are given priority. If there are multiple edge nodes that meet the conditions, the final judgment is made by combining the historical interaction service success rate of the edge nodes.

[0022] When dynamically configuring the multimodal interaction data cache set, the priority of the cached content is adjusted according to the real-time scene characteristics. For example, in the case of cycling in the rain, the tactile vibration encoding data of road water accumulation points and the visual annotation data of water avoidance routes are set to the highest priority. In the case of high-intensity training, the health supply point data and physical fitness warning data are set to the highest priority.

[0023] Establish a dynamic update mechanism for cached data. When new weather data, road condition data, or user physical fitness data are received, automatically clean up expired interactive data in the cache and supplement it with the latest scene adaptation data to ensure the timeliness and accuracy of the cache set.

[0024] Optionally, in the process of acquiring user multimodal interaction commands and parsing cycling intent information, and combining the configured cached paths with health data to generate multi-dimensional matching features for health path linkage, the method further includes:

[0025] When parsing multimodal interaction commands, passive command parsing based on user physiological state is added. When multiple abnormal physiological states are detected, such as the user's heart rate exceeding the preset heart rate safety value and blood oxygen being lower than the preset blood oxygen standard value, passive interaction commands for physiological abnormalities are automatically generated without the user having to actively trigger them.

[0026] When generating multi-dimensional matching features for health routes, two feature dimensions are added: health supply resource density and medical station accessibility. Health supply resource density is the number of water and energy supply points per unit path length, and medical station accessibility is the shortest cycling time from the path to the nearest medical station.

[0027] If the identified cycling intention is professional training, additional training objective adaptation features are incorporated when generating matching features. These features include the frequency of slope changes, road smoothness, and the degree of fit between these features and the user's training objectives, ensuring that the path meets the requirements of professional training.

[0028] Optionally, in the process of achieving closed-loop iterative optimization of the dynamic interaction capabilities of cycling maps by updating user profiles, health route linkage, and multimodal interaction models based on feedback information and incremental training mechanisms, the method also includes:

[0029] Feedback information is divided into three categories: individual-specific feedback, group-wide feedback, and emergency scenario feedback. Individual-specific feedback is only used to optimize the personalized model parameters corresponding to the user, group-wide feedback is used for parameter iteration of the global model, and emergency scenario feedback is used for optimization of specific emergency models.

[0030] To address the feedback data from emergency scenarios, a specialized optimization sub-model is established. This sub-model only receives emergency scenario feedback data related to emergency events, including emergency response time, rescue route adaptability, and emergency interaction effectiveness. Through specialized training, the model's service capabilities in emergency scenarios are improved.

[0031] During incremental training, a protection threshold is set for the model parameters. For model parameters corresponding to the user's core cycling preferences, the adjustment range must not exceed the protection threshold.

[0032] Optionally, after synchronously generating cross-device multimodal interactive content and outputting it to the user terminal and cycling equipment, the method further includes:

[0033] Based on the configured cached data and the generated personalized cycling routes, a digital twin model of the cycling scenario is constructed. The digital twin model of the cycling scenario integrates real-time weather data, road condition data, user physical fitness data, and equipment status data to realize a virtual pre-rehearsal of all elements of the cycling route.

[0034] In the digital twin model of cycling scenarios, the process of physical exertion, probability of risk event triggering, and response effect of interactive services under different cycling schemes are simulated, and a quantitative comparison report of multiple schemes is generated. The report includes multiple core indicators such as peak physical exertion, number of risk points, and interaction response time for each scheme.

[0035] Based on the optimal pre-simulation scheme selected by the user, the multimodal interactive content is finally calibrated. If visual interaction interference is predicted in a certain road segment during the pre-simulation, the interaction form of the corresponding road segment is adjusted in advance, and the visual cues are switched to tactile cues.

[0036] Secondly, this application provides an AI-based dynamic interactive system for cycling maps, employing the following technical solution:

[0037] An AI-based dynamic interactive cycling map system includes:

[0038] The multi-source data perception and synchronization module collects various types of data from user terminals, environmental monitoring equipment, road network service platforms, and cycling equipment sensors in real time at the multi-source data perception end. It extracts user cycling behavior data, physiological state data, regional real-time environmental data, cycling service resource data, and equipment performance parameters, and generates unified standardized data. Before standardization, emergency data is preprocessed locally at the edge node and triggers an immediate warning. Privacy data is encrypted and trained locally based on a federated learning mechanism. The standardized data is then transmitted to the cloud database through a distributed consistency synchronization mechanism.

[0039] The user profiling and conflict resolution module constructs multi-dimensional and accurate user profiles based on standardized user basic information, historical riding characteristics, equipment performance parameters, and team member physical fitness data, and predicts real-time riding needs. When there is a mismatch between profile tags and road difficulty or equipment performance, conflict resolution strategies are dynamically generated. The conflict resolution strategies are executed by the local preprocessing module deployed on the edge node. The local preprocessing module performs millisecond-level fast response to sudden data and performs batch clustering processing on periodic data.

[0040] After conflict resolution, the edge node cache configuration module, in a hybrid computing architecture where edge nodes and central servers collaborate, selects target edge nodes based on user geographic location information, real-time physical condition, and equipment adaptation scenarios, and dynamically configures a multimodal interactive data cache set.

[0041] The interaction iteration and path generation module acquires user multimodal interaction commands and parses cycling intent information. It combines cached paths and health data to generate multi-dimensional matching features for health path linkage. It determines personalized cycling paths and graded risk warning information through a time-series deep learning model. It simultaneously generates and outputs cross-device multimodal interaction content and acquires user feedback information. Based on the feedback information and an incremental training mechanism, it updates the user profile, health path linkage, and multimodal interaction models to achieve closed-loop iterative optimization of the dynamic interaction capabilities of the cycling map.

[0042] Thirdly, this application provides an electronic device that adopts the following technical solution:

[0043] An electronic device includes a processor in which a program for the AI-based dynamic interaction method for cycling maps described in any one of the preceding claims is running.

[0044] Fourthly, this application provides a storage medium, which adopts the following technical solution:

[0045] A storage medium storing a program for the AI-based dynamic interaction method for cycling maps as described in any one of the above.

[0046] In summary, this application includes at least one of the following beneficial technical effects:

[0047] Through a multi-source data collection system, user cycling behavior, physiological state, equipment performance, and team collaboration characteristics are comprehensively integrated to construct a multi-dimensional and accurate user profile. Combining individual physical differences, cycling preferences, equipment performance adaptation needs, and the collaborative characteristics of team cycling, dynamic adaptation conflict resolution strategies are generated to specifically match cycling needs in different scenarios, significantly improving the adaptability of personalized services to individual user and scenario needs. At the same time, through the real-time preprocessing and early warning of emergency data by edge nodes, the response lag problem of traditional cloud-based centralized processing is solved, effectively reducing cycling safety hazards. The full-element pre-simulation of the digital twin model further optimizes the scenario adaptation accuracy of path and interaction solutions.

[0048] By dynamically adjusting the priority and form of multimodal interactions in different scenarios, such as rainy days and high-speed cycling, the system adapts to the interaction needs of different scenarios, reducing cycling interference and ensuring safety. It relies on layered encryption and federated learning mechanisms to protect privacy and data security, and combines a dynamic cache update mechanism to improve interaction response efficiency. By classifying user feedback, a closed-loop iterative system is built to optimize personalized models, global models, and emergency sub-models respectively. At the same time, core preference parameter protection thresholds are set to ensure that model performance continues to improve without deviating from the core needs of users, achieving long-term optimization of personalized adaptation capabilities and overall service effectiveness. Attached Figure Description

[0049] Figure 1This is a flowchart illustrating an AI-based dynamic interaction method for cycling maps according to an exemplary embodiment.

[0050] Figure 2 This is a structural block diagram of an AI-based dynamic interactive cycling map system according to an exemplary embodiment. Detailed Implementation

[0051] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0052] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0053] This application discloses an AI-based dynamic interaction method for cycling maps, referring to... Figure 1 ,include:

[0054] The S100 collects various types of data in real time from user terminals, environmental monitoring equipment, road network service platforms, and cycling equipment sensors at the multi-source data sensing end. It extracts user cycling behavior data, physiological state data, regional real-time environmental data, cycling service resource data, and equipment performance parameters, and generates unified standardized data. Before standardization, emergency data is preprocessed locally at the edge node and triggers an immediate warning. Privacy data is encrypted and trained locally based on a federated learning mechanism. The standardized data is then transmitted to the cloud database through a distributed consistency synchronization mechanism.

[0055] The S100 execution process specifically includes:

[0056] Step 1, Real-time acquisition of multi-source data triggered:

[0057] The multi-source data sensing terminal initiates the data acquisition process, simultaneously connecting to four data sources through a preset data transmission protocol: user terminals (mobile phones, smart bracelets), environmental monitoring equipment (road monitoring, weather sensors), road network service platforms (urban traffic road network systems, cycling service station management platforms), and cycling equipment sensors (handlebar vibration sensors, frame stress sensors, tire pressure sensors). Following a preset acquisition frequency (millisecond-level acquisition of physiological data, second-level acquisition of environmental data, and minute-level acquisition of road network data), it captures various types of raw data in real time, ensuring the timeliness and comprehensiveness of data acquisition.

[0058] Step 2, Key Data Extraction and Preliminary Integration:

[0059] From the collected raw data of various types, five key data categories were accurately separated and extracted using data filtering and feature extraction algorithms:

[0060] User cycling behavior data includes historical cycling routes, changes in cycling speed, stop points, and role division during team cycling.

[0061] Physiological status data: including heart rate, blood oxygen saturation, physical exertion level, muscle fatigue feedback, etc.

[0062] Real-time regional environmental data: including real-time weather (temperature, precipitation, wind speed), road congestion level, road surface smoothness, and user-reported emergency road conditions (construction, obstacles), etc.

[0063] Cycling service resource data: including the location and operational status of water replenishment points, repair stations, rest areas, and medical stations along the route;

[0064] Equipment performance parameters include frame material durability, tire pressure, gear shift status, and component wear.

[0065] After extraction, the data is initially formatted to lay the foundation for subsequent standardization processing.

[0066] Step 3, Lightweight preprocessing and real-time alerts for emergency data at the edge:

[0067] A lightweight data processing engine is deployed at edge nodes (local user terminal, cycling equipment control module) to perform real-time analysis of the extracted key data.

[0068] First, emergency data is filtered out using preset emergency data identification rules (such as heart rate exceeding a preset safety threshold, sudden road construction markings, and equipment malfunction signals).

[0069] Lightweight preprocessing is performed on emergency data, including data noise reduction and key feature extraction (such as extracting the faulty component number and the location coordinates of the sudden road condition).

[0070] Instantly trigger local real-time alerts, with alert methods adapted to data types (physiological abnormalities trigger wristband vibration + voice prompts, abnormal road conditions trigger handlebar beeping + mobile phone pop-up windows, and equipment malfunctions trigger flashing equipment warning lights).

[0071] Non-urgent data is temporarily stored in the edge node cache area, awaiting subsequent batch processing.

[0072] Step 4, Local Encryption Training of Privacy Data (Based on Federated Learning Mechanism):

[0073] For the extracted privacy data, initiate a federated learning encryption process:

[0074] First, the privacy data is classified into layers. The user's core physiological data (heart rate, blood oxygen, health record) is classified as Level 1 privacy data, the user's cycling preferences and habits data is classified as Level 2 privacy data, and the publicly available cycling routes and service resource data is classified as Level 3 data.

[0075] Level 1 privacy data is only deployed locally on the terminal using a lightweight training model, and no raw data is transmitted after feature learning is completed; Level 2 privacy data is processed through an anonymization algorithm (removing personal identification information) and only feature parameters are extracted; Level 3 data does not require encryption.

[0076] Edge nodes transmit only the model training parameters of Level 1 privacy data, the de-identified feature parameters of Level 2 privacy data, and the Level 3 public data to the cloud to ensure user privacy and security.

[0077] Step 5, Data standardization and unification processing:

[0078] For all critical data that has undergone edge preprocessing (urgent data) and encryption (privacy data), a unified standardized operation is performed:

[0079] A data format calibration algorithm is used to convert data from different sources and in different formats (text, numerical, signal) into a unified data format;

[0080] Feature alignment technology can be used to unify the feature dimensions of similar data (such as unifying the speed units collected by different sensors and the road condition level classification standards reported by different platforms).

[0081] Remove redundant information and outliers from the data (such as invalid data falsely reported by sensors) to ensure the accuracy and consistency of the standardized data.

[0082] Step 6, Standardized data synchronization and transmission to the cloud:

[0083] Initiate a distributed consistency synchronization mechanism to transmit the standardized complete data to the cloud database:

[0084] During transmission, a fragmentation transmission strategy is adopted to split large amounts of data into multiple data fragments for parallel transmission, thereby improving transmission efficiency.

[0085] The integrity of each data segment is checked through a data verification mechanism. If transmission loss or error occurs, automatic retransmission is triggered.

[0086] After receiving the data, the cloud database categorizes and stores it according to data type and timestamp, and creates indexes to support subsequent fast queries and retrieval.

[0087] Through a complete process of "comprehensive collection of multi-source data - precise extraction of key data - rapid local response to emergency data - encrypted protection of privacy data - standardized integration of data - reliable synchronization in the cloud", the system not only achieves comprehensive coverage and real-time capture of cycling-related data, but also solves the problem of delayed response in emergency scenarios through edge preprocessing, ensures user privacy and security through federated learning mechanisms, and ensures data consistency and availability through standardized processing. Ultimately, it provides high-quality, secure, and timely data support for subsequent user profile construction, demand prediction, route planning, and multimodal interaction.

[0088] S200 constructs multi-dimensional and accurate user profiles and predicts real-time riding needs based on standardized user basic information, historical riding characteristics, equipment performance parameters, and team member physical fitness data. When there is a mismatch between profile tags and road difficulty or equipment performance, it dynamically generates conflict resolution strategies. The conflict resolution strategies are executed by the local preprocessing module deployed on the edge node. The local preprocessing module performs millisecond-level fast response to sudden data and performs batch clustering processing on periodic data.

[0089] The S200 execution process specifically includes:

[0090] Step 1, Core Input Data Filtering and Feature Extraction:

[0091] From the standardized data transmitted from the S100 to the cloud database, four core input data categories were precisely selected: basic user information, historical riding characteristics, equipment performance parameters, and team member physical fitness data. Feature extraction was then performed on each category of data.

[0092] User basic information: Extract static features such as age, height, weight, health status (no underlying diseases / physical fitness level), and years of cycling experience;

[0093] Historical cycling characteristics: Extract dynamic features such as frequently used cycling distances, preferred route types (flat / uphill / off-road), average cycling speed, longest continuous cycling time, and high-frequency stopping areas;

[0094] Equipment performance parameters: Extract equipment type (mountain bike / road bike / folding bike), core component durability (frame load-bearing capacity, tire grip), range of applicable road conditions, current component wear level, and other compatibility features;

[0095] Team member physical fitness data: Extract each team member's peak physical fitness, endurance limit, climbing advantage, flat road speed advantage, and coordination characteristics such as the degree of cooperation during historical collaborative riding.

[0096] After extraction, the feature data is normalized to eliminate differences in units, providing a unified input dimension for profile construction.

[0097] Step 2: Construct a multi-dimensional and precise user profile in layers:

[0098] Based on the extracted feature data, a multi-model fusion algorithm is used to construct a hierarchical user profile, which includes four core dimensions:

[0099] Individual basic dimensions integrate basic user information and static characteristics to form a basic tag set including "age-weight-fitness level-cycling experience";

[0100] In terms of cycling preferences, historical cycling characteristics are analyzed using clustering algorithms to generate a dynamic tag set of "route type preference - cycling distance preference - speed preference - stop scene preference";

[0101] In terms of equipment compatibility, a set of compatibility tags is constructed by combining equipment performance parameters with matching data from historical riding scenarios, based on "equipment type - compatible road conditions - wear tolerance - usage priority".

[0102] In terms of team collaboration, for team cycling scenarios, the system integrates team members' physical fitness data and historical collaboration characteristics to generate a set of collaboration tags for "team roles (leader / cruiser / closer) - complementary physical fitness directions - collaboration needs".

[0103] The various dimensional tags are interconnected, forming a complete and dynamically updated multi-dimensional user profile.

[0104] Step 3, Real-time Intelligent Prediction of Cycling Demand:

[0105] By combining the constructed multi-dimensional user profiles with real-time scene data (current time, geographical location, real-time weather, road network status), and using time-series prediction models and demand classification algorithms, the system predicts users' immediate cycling needs.

[0106] First, the cycling scenario type is determined by the scene matching algorithm (commuting / leisure sightseeing / professional training / team cycling).

[0107] Based on different scenario types and user profile tags, the core needs are predicted: commuting scenarios predict the needs of "speed, fewer traffic lights, and smooth road conditions", leisure scenarios predict the needs of "beautiful scenery, gentle road conditions, and convenient supplies", professional training scenarios predict the needs of "adaptable slope, distance met, and controllable intensity", and team scenarios predict the needs of "fitted physical fitness of members, convenient collaboration, and safety guarantee".

[0108] The output includes real-time cycling demand prediction results that include scenario type, core requirements, and auxiliary requirements.

[0109] Step 4, Adapt to intelligent conflict recognition:

[0110] The conflict detection engine is activated to verify the match between user profile tags and the difficulty of road sections and equipment performance in the current potential cycling routes, identifying two types of core adaptation conflicts:

[0111] Individual compatibility conflict: When the physical fitness level tag in the user profile is lower than the road difficulty level (e.g., a user with a beginner level of physical fitness is matched with a steep road section), or the equipment compatibility tag is incompatible with the road conditions (e.g., a road bike is matched with an off-road section), it is judged as an individual compatibility conflict.

[0112] Team compatibility conflict: When the physical fitness complementarity tags of team members do not match the difficulty distribution of the route (e.g., a team without climbing advantage is matched with a continuous climbing route), or the team equipment performance tags cannot meet the route requirements (e.g., a folding bike team is matched with a long mountain route), it is judged as a team compatibility conflict.

[0113] After conflict identification, key information such as conflict type, conflict label, and associated features are recorded to provide a basis for strategy generation.

[0114] Step 5, Dynamic conflict resolution strategy generation:

[0115] For the identified types of adaptation conflicts, an algorithm combining rule-based reasoning and case matching is used to dynamically generate targeted conflict resolution strategies:

[0116] Individual adaptation conflict resolution strategy: If the conflict is between physical fitness and road difficulty, a combination strategy of "road replacement (steep slope to gentle slope) - distance fine-tuning - increased supply point density" is generated based on the user's historical riding data of the same difficulty and physical fitness redundancy; if the conflict is between equipment and road conditions, a strategy of "alternative route recommendation (off-road conditions to paved road surface) - equipment usage precautions - road section travel speed suggestion" is generated.

[0117] Team adaptation conflict resolution strategy: If the conflict is between physical fitness complementarity and road section difficulty, generate a strategy of "road section splitting (split long steep slopes into multiple short steep slopes + gentle transition sections) - rest point optimization - dynamic adjustment of member roles"; if the conflict is between equipment and road section, generate a strategy of "equipment usage priority allocation (equipment adapted to the road section is used first) - road section passage order planning - auxiliary equipment recommendation".

[0118] Once the strategy is generated, key elements such as execution steps, associated data, and expected results should be clearly defined to ensure its feasibility.

[0119] Step 6: Conflict resolution strategy is implemented in a tiered manner at the edge:

[0120] The generated conflict resolution strategy is transmitted to the local preprocessing module deployed on the edge node. The module performs hierarchical processing based on the type of data associated with the strategy.

[0121] In response to sudden data (conflicts caused by real-time physical decline, equipment compatibility conflicts caused by sudden road conditions), the module activates a millisecond-level rapid response mechanism to immediately execute emergency handling steps in the strategy (such as real-time switching of alternative routes and push notifications for emergency equipment use) to ensure that conflicts are resolved in a timely manner.

[0122] For periodic data (regular conflicts caused by historical cycling preferences and predicted conflicts caused by team physical fitness trends), the module initiates a batch clustering processing mechanism to classify conflict strategies of the same type and execute them in batches (such as batch adjusting the recommendation priority of sections of the same difficulty and batch updating the equipment allocation rules for team members) to improve processing efficiency.

[0123] Once completed, the module will send the results to the cloud database for subsequent profile updates and strategy optimization.

[0124] Through the entire process of “feature extraction, multi-dimensional profile construction, real-time demand prediction, adaptation conflict identification, dynamic strategy generation, and edge-level execution”, the system not only achieves a precise upgrade of user profiles from “single-dimensional” to “individual + equipment + team”, ensuring the relevance of real-time cycling demand predictions, but also efficiently resolves the compatibility contradictions between profiles and road sections and equipment through targeted conflict resolution strategies and edge-level execution mechanisms. At the same time, it ensures rapid response to sudden conflicts and efficient batch processing of periodic conflicts.

[0125] After the conflict is resolved, the S300, in a hybrid computing architecture that combines edge nodes and central servers, selects target edge nodes by combining user geographic location information, real-time physical condition, and equipment adaptation scenarios, and dynamically configures a multimodal interactive data cache set.

[0126] The S300 execution process specifically includes:

[0127] Step 1, Hybrid Computing Architecture Initialization and Co-readiness:

[0128] Once the conflict is resolved, the hybrid computing architecture startup process, which coordinates edge nodes and the central server, is automatically triggered. The central server first establishes communication links with all candidate edge nodes (covering the edge node cluster around the user's current geographical location) and verifies the online status, service availability, and computing load of each edge node through a heartbeat detection mechanism. At the same time, it loads the user profile, real-time riding needs, and conflict resolution results output by S200, using them as the core reference for architecture collaborative decision-making, ensuring that the architecture has the basic conditions for accurate service from the very beginning.

[0129] Step 2, Acquisition and Validation of Core Decision Data:

[0130] Three types of key decision-making data are extracted from the cloud database, and their validity is verified in real time:

[0131] User geolocation information: Accurate coordinates are obtained through the user terminal's GPS and Beidou positioning modules. The positioning error is checked to see if it is within a preset threshold (e.g., ≤5 meters). If the error exceeds the threshold, a secondary positioning correction is triggered.

[0132] Real-time physical fitness data: Extract dynamic data such as user heart rate, blood oxygen, and physical exertion rate uploaded in real time from edge nodes, compare them with the physical fitness baseline data in S200, and confirm that the data is not outdated (the difference between the data generation time and the current time is ≤30 seconds).

[0133] Equipment Adaptation Scenario Data: Call the equipment performance parameters stored in S100 and the cycling scenario type determined by S200 (such as rainy day cycling, mountain training) to verify whether the matching relationship between equipment parameters and scenario type has been corrected through conflict resolution strategy.

[0134] Data that fails verification will trigger a re-acquisition process until valid data is obtained, after which the process will proceed to the next step.

[0135] Step 3, Multi-dimensional filtering and determination of target edge nodes:

[0136] Based on the three types of core data after verification, a multi-dimensional weighted scoring algorithm is used to filter target edge nodes:

[0137] The first dimension (geographic location adaptation): Taking the user's current coordinates as the center, calculate the physical distance of each candidate edge node. The closer the distance, the higher the weight. Edge nodes with a distance of ≤5 kilometers are selected first to enter the next round.

[0138] The second dimension (physical fitness status adaptation): If the user's real-time physical fitness status is "high consumption" or "fatigue", then increase the weight of low-latency edge nodes (requiring data transmission latency ≤ 100 milliseconds) to reduce service lag caused by transmission delay; if the physical fitness status is "good", then balance the latency and load weights.

[0139] The third dimension (equipment and scenario adaptation): For different equipment types and scenarios, select edge nodes that support corresponding data processing (e.g., smart helmet users need to select nodes that support AR visual data parsing, and mountain bike users need to select nodes that support complex road condition data caching).

[0140] The fourth dimension (service capability adaptation) further considers the current load of candidate nodes (load rate ≤70% is qualified) and historical service success rate (≥95% is qualified). Nodes that pass the screening of the first three dimensions are weighted and scored, and the node with the highest comprehensive score is selected as the target edge node.

[0141] Step 4: Determine the content of the multimodal interaction data cache:

[0142] Based on the S200's predicted real-time cycling demand, the current cycling scenario, and the storage and processing capabilities of the target edge nodes, the core content of the cache set is determined as follows:

[0143] Visual interaction data includes AR annotation data of cycling routes (such as route guidance arrows, risk point highlighting), text and images introducing scenic spots along the way, and real-time road condition image adaptation data, etc.

[0144] Haptic interaction data: including handlebar vibration coding data corresponding to different road conditions (water accumulation, steep slope, obstacles), haptic feedback intensity parameters for emergency warnings, etc.

[0145] Voice interaction data includes route guidance voice scripts, safety prompt voice clips, health replenishment suggestion voice content, etc.

[0146] Supporting data includes location-related data of service resources along the route (water replenishment points, maintenance stations), and early warning threshold data corresponding to the user's physical condition.

[0147] All cached content is associated with user profile tags to ensure that data accurately matches user needs.

[0148] Step 5, Dynamic configuration and priority sorting of cache collections:

[0149] After the target edge node receives the cached content list, it initiates the dynamic configuration process:

[0150] First, adjust the cache priority according to the type of cycling scenario: for rainy cycling scenarios, set "tactile vibration encoding of water accumulation points + visual annotation of water avoidance routes" as the highest priority; for high-intensity training scenarios, set "physical fitness warning data + health supply point association data" as the highest priority; and for team cycling scenarios, set "member location synchronization data + collaborative command interaction data" as the highest priority.

[0151] Then optimize the storage structure according to the data access frequency: store high-frequency access data (such as real-time traffic visual annotations and instant voice guidance) in the high-speed cache area of ​​the edge nodes, and store low-frequency access data (such as alternative route auxiliary information) in the ordinary storage area.

[0152] Finally, set cache expiration rules: automatically mark expiration time based on data timeliness (e.g., weather data is valid for 1 hour, traffic data is valid for 15 minutes) to avoid invalid data occupying storage resources.

[0153] Step 6, Cache synchronization verification and architecture status monitoring:

[0154] The central server synchronizes the configured cache set list to the target edge nodes. Upon receiving the data, the target edge nodes verify the integrity and format correctness of the data one by one.

[0155] If there is missing data or incorrect format, a retransmission request will be automatically sent to the central server, which will trigger the fragmented retransmission mechanism to supplement the data.

[0156] After successful verification, the target edge node returns a "cache ready" response to the central server, and the central server records the current cache configuration details (including node ID, cache content, priority, and expiration rules).

[0157] Simultaneously, real-time monitoring of the architecture status is initiated: continuously monitoring the load changes of target edge nodes, the frequency of cached data access, and the stability of the communication link with the central server. Once an anomaly such as node overload or link interruption occurs, the candidate edge node switching process is immediately triggered.

[0158] Through a complete process of "architectural collaborative initialization - decision data verification - precise target node selection - dynamic configuration of cached content - synchronous verification and status monitoring," a crucial bridge connecting "conflict resolution" and "personalized interaction and path generation" has been built. Its core functions are: to precisely select target edge nodes by combining user geolocation, real-time physical condition, and equipment scenario, thus shortening data transmission distance and latency; to ensure rapid response to high-demand data by dynamically configuring multimodal interaction caches and prioritizing them according to scenario; and to guarantee service stability through architecture status monitoring. Ultimately, this provides low-latency, highly adaptable, and highly available computing and storage support for subsequent S400 multimodal interaction parsing and personalized path generation.

[0159] The S400 acquires user multimodal interaction commands and parses cycling intent information. It combines cached paths and health data to generate multi-dimensional matching features for health path linkage. It determines personalized cycling paths and graded risk warning information through a time-series deep learning model. It simultaneously generates and outputs cross-device multimodal interaction content and acquires user feedback information. Based on the feedback information and an incremental training mechanism, it updates the three types of models: user profile, health path linkage, and multimodal interaction, to achieve closed-loop iterative optimization of the dynamic interaction capabilities of the cycling map.

[0160] The S400 execution process specifically includes:

[0161] Step 1: Comprehensive acquisition and validity verification of multimodal interaction commands:

[0162] The multimodal command acquisition interface is activated to synchronously receive various interactive commands from user terminals and cycling equipment. Command types include active commands (voice commands, gesture operation commands, and touch input commands) and passive commands (triggered commands based on physiological state data). The acquired commands are validated in real time: voice commands are processed to remove environmental interference and verify semantic integrity through noise reduction algorithms; gesture / touch commands are validated for operational logic coherence (e.g., to avoid accidental command touches); and passive commands are validated for the authenticity of trigger data (e.g., to exclude physiological data falsely reported by sensors). Commands that fail validation are re-acquired (e.g., prompting the user to repeat if the voice command is unclear), while valid commands are temporarily stored in the command cache area of ​​the target edge node.

[0163] Step 2, In-depth analysis of cycling intention information:

[0164] The intent parsing model is invoked, and combined with the multimodal interaction data cache set configured by the S300, the user profile generated by the S200, and real-time riding needs, the effective commands are analyzed in depth:

[0165] Active command parsing: Extract core keywords from commands (such as "commuting", "leisure sightseeing" and "avoiding steep slopes"), and combine them with cycling preference tags in user profiles to determine basic cycling intentions;

[0166] Passive command parsing: When abnormal physiological data such as a user's heart rate exceeding the safety threshold or blood oxygen being lower than the standard value are detected, the command is automatically parsed as a "physiological regulation need" intent; when an equipment malfunction warning signal is detected, the command is parsed as an "emergency handling need" intent.

[0167] Intent completion and confirmation: Combine real-time scene data (weather, road network status) to complete intent details (e.g., "commuting" intent is completed as "fast commuting, avoid congested sections"), and the accuracy of intent is confirmed through brief interactive feedback (e.g., voice inquiry "whether to prioritize routes with fewer traffic lights").

[0168] Step 3: Generate multi-dimensional matching features linked to health pathways:

[0169] Path data (slope, length, and road condition type of candidate routes) and health data (real-time user fitness, historical health baseline, and distribution of supplies along the route) are extracted from the S300 cache set and fused to generate multi-dimensional matching features:

[0170] Basic compatibility features: the degree to which the route difficulty matches the user's fitness level, the degree to which the equipment performance matches the road conditions, and the degree to which the route length matches the cycling needs.

[0171] Health-related characteristics: density of health supply resources per unit path length (number of water and energy supply points), cycling time from the path to the nearest medical station (accessibility of medical stations), and balance between estimated physical exertion and real-time physical redundancy of users.

[0172] Scene adaptation features: road surface anti-slip level in rainy weather, route smoothness in high-speed cycling, and ease of team collaboration in team cycling.

[0173] All features are quantized according to a uniform weighting rule to form a standardized feature vector.

[0174] Step 4: Generation of personalized cycling routes and tiered risk warning information:

[0175] The standardized feature vectors are input into a time-series deep learning model, which combines the user profile and immediate needs of the S200 to perform path planning and risk warning calculations.

[0176] Personalized route generation: Through a multi-objective optimization algorithm, while satisfying the riding intention, the route speed, health adaptability, and safety factor are balanced to output one optimal main route and 2-3 alternative routes. The route information includes detailed route coordinates, estimated riding time, and the location of key nodes (supply points, rest points).

[0177] Tiered risk warning judgment: Potential risks in the route are marked according to risk level (high, medium, low), high risk (such as steep slopes, construction sections, and sections with excessive physiological load), medium risk (such as bumpy road surface and sections with high traffic volume), and low risk (such as smooth paved road surface). Corresponding warning methods are matched for different risk levels (high risk triggers multi-device linkage warning, medium risk triggers voice + tactile warning, and low risk only visual marking).

[0178] After generation, the path and warning information will be synchronized to the central server and the target edge nodes.

[0179] Step 5: Customized generation and output of cross-device multimodal interactive content:

[0180] Based on the generated path, warning information, and S300 cache priority configuration, customized cross-device interactive content is generated:

[0181] Smart helmet: Outputs AR visual annotations (path guidance arrows, risk point highlighting, key node prompts), with brightness automatically adjusted according to ambient light;

[0182] Smart handlebars: Output tactile vibration codes (one short vibration for left turns, two short vibrations for right turns, and a long vibration for high-risk turns), with vibration intensity adapted to the user's physical condition (intensity reduced when fatigued).

[0183] User terminal (mobile phone / wristband): Outputs voice guidance (segmented route announcements, risk warnings), pop-up notifications (reminders of supply points, physical fitness alerts), and text-based route details;

[0184] Cycling Equipment Control Module: Outputs equipment compatibility prompts (such as suggestions for gear shifting when triggering on steep slopes);

[0185] When outputting data, the principle of "scene adaptation" should be followed. For example, AR brightness should be reduced when riding in the rain, and visual interaction should be reduced when riding at high speed to avoid distracting the user.

[0186] Step 6: Comprehensive collection and categorization of user feedback information:

[0187] Deploy a feedback collection mechanism to simultaneously acquire both proactive and reactive user feedback:

[0188] Proactive feedback: collected through user terminal rating portals (path satisfaction, interaction experience rating), voice feedback, text input, etc.

[0189] Passive feedback: Collect user behavior feedback (such as whether the user deviates from the recommended route, whether a certain type of warning is turned off, and whether the dwell time is consistent with the plan) and physiological status feedback (such as whether the physical exertion during the route cycling exceeds the estimate).

[0190] The feedback information is categorized by type: route adaptation feedback, interaction experience feedback, risk warning feedback, and emergency handling feedback. At the same time, it is associated with the corresponding cycling scenarios and user profile tags to form a structured feedback dataset.

[0191] Step 7: Update the three core models based on the incremental training mechanism:

[0192] Initiate the incremental training process by inputting the structured feedback dataset into the user profile model, health path linkage model, and multimodal interaction model:

[0193] User profile model updates: Individual specific feedback is used to fine-tune the weight of the user's profile tags (e.g., if a user reports a preference for off-road sections, increase the weight of the "off-road condition preference" tag), while group common feedback is used to optimize the global profile feature dimensions (e.g., if most users report that the adaptation tag for a certain type of equipment is inaccurate, adjust the feature extraction rules for that type of tag).

[0194] Health path linkage model update: Path adaptation feedback and health association feedback are used to optimize the matching feature weights (such as increasing the feature weight of "supply resource density" if users report that the supply point reminder is not timely), and emergency handling feedback is used to strengthen the path planning logic in high-risk scenarios;

[0195] Multimodal interaction model update: Interaction experience feedback is used to adjust the interaction form across devices (e.g., if the user feedback is that the tactile vibration is too strong, the vibration intensity threshold of the corresponding scene is reduced), and scene adaptation feedback is used to optimize the interaction priority configuration of different scenes.

[0196] During training, a core parameter protection threshold is set so that the adjustment range of model parameters corresponding to the user's core cycling preferences does not exceed the threshold, thus avoiding common feedback from overriding personalized needs.

[0197] Step 8: Implementation and effect verification of closed-loop iterative optimization:

[0198] The updated three types of models are synchronized to the entire technical architecture (multi-source data sensing end, edge nodes, and central server) and applied to the next round of data collection, profile building, node selection, and path planning. The central server records service metrics (path matching accuracy, interaction satisfaction, and early warning response efficiency) before and after the model update and verifies the optimization effect through comparative analysis. If the metrics do not meet the preset standards (such as a path matching accuracy improvement of less than 5%), a second incremental training is triggered, and iteration continues until the requirements are met, forming a closed-loop optimization mechanism of "collection-analysis-generation-feedback-update".

[0199] Through a complete workflow encompassing "command acquisition, intent parsing, feature generation, path and warning generation, interactive output, feedback acquisition, model update, and closed-loop optimization," and leveraging the low-latency caching support of the S300, the core implementation of dynamic interaction on cycling maps has been achieved. Its core functions are: to accurately address user-specific needs and scenario adaptation requirements through multimodal command parsing and multi-dimensional matching feature generation, outputting customized paths and cross-device interactive content; to drive continuous iteration of three core models through active and passive feedback acquisition and incremental training mechanisms, ensuring service capabilities are continuously optimized with user usage; and to enhance cycling safety and user experience through tiered risk warnings and scenario-based interaction adaptation, ultimately resolving the technical issues of insufficient personalized adaptation of existing cycling maps, lack of scenario-based adjustments for multimodal interaction, and inability to continuously optimize models, thus completing the closed-loop implementation of the entire solution.

[0200] Based on the solutions in the above embodiments of this application, the following is an illustration with reference to a case study: Personalized adaptation implementation in the mountain training scenario of a professional cycling team.

[0201] Suppose a professional cycling team plans to conduct mountain bike training. The team members include a leader with intermediate fitness levels, beginners with basic fitness levels, and experienced riders with lightly worn mountain bikes. The cycling scenario is a complex mountain environment with many steep slopes and the possibility of afternoon showers. The solution utilizes S100 multi-source data acquisition to comprehensively capture each member's physiological state data, such as heart rate and endurance limits; equipment performance parameters, such as bike frame load capacity and tire grip; and environmental and service data, such as real-time road conditions, weather forecasts, and the distribution of aid stations along the route. Simultaneously, it uses edge nodes to preprocess emergency data such as "sudden increases in a beginner's heart rate" and trigger vibration warnings. Based on federated learning, it locally encrypts members' health records and other private data to ensure data security and real-time response. The S200 extracts core features from standardized data to build an individual profile for each member, including "physical fitness level, off-road preference, and equipment wear and tear tolerance," as well as a team collaboration profile of "leader, cruiser, and finisher." It predicts the immediate needs for "steep slope adaptation, equipment protection, and team collaboration," while identifying two types of problems: "conflict between novice's physical fitness and the difficulty of continuous steep slopes" and "conflict between experienced rider's equipment wear and tear and adaptability to gravel roads." It dynamically generates conflict resolution strategies such as "splitting steep slopes into short slopes and gentle transition sections" and "prioritizing alternative routes with a high proportion of paved roads for experienced riders." These strategies are quickly executed by edge nodes to achieve precise adaptation between individuals, equipment, and road sections.

[0202] In terms of dynamic scene adaptation, the S300 combines the team's real-time geographical location, the state of beginners' physical exhaustion, and the scene requirements of riding in the rain to select the nearest, low-latency target edge nodes that support AR visual analysis, dynamically configure the cache set, and set "steep slope tactile vibration encoding, rainy day water avoidance route visual annotation, and supply point location association data along the way" as the highest priority to ensure rapid response to high demand data. The S400 receives the team's voice command to "prioritize completing training goals and ensure the safety of beginners," analyzes the core intent of professional training, and combines cached path and health data to generate multi-dimensional matching features including "slope suitability, physical exertion balance, and rain-resistant slip resistance." Through a temporal deep learning model, it plans a dual-path scheme with "the main path covering the core training steep slopes and alternative paths suitable for beginners' physical abilities." At the same time, it customizes multimodal interactive content for different members: pushes AR visual annotations of the team's overall physical status and global road conditions to the team leader, outputs voice prompts and handlebar vibrations for "slow down before steep slopes" to beginners, and pushes equipment adaptation prompts for "gear shifting suggestions on gravel roads" to experienced riders, perfectly adapting to the individual needs of different members and the team's collaborative needs.

[0203] During the ride, the solution continuously collected feedback information via the S400: novice riders reported that the steep slope warning was too late, experienced riders reported excessive vibration intensity, and the team as a whole reported that aid station reminders were not timely. These proactive feedbacks, along with the passive feedback of "novice riders deviating from their planned routes," were categorized and organized. Then, an incremental training mechanism was used to update three core models: the user profile model increased the weight of the "steep slope tolerance" tag for novices, the healthy route linkage model increased the weight of the "aid station density" feature, and the multimodal interaction model lowered the tactile vibration intensity threshold for experienced riders. When the team conducts similar training again, the solution can automatically optimize route planning, trigger steep slope warnings for novices 500 meters in advance, accurately push aid station locations, and adjust vibration intensity to an appropriate value. This achieves a closed-loop improvement from "initial adaptation" to "continuous optimization," completely solving the problem that traditional cycling maps with "uniform routes and fixed interactions" cannot adapt to individual team differences, equipment status, and complex scenarios, significantly improving the adaptability of personalized services to individual user and scenario needs.

[0204] In this embodiment of the application, the method further includes performing local lightweight preprocessing on emergency data at edge nodes before standardization and triggering immediate alerts, while simultaneously implementing local encrypted training on privacy data based on a federated learning mechanism:

[0205] Step 1: Integrate equipment-related data into the multi-source acquisition system and model construction:

[0206] When initiating the data acquisition process at the multi-source data sensing end, additional component wear data, historical fault records, and real-time operating status data of cycling equipment are included in the acquisition scope. These, along with the original five categories of key data (user cycling behavior data, physiological state data, etc.), constitute a complete acquisition system, ensuring real-time capture and initial integration of equipment-related data. A lightweight equipment health assessment model is deployed at the edge node. This model uses the collected equipment data as input and, by analyzing the correlation between component wear trends, historical faults, and real-time operating status, calculates the fault risk level (e.g., low risk, medium risk, high risk) of key equipment components such as the frame, tires, and transmission system, generating equipment fault prediction results.

[0207] Step 2, Fault Risk Assessment and Early Warning, and Data Attachment:

[0208] After the equipment health assessment model outputs the fault risk level, it is immediately compared with the preset fault risk threshold. If it is determined to be medium or high risk, an immediate warning is triggered at the edge node. The warning method is adapted to the equipment type (e.g., a high risk in the transmission system triggers a handlebar beep + mobile APP pop-up, and abnormal tire pressure triggers the equipment's own warning light to flash). At the same time, the equipment fault prediction result (including the name of the faulty component, risk level, and prediction basis) is added as an additional field to the key dataset to be standardized, ensuring that this information participates in the subsequent standardization process synchronously with other data.

[0209] Step 3, detailed execution of layered encryption for privacy data:

[0210] Based on local encryption training of privacy-preserving data using a federated learning mechanism, layered encryption processing is performed according to the data's sensitivity:

[0211] User core physiological data (heart rate, blood oxygen saturation, health records, muscle fatigue feedback) are classified as Level 1 privacy data. Lightweight training models are deployed only locally on the user terminal or cycling equipment. After feature learning is completed, only the model training parameters are retained, and no original core physiological data is transmitted to the cloud.

[0212] User cycling preference data (frequently used cycling distances, preferred route types, frequently visited areas, and team cycling role divisions) are classified as secondary privacy data. Personal identification information (such as mobile phone numbers and nicknames) is removed using an anonymization algorithm, and only feature parameters (such as the label weight of "preferring off-road routes" and the quantitative feature of "average cycling speed of 30km / h") are extracted.

[0213] The publicly available cycling route data (urban cycling greenways, public road network information, and distribution of public service stations along the route) are designated as Level 3 data, which do not require encryption and can be directly used for global training of the cloud model.

[0214] Edge nodes only package and transmit the model training parameters of the first-level privacy data, the de-identified feature parameters of the second-level privacy data, and the third-level public data to the cloud database.

[0215] Step 4: Preprocessed log generation and cloud synchronization:

[0216] After completing lightweight preprocessing of emergency data (including equipment failure warning data, physiological abnormality data, etc.), the edge node automatically generates a preprocessing log. The log clearly records the data type (such as equipment failure data, sudden road condition data), processing time (accurate to milliseconds), and warning triggering criteria (such as "tire pressure is lower than 1.8 bar, reaching the preset warning threshold" or "heart rate exceeds 180 beats / minute, meeting the emergency data identification rules"). Through a distributed consistency synchronization mechanism, the preprocessing log is synchronously transmitted to the cloud database. The cloud stores the logs according to timestamps and data types, providing data support for subsequent verification of the effectiveness of the edge preprocessing logic and optimization of preprocessing thresholds (such as adjusting the critical value of tire pressure warning and the judgment criteria for emergency heart rate data).

[0217] By integrating equipment data to build a fault prediction model, early warning of equipment faults and data correlation are achieved; by layering and encrypting privacy data for detailed processing, user privacy and security protection is further strengthened; by generating preprocessing logs and synchronizing them to the cloud, a basis for optimizing preprocessing logic is provided. Ultimately, based on the original data collection, preprocessing, and encrypted transmission, the comprehensiveness of data, privacy and security, and the optimizability of the preprocessing mechanism are improved.

[0218] In this embodiment of the application, the method further includes the following steps in the process of dynamically generating conflict resolution strategies:

[0219] Step 1, Building the core features of team collaboration:

[0220] Core information such as team members' historical cycling data and collaboration records are extracted from S100 standardized data to construct two types of team collaboration characteristics:

[0221] Complementary physical fitness features: Extract data such as average climbing speed, flat road endurance, and long-distance endurance limit of each member, and analyze the complementary advantages among members through feature matching algorithms to clarify the combination matching relationship of "member A's climbing advantage + member B's flat road advantage + member C's endurance advantage" and form a quantitative complementary physical fitness matrix.

[0222] Historical collaboration features: Based on data such as the team's past track synchronization rate, command response time, and efficiency in handling emergencies (such as members falling behind or sudden changes in road conditions), the team's tacit understanding is calculated through clustering algorithms (e.g., a track synchronization rate of ≥85% is considered high tacit understanding), and the emergency response speed level is determined by statistically analyzing the emergency response time (e.g., ≤3 seconds is considered a rapid response), forming a complete set of historical collaboration features.

[0223] Step 2, Refine and generate targeted conflict resolution strategies:

[0224] After S200 conflict identification is completed, refined strategies are generated for the two core conflict types:

[0225] Individual and route difficulty conflict: The system retrieves the user's historical cycling data for routes of similar difficulty (such as average speed, energy consumption rate, and completion status when cycling steep hills in the past), and combines this data with the user's current physical condition to generate a combined adaptation solution of "route replacement (replacing high-difficulty routes with routes of the same type but one level lower difficulty) + cycling rhythm suggestions (such as resting for 30 seconds every 500 meters on steep hills) + energy replenishment planning (sending water replenishment reminders 1 kilometer in advance)".

[0226] Team and equipment performance conflict: Based on the physical fitness complementarity feature matrix constructed in step 1, analyze the compatibility between each member's physical fitness advantages and the performance of existing equipment, and allocate equipment usage priority—prioritize the allocation of high-performance off-road equipment to members with outstanding climbing advantages in physical fitness complementarity, allocate durable equipment to members with endurance advantages, and allocate easy-to-operate equipment to members with fast emergency response speed. At the same time, generate "equipment adaptation route guide" (such as high-priority equipment adaptation to core training routes).

[0227] Step 3: Dynamically update team profiles and iterate strategies:

[0228] The team profile is updated at a preset interval (e.g., every 30 minutes or after completing a key section). At each interval, the real-time physical fitness data (heart rate, blood oxygen, and physical exertion rate) of all members is extracted and compared with the preset physical fitness threshold (based on the members' basic physical fitness level, such as judging members with a heart rate ≥170 beats / minute as having insufficient physical fitness).

[0229] If a member's physical fitness data drops to the physical fitness threshold, immediately adjust their profile label (e.g., change "good physical fitness" to "physical fatigue"), and simultaneously re-predict the member's immediate cycling needs (e.g., change from "complete training goals" to "ensure safety and reduce physical exertion").

[0230] Based on the updated member tags and the overall team profile, the adaptability of the existing cycling plan is re-verified. If new adaptability conflicts arise (such as a conflict between a fatigued member and a subsequent steep slope), the conflict resolution strategy is updated in real time (such as planning a temporary alternative route to bypass the steep slope for the member) and synchronized to the edge nodes for execution.

[0231] By refining the construction of team collaboration characteristics, targeted conflict resolution strategies, and dynamic update mechanisms, the personalized adaptation capability of the team in cycling scenarios has been further enhanced, the accuracy and timeliness of conflict resolution strategies have been improved, and the dynamic matching of the team's overall needs with individual differences, equipment performance, and road characteristics has been ensured.

[0232] In this embodiment of the application, the method further includes the following steps in dynamically configuring the multimodal interaction data cache set:

[0233] Step 1: Target edge node filtering dimension supplementation and refinement execution:

[0234] Based on the multi-dimensional filtering of S300 target edge nodes, two new core filtering dimensions have been added and the judgment logic has been refined:

[0235] Dimension 1 (Service Load Status Filtering): Through the communication link between the central server and the candidate edge nodes, the current service load data of each node (such as CPU utilization, memory utilization, and data processing queue length) is obtained in real time and compared with the preset service load thresholds (such as CPU utilization ≤70% and memory utilization ≤65%) to remove nodes with excessive load.

[0236] Dimension 2 (Data Transmission Latency Screening): Using a combination of ping test and data fragmentation transmission test, the actual data transmission latency between the user terminal and each candidate node is calculated, and the node with the shortest latency of ≤100 milliseconds is retained first.

[0237] Final determination: If multiple nodes still meet the criteria after the first two dimensions of screening, retrieve the historical interaction service success rate data of each node (such as the percentage of successful service responses in the past 3 months), select the node with a success rate ≥ 95% and the highest ranking as the target edge node to ensure the stability of node services.

[0238] Step 2, Dynamically adjust the priority of cached content based on specific scenarios:

[0239] Based on the determined S300 cache content, the cache priority is precisely adjusted according to the characteristics of the real-time riding scenario:

[0240] Rainy day cycling scenario determination: After confirming the scenario type through the real-time environmental data (rainfall intensity, road surface humidity) collected by S100, the "tactile vibration coding data of road water accumulation points" (such as different vibration frequencies corresponding to different water depths) and "visual annotation data of water avoidance routes" (such as AR blue highlighted water avoidance paths) are set to the highest priority to ensure that users can quickly obtain core safety interaction data.

[0241] High-intensity training scenario determination: After confirming the scenario type through the cycling intention (professional training) predicted by S200 and real-time physical fitness data (high consumption, high heart rate), the "health supply point data" (location, supply type, operation status) and "physical fitness warning data" (physical fitness consumption rate, warning threshold, supply suggestions) are set to the highest priority to adapt to the health protection needs during training.

[0242] Priority execution: Storage resources are allocated to target edge nodes according to the adjusted priority, and the highest priority data is stored in the high-speed cache to ensure millisecond-level call response.

[0243] Step 3: Implement the dynamic update mechanism for cached data.

[0244] Establish a real-time data monitoring and cache update linkage process:

[0245] Trigger condition monitoring: Continuously monitor new data transmitted by S100. When new meteorological data (such as changes in precipitation intensity, sudden increase in wind speed), road condition data (such as sudden construction, road repair) or user physical fitness data (such as sudden drop in heart rate, sudden increase in physical fitness rate) are detected, the cache is immediately updated.

[0246] Expired data cleanup: Identify expired interactive data in the cache (such as expired no-rain road condition labels and outdated physical fitness warning thresholds) by using data timeliness tags (such as 1-hour validity period for weather data and 15-minute validity period for road condition data), automatically perform cleanup operations, and release storage resources;

[0247] New data supplementation: The latest scene-adapted data (such as updated water accumulation point locations and adjusted physical fitness warning thresholds) will be supplemented to the storage area corresponding to the cache set according to the current scene priority to ensure that the cached data is consistent with the real-time scene and user status.

[0248] By refining the selection dimensions of target edge nodes, adjusting cache priorities according to scenarios, and establishing a dynamic update mechanism, the accuracy of edge node selection and service stability have been further improved, ensuring that cached data is highly adapted to real-time scenarios and user needs and is timely.

[0249] In this embodiment of the application, the method further includes, in the process of obtaining user multimodal interaction commands and parsing cycling intention information, and generating multi-dimensional matching features for health path linkage by combining the configured cached path and health data:

[0250] Step 1, Passive command parsing and execution based on user's physiological state:

[0251] Simultaneously, the real-time monitoring process of the user's physiological state is initiated while the multimodal command acquisition interface is launched:

[0252] The system acquires core physiological data such as user heart rate and blood oxygen saturation in real time from edge nodes and compares them with preset safe heart rate values ​​(such as 60-100 beats / minute for adults at rest and ≤180 beats / minute for exercise) and preset blood oxygen standard values ​​(such as ≥95%).

[0253] If any abnormal physiological indicator is detected (heart rate exceeding the safe value, blood oxygen below the standard value), the passive instruction generation logic is immediately triggered. Without the user's active operation, standardized passive interaction instructions are automatically generated. The instruction content clearly indicates the type of abnormality (such as "heart rate exceeding the standard" or "blood oxygen being low"), the current indicator value, and preliminary suggestions (such as "suggest slowing down and resting" or "need to replenish water in time").

[0254] The generated passive interaction commands and active interaction commands are stored together in the command cache area of ​​the target edge node and participate in the subsequent intent parsing process.

[0255] Step 2, Construction and Quantification of Supplementary Dimensions for Health Path Linkage Matching Features:

[0256] Based on the path data and health data extracted from the S300 cache set, two new feature dimensions were added and quantization was completed:

[0257] Health supply resource density dimension: The location data of water supply points and energy supply points along the route and the total path length data are retrieved from the cache. The quantitative value is calculated by the formula "Number of supply points per unit path length = (Number of water supply points + Number of energy supply points) ÷ Total path length (unit: points / km)", forming the standardized feature data of this dimension.

[0258] Medical site accessibility dimension: Based on the coordinates of the nearest medical site in the cache and the route data of the current path, the shortest time (in minutes) to ride from any node on the path to the medical site is calculated through a route planning algorithm. This serves as a quantitative feature of this dimension, ensuring that the data intuitively reflects the efficiency of emergency medical accessibility.

[0259] The two supplementary dimensional features are integrated with the original basic adaptation features and scenario adaptation features, and normalized according to a unified weighting rule to form a complete health path linkage multi-dimensional matching feature vector.

[0260] Step 3: Integrating features to adapt to training objectives in professional training scenarios:

[0261] After completing the cycling intention analysis, if the cycling intention is determined to be professional training, the training objective adaptation feature integration process is initiated:

[0262] Extract the professional training goals recorded in the user profile (such as improving hill climbing ability, long-distance endurance training, and improving speed on flat roads) to clarify the core training needs;

[0263] The training objective adaptation features are broken down into two core sub-dimensions: path slope change frequency (the number of road segments with a slope ≥5° per unit mileage) and road segment smoothness (quantified into 0-10 points based on the degree of road bumpiness, with 10 points being completely smooth).

[0264] The matching degree between sub-dimensions and training objectives is calculated through a feature matching algorithm: for example, for hill climbing training objectives, the matching degree weight of "slope change frequency" is increased, and for endurance training objectives, the matching degree weight of "proportion of long-distance flat road sections" is increased. Finally, the quantified matching degree data is integrated into the health path linkage multi-dimensional matching feature vector to ensure that the features are accurately aligned with professional training needs.

[0265] By adding passive instruction parsing for physiological abnormalities, supplementing health-related feature dimensions, and integrating professional training goal adaptation features, the comprehensiveness of cycling intention parsing and the accuracy of path matching features have been further improved, especially the adaptability of health protection and professional training scenarios has been strengthened.

[0266] In this embodiment of the application, the method further includes the following steps in the closed-loop iterative optimization process of updating user profiles, health route linkage, and multimodal interaction models based on feedback information and incremental training mechanisms to achieve dynamic interaction capabilities of cycling maps:

[0267] Step 1: Categorize, filter, and precisely define feedback information:

[0268] Based on the comprehensive collection and classification of user feedback information, a three-level classification was further completed according to feedback attributes and application scenarios:

[0269] Individual-specific feedback filtering: Extract unique feedback that belongs only to a single user (such as a user's exclusive off-road route preference adjustment or personalized vibration intensity requirements), link it to the corresponding personal account through user identification, and clarify that it is only used to optimize the user's personalized model parameters;

[0270] Group common feedback screening: Analyze all user feedback through clustering algorithms, extract the needs mentioned by multiple users (≥30% of the user group) (such as the delayed supply point reminders reported by most users, or the unreasonable interaction form in a certain scenario), eliminate individual difference factors, and form a group common feedback dataset;

[0271] Emergency scenario feedback filtering: Accurately extract feedback data related to emergency events (such as excessively long emergency response time, rescue routes deviating from the optimal path, and unclear emergency interaction prompts), and clarify the data dimensions including emergency response time, rescue route adaptability, and emergency interaction effectiveness, and define them separately as emergency scenario feedback datasets.

[0272] Step 2, Emergency-Specific Optimization Sub-Model Construction and Training:

[0273] For the emergency scenario feedback dataset, a specialized sub-model building and training process was initiated:

[0274] Sub-model architecture construction: A lightweight emergency-specific optimization sub-model is deployed on the central server. It adopts a neural network structure focused on emergency scenarios. The input layer only receives three types of data: emergency response duration, rescue route adaptability, and emergency interaction effectiveness. The output layer corresponds to emergency strategy adjustment parameters (such as early warning duration, rescue route planning weight, and emergency interaction form).

[0275] Data input and training iteration: The classified emergency scenario feedback data is divided into training set and test set in a 7:3 ratio and input into the sub-model for iterative training. The model parameters are optimized by gradient descent algorithm with the training objectives of "improving emergency response efficiency by ≥15% and rescue route adaptability by ≥90%" until the model test accuracy meets the target.

[0276] Sub-model integration application: The trained specialized optimization sub-model is integrated with the original health path linkage model and multimodal interaction model to form a synergy, which is triggered only in emergency scenarios to specifically handle emergency-related model optimization needs.

[0277] Step 3, Setting model parameter protection thresholds and executing incremental training:

[0278] Protection threshold setting: Based on the core cycling preference tags in the user profile (such as long-term off-road route preference, fixed cycling speed requirement, and custom equipment adaptation habits), by statistically analyzing the stability of users' historical feedback and the duration of preference persistence, protection thresholds are set for core parameters (such as the adjustment range of core preference tag weight ≤ ±10%) to avoid excessive interference of common feedback with individual core needs.

[0279] Categorized incremental training: Three types of feedback data are input into the corresponding models to perform optimization: individual-specific feedback is input into the user profile model to fine-tune the user's personalized parameters; group common feedback is input into the global health path linkage model and multimodal interaction model to complete global parameter iteration; emergency scenario feedback drives the update of emergency-specific optimization sub-models, and the optimization results are synchronized to the global emergency strategy library.

[0280] Parameter adjustment verification: During training, the model parameters corresponding to core cycling preferences are monitored in real time. If the adjustment is about to exceed the protection threshold, the limiting mechanism is automatically triggered, and the adjustment is performed at the upper limit of the threshold to ensure that the user's core needs are not diluted.

[0281] By classifying feedback information into three levels, precise adaptation and optimization are achieved for "individual-group-emergency". Emergency-specific sub-models enhance emergency scenario service capabilities. Parameter protection thresholds safeguard users' core personalized needs, further improving the targeting of model iterations, the professionalism of emergency services, and the stability of personalized experiences.

[0282] In this embodiment of the application, after synchronously generating cross-device multimodal interactive content and outputting it to the user terminal and cycling equipment, the method further includes:

[0283] Step 1, Construction and data fusion of digital twin model for cycling scenario:

[0284] After the initial generation of cross-device multimodal interaction content, the digital twin model construction process is initiated:

[0285] Model basic framework construction: Using the personalized cycling path generated by S400 as the spatial benchmark, a digital twin model framework that maps to the real cycling scene in a 1:1 ratio is constructed, including geospatial elements such as path coordinates, road segment features, and surrounding environment.

[0286] Multi-source data fusion injection: Real-time meteorological data (temperature, precipitation, wind speed) and road condition data (road surface smoothness, congestion status, risk point distribution) are extracted from the S300 cache set. Real-time user physical fitness data (heart rate, physical fitness consumption rate) and equipment status data (component wear level, operating parameters) are obtained from edge nodes. All data are injected into the model in real time through the data interface to achieve dynamic synchronization between the physical scene and the digital model.

[0287] Full-element pre-simulation capability activated: The model integrates a physics engine and behavior simulation algorithm, and has the ability to virtually reproduce the interaction of multiple elements such as "human-vehicle-road-environment" during cycling, laying the foundation for subsequent simulation analysis.

[0288] Step 2: Generating a virtual simulation and quantitative comparison report of multiple cycling options:

[0289] Based on the digital twin model, multi-scheme simulation and comparative analysis were initiated:

[0290] Simulation scenario setting: Load the optimal main path and alternative paths generated by S400, and set simulation parameters (such as user cycling speed, energy consumption rate, and environmental change trend) to be consistent with the real cycling scenario;

[0291] Core Dimension Simulation: Simulate the three core processes under each scheme respectively: physical energy consumption process (real-time calculation of physical energy consumption values ​​for different road sections and prediction of peak physical energy throughout the journey), risk event trigger probability (calculation of the probability of risks such as steep slope falls and equipment failures based on road conditions and environmental data), and interactive service response effect (simulating the transmission latency and user reception efficiency of multimodal interactive content).

[0292] Output of quantitative comparison report: Summarize the simulation results of each scheme and generate a quantitative comparison report containing core indicators such as "peak physical fitness (unit: kcal), number of risk points (high / medium / low risk are counted separately), interaction response time (unit: milliseconds), estimated riding time, and supply point suitability". The report is then pushed to the user terminal for selection after being sorted by overall suitability.

[0293] Step 3, Optimal Solution Confirmation and Multimodal Interaction Content Calibration:

[0294] User optimal solution selection: Users can view the quantitative comparison report through the terminal and select the optimal pre-rehearsal solution based on their own needs (such as prioritizing risk reduction and prioritizing control of physical exertion). The system receives the user's selection instruction and locks the corresponding solution.

[0295] Interactive content calibration execution: Based on the virtual pre-simulation results of the optimal solution, the adaptability of interactive content for each segment of the path is verified one by one. The focus is on identifying visual interactive interference scenarios (such as blurred AR markings under strong sunlight in rainy weather, visual cues in tunnels being easily ignored, and the risk of visual distraction when riding at high speed). If interference is predicted, the visual cues (such as AR arrows and pop-up text) of that segment are automatically switched to tactile cues (such as specific frequency vibration of the handlebars and vibration encoding of the wristband), while retaining necessary voice cues and adjusting the volume.

[0296] Post-calibration content synchronization: The calibrated multimodal interactive content is synchronized to all related devices such as user terminals, smart helmets, and smart handlebars to ensure that the output content of each device is consistent and adapted to the road scene, and finally completes the precise optimization of interactive content.

[0297] By constructing a digital twin model, we can achieve virtual pre-simulation of all elements of cycling scenarios and quantitative comparison of multiple solutions. Based on the optimal solution selected by the user, we can calibrate multimodal interactive content, effectively avoid visual interaction interference, and further improve the scene adaptation accuracy of interactive content and cycling safety.

[0298] This application discloses an AI-based dynamic interactive system for cycling maps, referring to... Figure 2 ,include:

[0299] The multi-source data perception and synchronization module 001 collects various types of data from user terminals, environmental monitoring equipment, road network service platforms, and cycling equipment sensors in real time at the multi-source data perception end. It extracts user cycling behavior data, physiological state data, regional real-time environmental data, cycling service resource data, and equipment performance parameters, and generates unified standardized data. Before standardization, emergency data is preprocessed locally at the edge node and triggers an immediate warning. Privacy data is encrypted and trained locally based on a federated learning mechanism. The standardized data is then transmitted to the cloud database through a distributed consistency synchronization mechanism.

[0300] The User Profile and Conflict Resolution Module 002 constructs a multi-dimensional and accurate user profile based on standardized user basic information, historical riding characteristics, equipment performance parameters, and team member physical fitness data, and predicts real-time riding needs. When there is a mismatch between profile tags and road difficulty or equipment performance, a conflict resolution strategy is dynamically generated. The conflict resolution strategy is executed by the local preprocessing module deployed on the edge node. The local preprocessing module performs millisecond-level fast response to sudden data and performs batch clustering processing on periodic data.

[0301] Edge node cache configuration module 003, after conflict resolution, in the hybrid computing architecture of edge nodes and central server collaboration, selects target edge nodes by combining user geographical location information, real-time physical status, and equipment adaptation scenarios, and dynamically configures a multimodal interactive data cache set.

[0302] The interaction iteration and path generation module 004 acquires user multimodal interaction commands and parses cycling intention information. It combines cached paths and health data to generate multi-dimensional matching features for health path linkage. It determines personalized cycling paths and graded risk warning information through a time-series deep learning model. It simultaneously generates and outputs cross-device multimodal interaction content and acquires user feedback information. Based on the feedback information and an incremental training mechanism, it updates the three types of models: user profile, health path linkage, and multimodal interaction, to achieve closed-loop iterative optimization of the dynamic interaction capabilities of the cycling map.

[0303] This application also discloses an electronic device, including a processor, wherein the processor runs a program for the AI-based dynamic interaction method for cycling maps described in any one of the above embodiments.

[0304] This application also discloses a storage medium storing a program for the AI-based dynamic interaction method for cycling maps described in any one of the above embodiments.

[0305] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. An AI-based dynamic interaction method for cycling maps, characterized in that, include: At the multi-source data sensing end, various types of data from user terminals, environmental monitoring equipment, road network service platforms, and cycling equipment sensors are collected in real time. User cycling behavior data, physiological state data, regional real-time environmental data, cycling service resource data, and equipment performance parameters are extracted and unified standardized data is generated. Before standardization, emergency data is preprocessed locally at the edge node and an immediate warning is triggered. Privacy data is encrypted and trained locally based on a federated learning mechanism. The standardized data is then transmitted to the cloud database through a distributed consistency synchronization mechanism for subsequent profile updates and strategy optimization. Based on standardized user basic information, historical cycling characteristics, equipment performance parameters, and team member physical fitness data, a multi-dimensional accurate user profile is constructed to predict real-time cycling needs. When there is a mismatch between profile tags and road difficulty or equipment performance, a conflict resolution strategy is dynamically generated. The conflict resolution strategy is executed by a local preprocessing module deployed on edge nodes. The local preprocessing module performs millisecond-level fast response to sudden data and performs batch clustering processing on periodic data. After the conflict is resolved, in the hybrid computing architecture that coordinates edge nodes and central servers, the target edge node is selected by combining user geographic location information, real-time physical condition, and equipment adaptation scenario, and a multimodal interactive data cache set is dynamically configured. All cached content is associated with user profile tags. The system acquires user multimodal interaction commands and parses cycling intent information. It combines cached paths and health data to generate multidimensional matching features for health path linkage. It uses a time-series deep learning model to determine personalized cycling paths and graded risk warning information. It simultaneously generates and outputs cross-device multimodal interaction content, acquires user feedback information, and updates the user profile, health path linkage, and multimodal interaction models based on the feedback information and an incremental training mechanism. This achieves closed-loop iterative optimization of the dynamic interaction capabilities of the cycling map. When generating multi-dimensional matching features for health routes, two additional feature dimensions are added: health supply resource density and medical station accessibility. Health supply resource density is the number of water and energy supply points per unit path length, and medical station accessibility is the shortest cycling time from the path to the nearest medical station.

2. The AI-based dynamic interaction method for cycling maps according to claim 1, characterized in that, Before standardization, the method involves performing local lightweight preprocessing on emergency data at edge nodes and triggering immediate alerts. Simultaneously, it includes local encrypted training of privacy-preserving data based on a federated learning mechanism. The collected data on component wear and tear, historical fault records, and real-time operating status of cycling equipment are integrated into a multi-source data acquisition system. An equipment health assessment model is built at the edge node. The fault risk level of key components of the equipment is calculated through the equipment health assessment model. If the fault risk level reaches the preset fault risk threshold, an immediate warning is triggered, and the equipment fault prediction result is added to the additional field of the standardized data. The privacy data is encrypted in layers. The user's core physiological data is classified as Level 1 privacy data, which is used for model training only on the terminal side and no raw data is transmitted. The user's cycling preference data is classified as Level 2 privacy data, which is used for transmitting only the de-identified feature parameters to the cloud database. The publicly available cycling route data is classified as Level 3 data, which is used for the global training of the cloud model. After completing emergency data preprocessing, the edge node generates a preprocessing log, which includes data type, processing time, and early warning triggering basis. The log is synchronized to the cloud database for subsequent verification of the effectiveness of the edge preprocessing logic and optimization of preprocessing thresholds.

3. The AI-based dynamic interaction method for cycling maps according to claim 2, characterized in that, The method also includes the following in the process of dynamically generating conflict resolution strategies: Construct team members' complementary physical fitness characteristics and historical collaboration characteristics. Complementary physical fitness characteristics include the matching relationship of advantages in climbing, flat roads, and endurance. Historical collaboration characteristics cover team tacit understanding and emergency response speed. To address conflicts between individual and road segment difficulty, historical data of users on road segments of similar difficulty are used to generate an adaptation solution. To address conflicts between team and equipment performance, equipment usage priority is allocated based on complementary physical abilities. The team's overall profile is updated according to a preset duration. If a member's physical fitness drops to the physical fitness threshold, the profile label of the corresponding member is adjusted and the cycling needs are re-predicted and the conflict resolution strategy is updated.

4. The AI-based dynamic interaction method for cycling maps according to claim 3, characterized in that, The method also includes the following in the process of dynamically configuring the multimodal interaction data cache set: When selecting target edge nodes, two filtering dimensions are added: the service load status of the node and the data transmission latency. Edge nodes with service load below the preset service load threshold and the shortest data transmission latency are given priority. If there are multiple edge nodes that meet the conditions, the final judgment is made by combining the historical interaction service success rate of the edge nodes. When dynamically configuring the multimodal interaction data cache set, the priority of the cached content is adjusted according to the real-time scene characteristics. For example, in the case of cycling in the rain, the tactile vibration encoding data of road water accumulation points and the visual annotation data of water avoidance routes are set to the highest priority. In the case of high-intensity training, the health supply point data and physical fitness warning data are set to the highest priority. Establish a dynamic update mechanism for cached data. When new weather data, road condition data, or user physical fitness data are received, automatically clean up expired interactive data in the cache and supplement it with the latest scene adaptation data to ensure the timeliness and accuracy of the cache set.

5. The AI-based dynamic interaction method for cycling maps according to claim 4, characterized in that, In the process of acquiring user multimodal interaction commands and parsing cycling intent information, and combining the configured cached paths with health data to generate multi-dimensional matching features for health path linkage, the method also includes: When parsing multimodal interaction commands, passive command parsing of user physiological state is added. When multiple abnormal physiological states such as user heart rate exceeding the preset heart rate safety value and blood oxygen below the preset blood oxygen standard value are detected, passive interaction commands for physiological abnormalities are automatically generated without user active triggering. If the identified cycling intention is professional training, additional training objective adaptation features are incorporated when generating matching features. These features include the frequency of slope changes, road smoothness, and the degree of fit between these features and the user's training objectives, ensuring that the path meets the requirements of professional training.

6. The AI-based dynamic interaction method for cycling maps according to claim 1, characterized in that, In the process of achieving closed-loop iterative optimization of the dynamic interactive capabilities of cycling maps by updating user profiles, health route linkage, and multimodal interaction models based on feedback information and incremental training mechanisms, the method also includes: Feedback information is divided into three categories: individual-specific feedback, group-wide feedback, and emergency scenario feedback. Individual-specific feedback is only used to optimize the personalized model parameters corresponding to the user, group-wide feedback is used for parameter iteration of the global model, and emergency scenario feedback is used for optimization of specific emergency models. To address the feedback data from emergency scenarios, a specialized optimization sub-model is established. This sub-model only receives emergency scenario feedback data related to emergency events, including emergency response time, rescue route adaptability, and emergency interaction effectiveness. Through specialized training, the model's service capabilities in emergency scenarios are improved. During incremental training, a protection threshold is set for the model parameters. For model parameters corresponding to the user's core cycling preferences, the adjustment range must not exceed the protection threshold.

7. The AI-based dynamic interaction method for cycling maps according to claim 1, characterized in that, After synchronously generating cross-device multimodal interactive content and outputting it to the user terminal and cycling equipment, the method also includes: Based on the configured cached data and the generated personalized cycling routes, a digital twin model of the cycling scenario is constructed. The digital twin model of the cycling scenario integrates real-time weather data, road condition data, user physical fitness data, and equipment status data to realize a virtual pre-rehearsal of all elements of the cycling route. In the digital twin model of cycling scenarios, the process of physical exertion, probability of risk event triggering, and response effect of interactive services under different cycling schemes are simulated, and a quantitative comparison report of multiple schemes is generated. The report includes multiple core indicators such as peak physical exertion, number of risk points, and interaction response time for each scheme. Based on the optimal pre-simulation scheme selected by the user, the multimodal interactive content is finally calibrated. If visual interaction interference is predicted in a certain road segment during the pre-simulation, the interaction form of the corresponding road segment is adjusted in advance, and the visual cues are switched to tactile cues.

8. An AI-based dynamic interactive cycling map system, characterized in that, include: The multi-source data perception and synchronization module collects various types of data in real time from user terminals, environmental monitoring equipment, road network service platforms, and cycling equipment sensors at the multi-source data perception end. It extracts user cycling behavior data, physiological state data, regional real-time environmental data, cycling service resource data, and equipment performance parameters, and generates unified standardized data. Before standardization, emergency data is preprocessed locally at the edge node and triggers an immediate warning. Privacy data is encrypted and trained locally based on a federated learning mechanism. The standardized data is then transmitted to the cloud database through a distributed consistency synchronization mechanism for subsequent profile updates and strategy optimization. The user profiling and conflict resolution module constructs multi-dimensional and accurate user profiles based on standardized user basic information, historical riding characteristics, equipment performance parameters, and team member physical fitness data, and predicts real-time riding needs. When there is a mismatch between profile tags and road difficulty or equipment performance, conflict resolution strategies are dynamically generated. The conflict resolution strategies are executed by the local preprocessing module deployed on the edge node. The local preprocessing module performs millisecond-level fast response to sudden data and performs batch clustering processing on periodic data. After conflict resolution, the edge node cache configuration module selects the target edge node in the hybrid computing architecture that coordinates edge nodes and central servers, combining user geographic location information, real-time physical fitness status, and equipment adaptation scenarios, and dynamically configures a multimodal interactive data cache set. All cached content is associated with user profile tags. The interaction iteration and path generation module acquires user multimodal interaction commands and parses cycling intention information. It combines cached paths and health data to generate multi-dimensional matching features for health path linkage. It determines personalized cycling paths and graded risk warning information through a time-series deep learning model. It simultaneously generates and outputs cross-device multimodal interaction content and acquires user feedback information. Based on the feedback information and an incremental training mechanism, it updates the three types of models: user profile, health path linkage, and multimodal interaction, to achieve closed-loop iterative optimization of the dynamic interaction capabilities of the cycling map. When generating multi-dimensional matching features for health routes, two additional feature dimensions are added: health supply resource density and medical station accessibility. Health supply resource density is the number of water and energy supply points per unit path length, and medical station accessibility is the shortest cycling time from the path to the nearest medical station.

9. An electronic device, characterized in that, Includes a processor, wherein the processor runs a program for an AI-based dynamic interaction method for cycling maps as described in any one of claims 1-7.

10. A storage medium, characterized in that, The program stores the AI-based dynamic interaction method for cycling maps as described in any one of claims 1-7.

Images