An ebike user behavior prediction and personalized power adjustment method

Abstract

The application provides an Ebike user behavior prediction and personalized power adjustment method, which comprises: collecting Ebike multi-dimensional operation data, the multi-dimensional operation data comprising user static physiological data, dynamic riding behavior data and scene perception data, and generating a user riding behavior fingerprint through feature engineering processing on the multi-dimensional operation data; generating a scene-intention prediction result through multi-scale time series modeling and intention decoding processing on the riding behavior fingerprint; calling a personalized power adjustment rule library, the rule library integrating a scenario-based power baseline, user behavior adaptation rules and battery protection constraint rules, performing dynamic calibration of power parameters on the scene-intention prediction result to generate adaptive power adjustment instructions; performing power parameter calibration on the adaptive power adjustment instructions, feeding back the battery energy state to the adjustment module synchronously, and dynamically adjusting the assistance ratio and the torque change rate to perform Ebike power output and user behavior, driving scene and battery state collaborative adaptation.

Description

Technical Field

[0001] This application relates to the field of electric-assisted bicycle control technology, and more specifically, to an Ebike user behavior prediction and personalized power adjustment method. Background Technology

[0002] With the popularization of the concept of green travel, ebikes have become an important tool for urban commuting, outdoor fitness, and leisure travel. Users are increasingly demanding higher adaptability and comfort in power output during riding. Different users have significant differences in physiological characteristics and riding habits, and their power needs vary even in the same riding scenario. This requires ebikes to have power adjustment that can accurately match individual user characteristics and scenario changes.

[0003] In existing technologies, a common Ebike power adjustment scheme is to preset a fixed power mode and determine the assistance ratio by switching gears. The core idea is to divide the motor's rated power into a limited number of gears, each corresponding to a fixed assistance coefficient. The user manually switches gears based on subjective feeling, and the system outputs a fixed proportion of assistance power only according to the current gear. This scheme simplifies the control logic to achieve basic assistance functionality, eliminating the need for complex data processing and model calculations.

[0004] However, this fixed-mode power adjustment scheme has obvious technical flaws: it cannot dynamically adjust the power output based on individual user characteristics and real-time riding behavior, resulting in poor power adaptability. For example, a fixed assist ratio is difficult to simultaneously meet the power needs of users of different heights and weights, and it cannot match the user's real-time changes in cadence and pedaling habits, often resulting in a mismatch between assist and human power, affecting riding smoothness and comfort. This problem is precisely the core pain point that the technical solution in this application aims to solve. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides an Ebike user behavior prediction and personalized power adjustment method to alleviate the problem of insufficient power adaptability in the prior art.

[0006] The technical solutions provided in this application are as follows: A method for predicting Ebike user behavior and personalizing motivation includes the following steps: Step 1: Collect Ebike multi-dimensional operation data, which includes user static physiological data, dynamic cycling behavior data and scene perception data. The multi-dimensional operation data is processed by feature engineering to generate the user's cycling behavior fingerprint. Step 2: Generate scene-intent prediction results by processing the cycling behavior fingerprint through multi-scale temporal modeling and intent decoding; Step 3: Call the personalized power adjustment rule library, which integrates scenario-based power baseline, user behavior adaptation rules and battery protection constraint rules, and perform dynamic calibration of power parameters on the scenario-intent prediction results to generate adaptive power adjustment instructions; Step 4: Perform power parameter calibration on the adaptive power adjustment command, and simultaneously feed back the battery energy status to the adjustment module to dynamically adjust the assist ratio and torque change rate, so as to coordinate the Ebike power output with user behavior, driving scenario and battery status.

[0007] Optionally, step 1 includes: Step 11: Clean and integrate user static physiological data, dynamic cycling behavior data and scene perception data, remove abnormal and interfering data and associate them with user identifiers to generate a set of cycling adaptation and compliance features. Step 12: Construct cycling behavior trend features and cycling parameter correlation features for the cycling adaptation compliance feature set. The cycling behavior trend features reflect the changing patterns of cycling parameters, and the cycling parameter correlation features reflect the collaborative relationship between different cycling parameters, so as to generate cycling time-series correlation feature groups. Step 13: Perform weighted adaptive fusion on the cycling time-series associated feature group to dynamically adjust the weight ratio of static physiological features, dynamic behavioral features and scene perception features according to the scene type, and generate cycling behavior fingerprint.

[0008] Optionally, step 11 includes: Step 111: Standardize the format of the user's static physiological data, unify the data measurement units and storage format, in order to generate standardized physiological data for cycling adaptation; Step 112: Perform outlier removal on dynamic cycling behavior data. Based on the behavior rationality rule base, determine and filter parameter values ​​that exceed the normal cycling range to generate regular cycling behavior data. Step 113: Remove redundant information from the scene perception data, retain the scene parameters associated with cycling behavior, and associate and bind the standardized physiological data for cycling adaptation with the regularized behavioral data for cycling to generate a set of cycling adaptation compliance features.

[0009] Optionally, step 12 includes: Step 121: Calculate the rate of change of time-series parameters in the cycling adaptation compliance feature set, capture the increase or decrease trend of parameters over time, and generate cycling behavior trend features; Step 122: Perform correlation analysis on different types of cycling parameters in the cycling adaptation compliance feature set, construct the correlation mapping relationship between parameters, and generate cycling parameter correlation features; Step 123: Integrate the cycling behavior trend features and cycling parameter correlation features, classify and sort them according to data dimensions to generate cycling time-series correlation feature groups.

[0010] Optionally, step 13 includes: Step 131: Construct a scene type recognition model to classify scene perception data to determine the current cycling scene type; Step 132: Set feature weight allocation rules for the scene type, wherein the weight ratio of relevant features is adjusted for different scenes; Step 133: Perform weighted fusion operation on each feature in the cycling time-series associated feature group according to the weight allocation rule to generate a cycling behavior fingerprint that matches the current cycling scenario type.

[0011] Optionally, step 2 includes: Step 21: Perform multi-scale convolutional feature extraction on cycling behavior fingerprints. Local and global behavioral features are captured in parallel by using a first-size convolutional kernel and a second-size convolutional kernel. The size of the second-size convolutional kernel is larger than that of the first-size convolutional kernel to generate a multi-scale feature map of cycling. Step 22: Input the cycling multi-scale feature map into a bidirectional temporal network to mine the sequential dependencies of behavioral features in order to generate a cycling temporal correlation feature vector; Step 23: Perform intent decoding on the cycling time-series associated feature vector and generate scene-intent prediction results by combining scene feature matching.

[0012] Optionally, step 21 includes: Step 211: Use a first-size convolutional kernel to extract local features from the cycling behavior fingerprint to obtain the original local cycling features, and perform association enhancement processing on the original local cycling features based on the correlation of instantaneous cycling parameters to generate a cycling local feature map; Step 212: Use a second-size convolutional kernel to extract global features from the cycling behavior fingerprint to obtain the original global cycling features. The size of the second-size convolutional kernel is larger than that of the first-size convolutional kernel. Perform time-series trend aggregation processing on the original global cycling features to determine the global cycling trend features based on the behavioral change trends within continuous time periods, so as to generate a global cycling feature map. Step 213: Perform channel attention weighted fusion on the local feature map of cycling and the global feature map of cycling to generate a multi-scale feature map of cycling.

[0013] Optionally, step 22 includes: Step 221: Divide the multi-scale feature map of cycling into cycling time-series feature segments according to the time dimension; Step 222: Input the cycling time-series feature fragments into the forward and reverse paths of the bidirectional temporal network, and extract the influence of historical behavior and the trend of future behavior to obtain the bidirectional cycling path output features. Step 223: Fuse the output features of the bidirectional cycling path to construct a feature representation containing complete temporal dependencies, so as to generate a cycling temporal correlation feature vector.

[0014] Optionally, step 23 includes: Step 231: Divide the cycling behavior into stages and set the attention weight allocation ratio for different stages; Step 232: Weight the cycling time-series related feature vectors according to the assigned attention weights to enhance the expression strength of core features related to intent in order to generate a cycling intent-related weighted feature vector; Step 233: Match the weighted feature vector associated with riding intention with the preset scene feature template to identify the current riding scene and decode the user's power demand to generate scene-intention prediction results.

[0015] Optionally, step 3 includes: Step 31: Match the scenario type in the scenario-intention prediction result with the scenario-based power baseline in the personalized power adjustment rule base to determine the range of basic power parameters for riding; Step 32: Based on the range of basic cycling power parameters, call the user behavior adaptation rules to adjust the power demand of the user in the scenario-intent prediction results to obtain the adjusted cycling power parameters by adjusting the assist ratio and torque change rate parameters. Step 33: Limit the power parameters after riding adjustment according to the battery protection constraint rules to generate an adaptive power adjustment command.

[0016] Optionally, step 4 includes: Step 41: Perform hierarchical analysis on the charging and discharging parameters, assist ratio parameters, and torque change rate parameters in the adaptive power adjustment command, and extract the reference values ​​and adjustment thresholds of each parameter to generate a power adjustment parameter analysis set; Step 42: Real-time collection of remaining battery power, cell voltage consistency and energy pool working status, integration to form a battery energy state dataset, and feeding the battery energy state dataset back to the adjustment module to trigger the parameter dynamic calibration logic; Step 43: Based on the battery energy state dataset and the power adjustment parameter analysis set, construct a "demand-supply" adaptation model. When the battery has sufficient remaining power, adjust the assist ratio according to user behavior characteristics to match the power exertion habit. When the battery has low remaining power, optimize the torque change rate while maintaining the basic power. Step 44: Based on the road conditions of the current driving scenario, perform scenario adaptation verification on the adjusted assist ratio and torque change rate, and measure the smoothness of the power output connection in different stages of start-stop, acceleration, and constant speed, so as to achieve dynamic collaborative adaptation between Ebike power output and user behavior, driving scenario and battery status.

[0017] The technical solution provided in this application has the following technical advantages: This application provides precise user characteristic support for personalized power adjustment by collecting multi-dimensional operational data and generating cycling behavior fingerprints. Cycling behavior fingerprints integrate static physiological data, dynamic cycling behavior data, and scene perception data, comprehensively depicting individual user characteristics and cycling habits. This differs from the limitations of traditional fixed-mode systems that rely solely on gear parameters, making power adjustment more comprehensive and tailored to the user's actual situation.

[0018] Based on multi-scale temporal modeling and intent decoding of cycling behavior fingerprints, this system enables the prediction of users' short-term power needs. This process simultaneously captures local features and global trends of cycling behavior, uncovering dependencies between actions and enabling early identification of user intentions such as starting, stopping, and accelerating. This avoids the assistance delay problem caused by traditional solutions that rely solely on passive responses to real-time data, resulting in a higher degree of alignment between power output and the user's effort intentions.

[0019] By invoking a personalized power adjustment rule library for dynamic parameter calibration, the adaptability and safety of power output across various scenarios are ensured. The rule library integrates scenario-based power baselines, user behavior adaptation rules, and battery protection constraint rules. This allows it to match the power demands of different driving scenarios while also adapting to individual user driving habits. Furthermore, by limiting parameter boundaries through battery protection constraints, a balance between adaptability and safety is achieved, overcoming the shortcomings of traditional fixed-mode systems that cannot accommodate the needs of multiple scenarios and users.

[0020] By calibrating adaptive power adjustment commands and adjusting battery energy status feedback, a coordinated adaptation of power output with user behavior, driving scenarios, and battery status is achieved. This process dynamically adjusts the assist ratio and torque change rate, optimizing power distribution based on battery energy status. While ensuring smooth riding, it rationally utilizes battery energy, further enhancing the comfort and range adaptability of the riding experience, demonstrating superior adaptability compared to traditional fixed power modes.

[0021] Compared with traditional fixed-mode power adjustment solutions, the proposed solution forms a closed-loop chain from user characteristic profiling, intent prediction, rule adaptation to dynamic calibration, breaking through the limitations of the traditional "one-size-fits-all" approach. Through multi-dimensional data fusion and precise processing, the power output can actively adapt to the individual characteristics and real-time riding needs of different users, effectively solving the technical problem of poor power adaptability in traditional solutions and significantly improving riding comfort and smoothness. Attached Figure Description

[0022] Figure 1 This application provides a method for optimizing and controlling the battery energy distribution of Ebikes in multiple scenarios.

[0023] Figure 2 This is an electronic device according to an embodiment of the present application. Detailed Implementation

[0024] like Figure 1 As shown, an Ebike user behavior prediction and personalized motivation adjustment method includes the following steps: Step 1: Collect Ebike multi-dimensional operation data, which includes user static physiological data, dynamic cycling behavior data and scene perception data. The multi-dimensional operation data is processed by feature engineering to generate the user's cycling behavior fingerprint. Step 2: Generate scene-intent prediction results by processing the cycling behavior fingerprint through multi-scale temporal modeling and intent decoding; Step 3: Call the personalized power adjustment rule library, which integrates scenario-based power baseline, user behavior adaptation rules and battery protection constraint rules, and perform dynamic calibration of power parameters on the scenario-intent prediction results to generate adaptive power adjustment instructions; Step 4: Perform power parameter calibration on the adaptive power adjustment command, and simultaneously feed back the battery energy status to the adjustment module to dynamically adjust the assist ratio and torque change rate, so as to coordinate the Ebike power output with user behavior, driving scenario and battery status.

[0025] Optionally, step 1 includes: Step 11: Clean and integrate user static physiological data, dynamic cycling behavior data and scene perception data, remove abnormal and interfering data and associate them with user identifiers to generate a set of cycling adaptation and compliance features. Step 12: Construct cycling behavior trend features and cycling parameter correlation features for the cycling adaptation compliance feature set. The cycling behavior trend features reflect the changing patterns of cycling parameters, and the cycling parameter correlation features reflect the collaborative relationship between different cycling parameters, so as to generate cycling time-series correlation feature groups. Step 13: Perform weighted adaptive fusion on the cycling time-series associated feature group to dynamically adjust the weight ratio of static physiological features, dynamic behavioral features and scene perception features according to the scene type, and generate cycling behavior fingerprint.

[0026] Optionally, step 11 includes: Step 111: Standardize the format of the user's static physiological data, unify the data measurement units and storage format, in order to generate standardized physiological data for cycling adaptation; Step 112: Perform outlier removal on dynamic cycling behavior data. Based on the behavior rationality rule base, determine and filter parameter values ​​that exceed the normal cycling range to generate regular cycling behavior data. Step 113: Remove redundant information from the scene perception data, retain the scene parameters associated with cycling behavior, and associate and bind the standardized physiological data for cycling adaptation with the regularized behavioral data for cycling to generate a set of cycling adaptation compliance features.

[0027] The technical essence of step 11 is to transform scattered user static physiological data, dynamic cycling behavior data, and scene perception data into a regular and related set of cycling adaptation and compliance features through Ebike's scenario-specific multi-dimensional data preprocessing mechanism. This is different from the traditional general data preprocessing mode that ignores the characteristics of cycling scenarios. It not only ensures data quality but also establishes a strong correlation between data and users and cycling behavior, providing a highly adaptable data foundation for subsequent feature construction and behavior fingerprint generation.

[0028] Preferably, the specific implementation process of step 111 is as follows: First, the core dimensions of the user's static physiological data are defined, including physiological parameters related to cycling power such as height, weight, and limb length. These data are entered by the user through the Ebike companion APP and associated with a unique user identifier. The user's static physiological data undergoes format standardization processing. First, a data format mapping table specific to cycling scenarios is constructed, clarifying the standard units of measurement for each physiological parameter (e.g., height in length units, weight in mass units) and a unified storage format (e.g., fixed decimal places). Then, each data entry is converted in unit and formatted according to the mapping table to generate preliminary standardized data. The preliminary standardized data undergoes validity verification. Based on the distribution range of physiological parameters of the Ebike cycling population, a verification threshold is set to remove abnormal data exceeding a reasonable range, ultimately obtaining cycling-adaptive standardized physiological data. This data ensures the comparability and usability of physiological data from different users.

[0029] Preferably, in the specific technical implementation of step 112, the dynamic cycling behavior data comes from sensors deployed on the Ebike pedal axles, bottom bracket, handlebars, and wheels, including parameters that change in real time such as cadence, pedaling force, speed, and riding posture. A behavior rationality rule library specifically designed for cycling scenarios is created. The rule library is categorized by cycling behavior type, covering cadence rationality rules, pedaling force matching rules, and speed correlation rules. For example, the cadence rationality rule sets a normal fluctuation range based on different cycling scenarios (such as flat roads and uphill climbs), and the pedaling force matching rule establishes the correlation logic between pedaling force and speed and gradient (such as pedaling force should not exceed the normal range of human force exertion at low speeds). The dynamic cycling behavior data is verified frame by frame, comparing each data point with the corresponding rule in the behavior rationality rule library. Parameter values ​​deemed abnormal (such as instantaneous cadence exceeding the normal range without corresponding speed changes) are directly filtered out, while valid data conforming to the rules is retained. This data is then integrated to generate regular cycling behavior data, which accurately reflects the user's actual cycling behavior state.

[0030] Preferably, in the specific technical implementation of step 113, the scene perception data is collected by the environmental sensors mounted on the Ebike, including parameters such as slope, road surface smoothness, ambient wind speed, and light intensity. Based on the analysis of Ebike riding behavior influencing factors, scene parameters strongly related to riding effort and power demand (such as slope and wind speed) are selected, and redundant information with less impact on riding behavior adaptation, such as light intensity, is removed to obtain a simplified scene parameter set. User identifiers are extracted from the standardized physiological data for riding adaptation, and timestamps are extracted from the regularized riding behavior data. The simplified scene parameter set is associated with the regularized riding behavior data by timestamp, and then the standardized physiological data for riding adaptation is bound to the user identifier to construct a three-dimensional associated data structure of "user physiology - riding behavior - scene parameters". Finally, a set of riding adaptation compliance features is generated. This set realizes the organic integration of multi-dimensional data and provides complete data support for subsequent feature construction.

[0031] Optionally, step 12 includes: Step 121: Calculate the rate of change of time-series parameters in the cycling adaptation compliance feature set, capture the increase or decrease trend of parameters over time, and generate cycling behavior trend features; Step 122: Perform correlation analysis on different types of cycling parameters in the cycling adaptation compliance feature set, construct the correlation mapping relationship between parameters, and generate cycling parameter correlation features; Step 123: Integrate the cycling behavior trend features and cycling parameter correlation features, classify and sort them according to data dimensions to generate cycling time-series correlation feature groups.

[0032] The technical essence of step 12 is to address the temporal dynamics and cross-dimensional correlations of parameters in the Ebike cycling scenario. Through hierarchical feature construction and structured integration, deep features with behavioral semantics are mined from the cycling adaptation and compliance feature set. Unlike the traditional general feature construction model that only pursues the number of features, the generated cycling temporal correlation feature group not only retains the temporal change pattern of parameters, but also highlights the collaborative logic between parameters, providing core feature support for the accurate generation of cycling behavior fingerprints.

[0033] Preferably, the specific implementation process of step 121 is as follows: First, time-series parameters are selected from the cycling adaptation compliance feature set, including parameters that continuously change during cycling, such as cadence, pedaling force, speed, cycling posture angle, and gradient. Each time-series parameter is associated with a timestamp and a user identifier. For each time-series parameter, a sliding time window is used for segmentation. The window length is set according to the sensitivity of the cycling parameter to change (e.g., the window length for cadence and pedaling force is shorter than the window length for speed and gradient). For example, the sliding window length for cadence is set to (0.5-1 second), and the sliding window length for speed is set to (1-2 seconds). The rate of change index is calculated for the time-series parameter within each window, including the maximum rate of change, average rate of change, and direction of change (rising, falling, stable). At the same time, the extreme value difference (the difference between the peak and the trough) and the proportion of stable segments are extracted. These indicators are concatenated in window order to form a single-parameter time-series trend vector. The dimensions of the single-parameter time-series trend vectors of all time-series parameters are unified and integrated to generate cycling behavior trend features. These features can accurately reflect the changing patterns of various dynamic parameters during the user's cycling process.

[0034] Preferably, in the specific technical implementation of step 122, the parameters in the cycling adaptation compliance feature set are first classified into three categories: dynamic behavioral parameters (cadence, pedaling force, speed, etc.), static physiological parameters (height, weight, etc.), and scene environment parameters (slope, road resistance, etc.). A parameter correlation analysis framework specific to cycling scenarios is designed, and differentiated correlation analysis logic is set for combinations of different categories of parameters: time-series cross-correlation analysis is used between dynamic behavioral parameters to capture the coordinated change relationship of different parameters in the time dimension (such as the time lag relationship between the increase in cadence and the increase in speed); partial correlation analysis is used between dynamic behavioral parameters and static physiological parameters to eliminate the interference of scene factors and focus on the influence of physiological characteristics on cycling behavior (such as the correlation between weight and pedaling force); segmented correlation analysis is used between dynamic behavioral parameters and scene environment parameters, and the correlation degree is calculated after dividing the interval according to the scene type (flat road, uphill, downhill) to avoid cross-scene correlation distortion (such as the correlation between slope and pedaling force in uphill scenarios). Based on the correlation analysis results, a parameter correlation mapping matrix is ​​constructed. The rows and columns of the matrix correspond to different cycling parameters. The matrix elements represent the correlation strength (range 0-1) and correlation type identifier (different numbers are used to distinguish positive correlation, negative correlation, and nonlinear correlation) between the two corresponding parameters. The matrix is ​​flattened into a vector form to generate cycling parameter correlation features. These features can fully reflect the synergistic relationship between different parameters.

[0035] Preferably, in the specific technical implementation of step 123, the cycling behavior trend features and cycling parameter correlation features are first standardized. The Min-Max normalization method is used to map the values ​​of the two types of features to the same range, avoiding the imbalance of feature weights during subsequent fusion due to differences in units. The standardized features are classified and sorted according to feature semantics, into two main categories: time-series trend features and parameter correlation features. The time-series trend features are sorted according to the priority of the parameters' influence on cycling intention (e.g., pedal force trend features and speed trend features take precedence over gradient trend features), and the parameter correlation features are sorted according to the strength of the correlation (e.g., cadence-speed correlation features and pedal force-gradient correlation features take precedence over height-speed correlation features). The two types of features after sorting are concatenated in the order of "temporal trend type first, parameter correlation type second" to construct a two-dimensional cycling temporal correlation feature group. The rows of this feature group represent different feature dimensions, and the columns represent different time segments or scene intervals. The elements at the intersection of rows and columns are the standardized values ​​of the corresponding features in that time segment or scene interval. This not only ensures the temporal continuity of the features but also reflects the semantic correlation of the features, providing structured and high-value feature input for the weight adaptive fusion in step 13.

[0036] Optionally, step 13 includes: Step 131: Construct a scene type recognition model to classify scene perception data to determine the current cycling scene type; Step 132: Set feature weight allocation rules for the scene type, wherein the weight ratio of relevant features is adjusted for different scenes; Step 133: Perform weighted fusion operation on each feature in the cycling time-series associated feature group according to the weight allocation rule to generate a cycling behavior fingerprint that matches the current cycling scenario type.

[0037] The technical essence of step 13 is based on the differentiated needs of Ebike riding scenarios. Through accurate scenario identification, exclusive weight rule formulation and intelligent feature fusion, a riding behavior fingerprint that is highly adapted to the current scenario is generated. Unlike the traditional fixed weight fusion mode that ignores scenario differences, it realizes the deep binding of feature fusion with scenario and user behavior, making the riding behavior fingerprint more scenario-specific and individual-identifiable, providing an accurate feature foundation for subsequent intention prediction and power adjustment.

[0038] Preferably, the specific implementation process of step 131 is as follows: First, the core scene differentiation parameters in the scene perception data are identified, including parameters strongly related to cycling power requirements such as slope, road surface smoothness, speed range, start-stop frequency, and turning frequency. A scene type recognition model specific to cycling scenarios is designed. This model includes a feature preprocessing layer, a scene feature enhancement layer, and a multi-classification output layer. The feature preprocessing layer normalizes the scene perception data to eliminate differences in parameter dimensions. The scene feature enhancement layer increases the weight of key differentiation parameters such as slope and start-stop frequency through an attention mechanism, suppressing interference from secondary parameters. The multi-classification output layer uses a designed multi-scene classifier to classify scenes into common types such as flat road commuting scenarios, uphill scenarios, downhill scenarios, leisure cycling scenarios, and fitness cycling scenarios. The model is trained using large-scale Ebike cycling scene sample data. The sample data covers scene perception parameters and corresponding scene type labels under different scenarios. During training, the attention weights and classifier thresholds are dynamically adjusted to improve the model's ability to differentiate similar scenarios. The real-time collected scene perception data is input into the trained scene type recognition model, which outputs the current cycling scene type to ensure the accuracy of scene classification and its relevance to actual cycling needs.

[0039] Preferably, in the specific technical implementation of step 132, based on the power demand characteristics of different cycling scenarios, a scenario-specific feature weight allocation rule is formulated. For flat road commuting scenarios, the core requirement is smoothness and effortlessness, with frequent starts and stops and a high demand for timely power response. Therefore, the weight of dynamic behavioral characteristics (such as cadence change trends and speed adjustment characteristics) is increased, while the weight of static physiological characteristics is appropriately reduced. For hill climbing scenarios, strong power support is required and is closely related to user weight and pedaling force output. Therefore, the weight of static physiological characteristics (such as weight) and pedaling force-related characteristics in dynamic behavioral characteristics is increased. For leisure cycling scenarios, comfort is the core, and the requirement for stable power output is high. Therefore, the weight of static physiological characteristics (such as height and limb length) and scenario perception characteristics (such as road surface smoothness) is increased. For fitness cycling scenarios, the user actively exerts force, with assistance only as a supplement. Therefore, the weight of dynamic behavioral characteristics (such as cadence stability and force exertion rhythm characteristics) is significantly increased. A feature weight allocation table is constructed, which clarifies the basic weight ratio of static physiological features, dynamic behavioral features, and scene perception features under each scenario type. At the same time, a dynamic adjustment interface is reserved so that the weight allocation can be further optimized based on user riding feedback, forming a complete feature weight allocation rule.

[0040] Preferably, in the specific technical implementation of step 133, all features in the cycling time-series associated feature group are first extracted and classified and labeled according to static physiological features, dynamic behavioral features, and scene perception features. Based on the current cycling scene type determined in step 131, the corresponding basic weight ratios of the three types of features are retrieved from the feature weight allocation rules. Then, for each specific sub-feature within each feature category, sub-weights are allocated according to their close correlation with the current scene's power requirements (e.g., in dynamic behavioral features, the sub-weight of pedaling force is higher than the sub-weight of cadence in an uphill scene). For each feature in the cycling time-series associated feature group, its category weight is multiplied by its sub-feature weight to obtain the final fusion weight of that feature. A weighted summation method is used to perform a fusion operation on all features in the cycling time-series associated feature group. The value of each feature is multiplied by its corresponding final fusion weight and then summed to obtain a one-dimensional fusion feature vector. The fused feature vector is normalized and its values ​​are mapped to a fixed interval to generate a cycling behavior fingerprint that matches the current cycling scenario. This cycling behavior fingerprint not only retains the individual user's cycling behavior characteristics, but also fully reflects the demand orientation of the current scenario, providing accurate and scenario-specific feature input for the subsequent multi-scale temporal modeling in step 2.

[0041] In the technical solution of this application, the core of step 133 is to complete feature fusion and generate cycling behavior fingerprints based on the weight allocation rules output by the scene type. The integration logic of the personalized power adjustment rule library corresponds to step 3. The two are connected by a technical link through "cycling behavior fingerprint → scene-intent prediction result". The following details the integration process of the personalized power adjustment rule library with scene-based power baseline, user behavior adaptation rules, and cycling behavior fingerprints from three aspects: technical essence, layered integration logic, and scene-based adaptation mechanism, combined with the Ebike cycling scenario: Cycling behavior fingerprint is a digital representation of a user's individual characteristics and the features of the current scenario, including weighted fusion information of static physiological features, dynamic behavioral features, and scenario perception features; scenario-based power baseline is the basic power parameter threshold under different cycling scenarios, which is the underlying constraint for power adjustment; user behavior adaptation rules are personalized parameter adjustment criteria extracted based on cycling behavior fingerprint, which is a bridge connecting the general baseline and individual needs.

[0042] The essence of the integrated personalized power adjustment rule base is to use a scenario-based power baseline as the basic framework and user behavior adaptation rules as the basis for adjustment. It transforms the feature information of riding behavior fingerprints into personalized power parameters and finally outputs adaptive power adjustment commands, which is different from the limitations of traditional rule bases that rely on a single scenario or user parameters.

[0043] Preferably, the integration process of the personalized power adjustment rule base is performed in three steps: "baseline embedding - rule mapping - fingerprint adaptation", as follows: First, common Ebike riding scenarios (flat road commuting, climbing, descending, leisure riding, fitness riding) are classified. For the power demand characteristics of each scenario, a corresponding range of basic power parameters is set to form a scenario-based power baseline.

[0044] For example, the baseline for hill climbing scenarios includes the minimum torque threshold (ensuring climbing power), the maximum assist ratio upper limit (avoiding motor overload), and the torque change rate range (preventing sudden power changes); while the baseline for flat road commuting scenarios focuses on the median assist ratio (balancing effort saving and range) and the torque stability coefficient (improving riding comfort).

[0045] The baseline parameters for the different scenarios are stored in the rule base according to the structure of "scenario type-parameter category-parameter threshold" to form the basic module of the rule base, providing constraint boundaries for subsequent personalized adjustments.

[0046] Based on the feature dimensions of cycling behavior fingerprints, we extract the correlation patterns between user behavior and power parameters, and construct user behavior adaptation rules.

[0047] Specifically, core features are extracted from cycling behavior fingerprints: static physiological features (height, weight), dynamic behavioral features (cadence change trend, pedaling stability, speed adjustment rhythm), and scene perception features (slope change, road surface smoothness). Parameter adjustment rules are set for each feature dimension.

[0048] For example, for users with larger body weight (static physiological characteristics), in the climbing scenario, the rule base triggers the "assist ratio increase rule", which increases the assist ratio according to the weight gradient within the upper limit of the scenario-based power baseline; for users with large cadence fluctuations (dynamic behavioral characteristics), the "torque change rate slowing rule" is triggered to reduce the sensitivity of torque to cadence changes and improve the stability of power output.

[0049] These rules are stored in the rule base according to the structure of "feature dimension-scenario type-adjustment range" to form the personalized adjustment module of the rule base.

[0050] After step 2 outputs the scenario-intent prediction result, the rule base first retrieves the corresponding scenario-based dynamic baseline based on the predicted scenario type to determine the range of basic dynamic parameters. Then, the cycling behavior fingerprint features corresponding to the scenario-intent prediction result are extracted and matched with the corresponding rules in the user behavior adaptation rule module. Finally, using the scenario-based power baseline as the constraint boundary, the basic parameters are adjusted in a personalized manner according to the user behavior adaptation rules: if the adjusted parameters are within the baseline range, the adjusted power parameters are directly generated; if they exceed the baseline range, the boundary is limited according to the baseline threshold, and finally, an adaptive power adjustment command is generated.

[0051] For example, for the same user in uphill and flat road scenarios, the rule base will output differentiated power parameters based on the baseline of different scenarios and the user's cycling behavior fingerprint (such as weight and pedaling force characteristics); for different users in the same scenario (such as light weight vs. heavy weight), the rule base will adjust the assist ratio and torque change rate within the same scenario baseline based on their respective cycling behavior fingerprints to ensure that the power output meets both the needs of the scenario and the individual's cycling habits.

[0052] Meanwhile, the dynamic update feature of cycling behavior fingerprints (which is continuously optimized as user cycling data accumulates) will drive the iteration of user behavior adaptation rules in the rule base, further enhancing the personalization of power adjustment.

[0053] Optionally, step 2 includes: Step 21: Perform multi-scale convolutional feature extraction on cycling behavior fingerprints. Local and global behavioral features are captured in parallel by using a first-size convolutional kernel and a second-size convolutional kernel. The size of the second-size convolutional kernel is larger than that of the first-size convolutional kernel to generate a multi-scale feature map of cycling. Step 22: Input the cycling multi-scale feature map into a bidirectional temporal network to mine the sequential dependencies of behavioral features in order to generate a cycling temporal correlation feature vector; Step 23: Perform intent decoding on the cycling time-series associated feature vector and generate scene-intent prediction results by combining scene feature matching.

[0054] Optionally, step 21 includes: Step 211: Use a first-size convolutional kernel to extract local features from the cycling behavior fingerprint to obtain the original local cycling features, and perform association enhancement processing on the original local cycling features based on the correlation of instantaneous cycling parameters to generate a cycling local feature map; Step 212: Use a second-size convolutional kernel to extract global features from the cycling behavior fingerprint to obtain the original global cycling features. The size of the second-size convolutional kernel is larger than that of the first-size convolutional kernel. Perform time-series trend aggregation processing on the original global cycling features to determine the global cycling trend features based on the behavioral change trends within continuous time periods, so as to generate a global cycling feature map. Step 213: Perform channel attention weighted fusion on the local feature map of cycling and the global feature map of cycling to generate a multi-scale feature map of cycling.

[0055] The technical essence of step 21 is to extract local instantaneous features and global trend features through differential scale convolution, targeting the temporal characteristics and multi-dimensional features of Ebike cycling behavior fingerprints. It also combines channel attention mechanism to achieve intelligent feature fusion. Unlike the traditional single-scale convolution mode that ignores the differences in feature granularity, it comprehensively captures the detailed changes and overall patterns of cycling behavior, providing multi-scale feature support with rich hierarchical information for subsequent temporal dependency mining.

[0056] Preferably, the specific implementation process of step 211 is as follows: First, the feature structure of the cycling behavior fingerprint is defined, which is in the form of a one-dimensional vector, containing a weighted fusion value of static physiological features, dynamic behavioral features, and scene perception features. Each feature dimension is associated with timestamp information. A first-size convolutional kernel is designed, the size of which is set according to the response cycle of the instantaneous cycling behavior (e.g., covering 2-3 consecutive sampling points). For example, the size of the first-size convolutional kernel is set to (2-4). This size can accurately capture local short-term behavioral features such as sudden changes in pedal frequency and instantaneous increases in pedal force. The first-size convolutional kernel is used to perform sliding convolution operations on the cycling behavior fingerprint. The convolution stride is matched with the Ebike data sampling frequency (e.g., when the sampling frequency is 10Hz, the stride is set to 1), outputting the original local cycling features. This feature reflects the changes in behavioral features within a single short-term window. A correlation matrix of instantaneous cycling parameters is constructed. The rows and columns of the matrix represent different cycling parameters (such as cadence, pedal force, and speed) within the same time window. The matrix elements represent the instantaneous correlation strength between two corresponding parameters. Based on this matrix, the original local cycling features are subjected to correlation enhancement processing. The feature dimensions with correlation strength higher than a set threshold are weighted and enhanced to generate a local cycling feature map. This map highlights the feature expression of the coordinated changes of parameters in instantaneous behavior.

[0057] Preferably, in the specific technical implementation of step 212, a second-size convolutional kernel is designed, which is larger than the first-size convolutional kernel. The size is set according to the formation cycle of the cycling behavior trend (e.g., covering 10-20 consecutive sampling points). For example, the size of the second-size convolutional kernel is set to (10-20), which can cover long-term behavior intervals such as the continuous power exertion segment during uphill climbing and the stable segment of constant speed riding. The second-size convolutional kernel is used to perform sliding convolution operations on the cycling behavior fingerprint. The convolution step size is set according to the long-term window division requirements (e.g., a step size of 5), outputting the global original cycling features, which reflect the overall behavior state within the long-term window. Temporal trend aggregation processing is performed on the global original cycling features, dividing them into continuous sub-intervals in chronological order. The direction, magnitude, and duration of feature changes within each sub-interval are calculated to determine the global cycling trend features based on the behavior change trends within continuous time periods, such as the upward trend of pedaling force features in a continuous uphill scenario and the stable trend of vehicle speed features in a flat road cruising scenario. By arranging the global trend features of cycling in chronological order, a global feature map of cycling is generated, which can clearly show the evolution of cycling behavior over a long time scale.

[0058] Preferably, in the specific technical implementation of step 213, the local cycling feature map and the global cycling feature map are first subjected to dimensional unification processing to ensure that the number of feature channels in the two types of maps is consistent, laying the foundation for fusion operation. A channel attention fusion module is designed, which includes a feature importance evaluation layer and a weighted fusion layer. The feature importance evaluation layer evaluates the contribution of each channel feature in the local cycling feature map and the global cycling feature map to the intention prediction by calculating the information entropy and variance of each feature channel. Channels with higher information entropy and variance correspond to higher contribution. For example, in the climbing scenario, the contribution of the global pedal force trend feature channel is higher than that of the local instantaneous feature channel, and in the start-stop scenario, the contribution of the local pedal frequency change feature channel is higher than that of the global trend feature channel. The weighted fusion layer assigns a corresponding fusion weight to each feature channel of the two types of maps according to the feature importance evaluation results. The feature values ​​of the corresponding channels in the local cycling feature map and the global cycling feature map are multiplied by the fusion weights respectively and then summed to obtain the fused feature channel values. All fused feature channel values ​​are arranged in their original sequence to generate a multi-scale feature map of cycling. This map retains the detailed features of local instantaneous behavior and integrates the trend features of global long-term behavior, providing comprehensive and rich feature input for the temporal dependency mining in step 22.

[0059] First, it should be noted that the core of step 212 is generating a global cycling feature map, while the intelligent feature fusion function of the channel attention mechanism is reflected in step 213. Its core is to accurately assess the importance of each channel feature in both the local and global cycling feature maps, assign differentiated weights to different feature channels to complete the fusion, ensuring that the fused features not only meet the needs of different Ebike cycling scenarios but also accurately match the details and trends of user cycling behavior. The following section, using an Ebike cycling scenario, details the specific process of this mechanism to achieve intelligent feature fusion step by step: 1. Unified preprocessing of feature map dimensions Before the channel attention mechanism performs fusion, the local and global cycling feature maps need to be standardized in terms of dimensions. The local feature map consists of instantaneous parameter correlation features within a short window, including cadence mutation feature channels and instantaneous pedal force coordination feature channels. The global feature map consists of behavioral trend features over a long time interval, including continuous uphill pedal force increase feature channels and flat road constant speed stability feature channels. The initial number of channels and feature dimensions of the two types of maps often differ. To address this issue, the map with more channels is used as a benchmark. Channel completion is performed on the map with fewer channels, and the feature values ​​of the completed channels are set to the average parameter values ​​for the corresponding scenario (e.g., in a flat road scenario, the feature values ​​of the completed channels are set to the average flat road cycling parameters; for example, the average instantaneous pedal force on a flat road is set to 30 Newtons). Simultaneously, the feature values ​​of both types of maps are normalized to the same numerical range (e.g., 0 - 1), generating standardized local and global feature maps with unified dimensions, providing a unified data foundation for subsequent attention weight allocation.

[0060] 2. Importance assessment of characteristic channels Based on the characteristics of Ebike cycling scenarios, a feature importance evaluation layer is designed to quantify the information value of each channel in both the standardized local feature map and the standardized global feature map. First, the information entropy and variance of each channel feature are calculated. Information entropy reflects the dispersion of feature values ​​within a channel, while variance reflects the fluctuation range of the feature. Combining these two factors allows us to determine the reference value of the feature for predicting cycling intentions. For example, in start-stop scenarios, the cadence change feature channel in the standardized local feature map has high information entropy and variance due to the drastic cadence fluctuations during start-stop; while the uniform speed trend feature channel in the standardized global feature map has lower information entropy and variance due to the smooth fluctuations. Then, the evaluation results are adjusted based on scenario type, with preset channel evaluation correction coefficients for different scenarios. For example, the correction coefficient for uphill scenarios is set to 1.2, and for downhill scenarios, it is set to 0.8. Multiplying the correction coefficient by the calculated information entropy and variance of each channel yields the final importance score for each channel, thus clarifying the differences in contribution of different channels in the current scenario.

[0061] 3. Differentiated fusion weight allocation Based on the final importance score of each channel, the weighted fusion layer assigns corresponding fusion weights. A fixed total weight is set, and weight values ​​are allocated according to the proportion of importance scores; channels with higher importance scores receive greater weights. For example, in a hill-climbing scenario, the continuous increasing step force feature channel in the standardized global feature map accounts for 30% of the score, so it is assigned a weight of 0.3; the instantaneous peak step force feature channel in the standardized local feature map accounts for 25% of the score, so it is assigned a weight of 0.25; while the flatness feature channel, which has low correlation with hill climbing in both types of maps, accounts for only 5% of the score, so it is assigned a weight of 0.05. For channels with scores below a set threshold, a basic weight is retained to avoid feature loss. For example, if the threshold is set to 3%, channels below this value are uniformly assigned a basic weight of 0.03, ensuring that the fused features highlight core information without omitting auxiliary features.

[0062] 4. Weighted fusion operation of dual-spectrum features After weight allocation, a channel-by-channel weighted fusion is performed on the standardized local feature map and the standardized global feature map. The feature value of each channel in the standardized local feature map is multiplied by its corresponding assigned weight to obtain the locally weighted feature channel; the feature value of each channel in the standardized global feature map is multiplied by its corresponding assigned weight to obtain the globally weighted feature channel. Then, the local weighted feature channel and the globally weighted feature channel corresponding to the location are numerically superimposed. For example, the feature values ​​of the local peak pedal force weighted channel and the globally increasing pedal force weighted channel in an uphill scenario are added point by point to obtain the fused pedal force feature channel. After performing the superposition operation on all corresponding channels in sequence, all fused channels are arranged according to their original feature temporal order to generate a multi-scale cycling feature map. This map, through differential fusion using a channel attention mechanism, retains both the detailed features of instantaneous cycling behavior and integrates the trend features of long-term cycling, providing high-quality feature input for subsequent mining of temporal dependencies in cycling behavior.

[0063] Optionally, step 22 includes: Step 221: Divide the multi-scale feature map of cycling into cycling time-series feature segments according to the time dimension; Step 222: Input the cycling time-series feature fragments into the forward and reverse paths of the bidirectional temporal network, and extract the influence of historical behavior and the trend of future behavior to obtain the bidirectional cycling path output features. Step 223: Fuse the output features of the bidirectional cycling path to construct a feature representation containing complete temporal dependencies, so as to generate a cycling temporal correlation feature vector.

[0064] Optionally, step 23 includes: Step 231: Divide the cycling behavior into stages and set the attention weight allocation ratio for different stages; Step 232: Weight the cycling time-series related feature vectors according to the assigned attention weights to enhance the expression strength of core features related to intent in order to generate a cycling intent-related weighted feature vector; Step 233: Match the weighted feature vector associated with riding intention with the preset scene feature template to identify the current riding scene and decode the user's power demand to generate scene-intention prediction results.

[0065] The technical essence of step 22 is to address the temporal continuity and correlation of Ebike cycling behavior. Through dynamic segmentation, bidirectional temporal modeling, and feature fusion, it uncovers the complete temporal dependencies hidden in the multi-scale feature map of cycling. Unlike traditional unidirectional temporal models that only focus on historical influences, the generated cycling temporal correlation feature vector can comprehensively capture the effect of historical behavior on the present, the prediction of the future by current behavior, and the bidirectional constraints between behaviors, providing feature support rich in temporal logic for subsequent intent decoding.

[0066] Preferably, the specific implementation process of step 221 is as follows: First, the structure of the cycling multi-scale feature map is analyzed. This map is in two-dimensional data form, with rows representing feature channels (such as local cadence change feature channels, global pedal force increase feature channels, etc.), and columns representing sampling points on the time series. The element at the intersection of rows and columns is the value of the corresponding feature at that sampling point. Based on the temporal characteristics of Ebike cycling behavior, a dynamic window partitioning mechanism is designed. The window length is adaptively adjusted according to the cycling scenario type. For example, the window length for urban road scenarios with frequent starts and stops is set to a shorter interval (such as 5-10 sampling points), while the window length for highway scenarios with uniform speed is set to a longer interval (such as 15-20 sampling points). The window sliding step size is set to 1 / 2 of the window length (such as 5 steps when the window length is 10). The cycling multi-scale feature map is segmented in the time dimension according to the dynamic window partitioning mechanism. Each window corresponds to a continuous cycling behavior segment. All feature channel data within each window are extracted to generate cycling time-series feature segments. Each cycling time-series feature segment retains the corresponding timestamp range and scene identifier to ensure the temporal correlation and scene correspondence of the segments.

[0067] Preferably, in the specific technical implementation of step 222, a bidirectional temporal network specifically designed for cycling scenarios is used. This network includes a forward path feature extraction layer, a reverse path feature extraction layer, and a path interaction layer. Both the forward and reverse path feature extraction layers are composed of stacked gated recurrent units (GRUs). The GRUs control the transmission and forgetting of feature information through update and reset gates, adapting to the temporal dynamic changes of cycling behavior. Cycling temporal feature segments are input into the forward path feature extraction layer in timestamp order, extracting features segment by segment from the past to the present, capturing the influence of historical cycling behavior on current behavior, such as the driving relationship between the pedal force increase feature in the previous segment and the speed feature in the current segment, generating a forward path feature vector. Cycling temporal feature segments are input into the reverse path feature extraction layer in reverse timestamp order, extracting features segment by segment from the present to the future, predicting the trend guidance of the current behavior on subsequent cycling, such as the correlation between the speed increase feature in the current segment and the cadence adjustment in the subsequent segment, generating a reverse path feature vector. The path interaction layer performs preliminary association operations on the feature vectors of the forward path and the feature vectors of the reverse path, calculates the similarity between the two types of vectors, provides an interactive basis for subsequent fusion, and finally obtains the bidirectional cycling path output features, which contain bidirectional temporal information of historical influence and future trends.

[0068] Preferably, in the specific technical implementation of step 223, a temporal feature fusion module is constructed, which includes a feature importance evaluation layer and a weighted fusion layer. The feature importance evaluation layer, considering the current cycling scenario type and behavioral stage, evaluates the contribution of the forward and reverse path feature vectors in the bidirectional cycling path output features. For example, in an uphill scenario, the historical power accumulation reflected by the forward path feature vector has a greater impact on the current power demand, and is assigned a higher contribution (e.g., 60%), while the future uphill trend prediction reflected by the reverse path feature vector is assigned a lower contribution (e.g., 40%). In a downhill scenario, the future deceleration trend demand reflected by the reverse path feature vector is assigned a higher contribution (e.g., 55%), while the forward path feature vector is assigned a lower contribution (e.g., 45%). The weighted fusion layer assigns fusion weights based on the contribution evaluation results, multiplying the forward and reverse path feature vectors by their corresponding weights and then summing the results to obtain a preliminary fused feature vector. Temporal smoothing is performed on the initial fused feature vector to eliminate abrupt interference after the fusion of features from different paths. A feature expression containing complete temporal dependencies is constructed, and a cycling temporal correlation feature vector is generated. This vector reflects both the cumulative impact of historical cycling behavior and the predictive information of future cycling trends, providing comprehensive temporal feature support for the intent decoding in step 23.

[0069] The core of this bidirectional temporal network, designed specifically for cycling scenarios, revolves around the behavioral temporal characteristics of different Ebike cycling scenarios. It creates a network architecture adapted to the specific needs of each scenario. Through customized layer structures, scenario-based parameter configurations, and targeted training optimization, the network can accurately uncover the bidirectional temporal dependencies of cycling features in different scenarios. This distinguishes it from general bidirectional temporal networks that lack scenario adaptability. The specific design is detailed below from four aspects: overall architecture construction, customization of each core layer, scenario-based parameter configuration, and a dedicated training process. 1. Construct a bidirectional temporal network architecture that fits the cycling scenario. The overall architecture comprises an input adaptation layer, a bidirectional feature extraction layer, a scene interaction layer, and an output layer, forming a complete processing chain for adapting cycling data. The input adaptation layer receives the cycling time-series feature fragments generated in step 221. These fragments contain multi-dimensional features from different scenarios. This layer first maps the feature dimensions of the fragments to the network adaptation dimensions, while embedding scene identifier codes (e.g., 001 for uphill scenarios, 002 for flat road scenarios, and 003 for downhill scenarios) to allow the network to perceive the current scene type. The bidirectional feature extraction layer, as the core layer, consists of parallel forward and reverse feature extraction branches, capturing the influence of historical behavior and future behavior trends, respectively. The scene interaction layer strengthens the scene correlation of the bidirectional features, eliminating biases in feature extraction under different scenarios. The output layer transforms the fused bidirectional features into fixed-dimensional feature vectors, providing adaptation input for subsequent generation of cycling time-series correlated feature vectors. Layers are connected through scene feature feedback links to ensure that inter-layer processing is consistent with the current cycling scenario.

[0070] 2. Customized bidirectional feature extraction core layer adapted to cycling behavior Both branches of the bidirectional feature extraction layer employ customized stacked gated loop units. The number of gated loop units is adapted to the scene complexity; for complex scenes (such as congested urban roads), the stacking number is set to more (e.g., 6 layers), while for simple scenes (such as straight roads in the suburbs), it is set to less (e.g., 3 layers). The forward feature extraction branch receives cycling time-series feature segments in forward chronological order. It filters historical key features through update gates, such as retaining consistently stable cadence features in flat road scenes and forgetting occasional pedal force fluctuations. It also adjusts the influence of historical features on the current output through reset gates, such as weakening the influence of low pedal force features from previous flat roads in the early stages of an uphill climb. The reverse feature extraction branch receives cycling time-series feature segments in reverse chronological order. Its gated loop unit adds a trend prediction submodule. This module predicts subsequent behavior trends based on the features of the current segment. For example, if it detects an increase in speed on the current downhill section, it predicts possible braking and deceleration behavior and extracts deceleration-related features in advance. Meanwhile, both side roads are embedded with cycling parameter constraint modules. These modules preset reasonable ranges for cycling parameters under different scenarios and correct features that exceed the range, ensuring that the extracted features conform to the cycling behavior patterns.

[0071] 3. Configure network parameters and interaction rules that match the characteristics of the scenario. Differentiated parameters and interaction rules are customized for different cycling scenarios. Regarding parameters, different activation functions for gated recurrent units are set for different scenarios. For example, the Leaky ReLU function is used for scenarios with frequent starts and stops to avoid gradient vanishing in low feature value regions; the tanh function is used for constant speed scenarios to adapt to smooth feature changes. Scenario-specific learning rates are set: a higher learning rate (e.g., 0.005) is set for scenarios with variable road conditions to accelerate the network's adaptation to sudden behavioral changes; a lower learning rate (e.g., 0.001) is set for scenarios with stable road conditions to ensure the stability of feature extraction. Regarding interaction rules, the scenario interaction layer sets interaction weights for bidirectional features for different scenarios. In uphill scenarios, the weight of historical effort features extracted from forward branches is set higher (e.g., 0.6), while the weight of future trend features from reverse branches is set lower (e.g., 0.4); in downhill scenarios, the weights are reversed. Simultaneously, the interaction layer calculates the scenario matching degree of the bidirectional features. When the matching degree is lower than a set threshold (e.g., 0.5), branch feature re-extraction is triggered to ensure that both bidirectional features fit the scenario.

[0072] 4. Design a personalized online training process tailored to cycling data. The training data consists of time-series feature samples from various cycling scenarios, including uphill, downhill, flat roads, and congested sections. Each scenario sample is divided into a training set and a validation set in a 1:1 ratio. During training, scenario identifiers are first bound to sample features and input into the network. The time-series correlation error of cycling behavior is used as the loss function, which is customized according to the scenario. For example, in uphill scenarios, the time-series correlation error between pedaling force and speed is emphasized, while in downhill scenarios, it is emphasized, as is the time-series correlation error between speed and braking frequency. During training, the network is tested with the validation set after each training round. If the error in a certain scenario exceeds a threshold, the training rounds for that scenario are increased (e.g., an additional 20 rounds are added). After training, a scenario-parameter mapping table is constructed to record the optimal network parameter configuration for different scenarios. During network runtime, the network can quickly call the corresponding parameters based on the input scenario identifier without retraining, significantly improving the efficiency and accuracy of feature extraction in different cycling scenarios.

[0073] User behavior adaptation rules are power parameter adjustment guidelines customized based on Ebike user riding behavior characteristics. They are designed around individual user riding habits and specifically include three types of rules: 1. Rules for adapting force application habits For users with "stable power delivery" (pedal force fluctuation <10% as shown by cycling behavior fingerprint): within the range of the assist ratio of the scenario-based power baseline, take the range of ±5% of the midpoint of the range (e.g., if the baseline range is 20%-50%, adjust it to 30%-40%) to ensure stable power output.

[0074] For users who use intermittent power (pedal force fluctuation ≥20% as shown by cycling behavior fingerprint): within the range of the assistance ratio of the scenario-based power baseline, shift 10%-15% towards the higher end (e.g., if the baseline range is 40%-70%, adjust to 50%-70%) to match the power supplementation needs of intermittent power use.

[0075] 2. Cycling Pace Adaptation Rules For users with "high-frequency start-stop" behavior (start-stop frequency ≥ 3 times / minute as displayed by fingerprint of riding behavior): Adjust the torque change rate of the scenario-based power baseline to the higher end by 10%-20% (e.g., when the baseline range is 1-1.5 N·m / s, adjust it to 1.3-1.5 N·m / s) to improve power response speed.

[0076] For "constant speed cruising" users (riding behavior fingerprint display speed fluctuation <5%): adjust the torque change rate of the scenario-based power baseline to the lower end by 10%-20% (e.g., when the baseline range is 1-1.5 N·m / s, adjust to 1-1.2 N·m / s) to ensure smooth power output.

[0077] 3. Dynamic response preference adaptation rules For users with "high response demand" (the scenario-intention prediction results show that the acceleration demand accounts for ≥60%): within the range of the assistance ratio of the scenario-based power baseline, take the higher value segment (e.g., when the baseline range is 20%-50%, adjust it to 40%-50%) to enhance the power output during the acceleration phase.

[0078] For users with "low response demand" (the scenario-intention prediction results show that the proportion of constant speed demand is ≥70%): within the range of the assistance ratio of the scenario-based power baseline, take the lower value (e.g., when the baseline range is 20%-50%, adjust it to 20%-30%) to balance power output and range.

[0079] The personalized power adjustment rule base integrates these three elements through the logic of "hierarchical constraints + priority adaptation". The specific process is as follows: 1. Underlying Constraints: Scenario-Based Dynamic Baseline The rule base is based on a scenario-based power baseline framework. It prioritizes matching the corresponding baseline parameters (assist ratio range, torque change rate range, etc.) according to the current riding scenario type to determine the basic range of power parameters and ensure that the power output matches the core needs of the scenario (such as high assist in climbing scenarios and low assist in downhill scenarios).

[0080] 2. Mid-layer adjustments: Optimization of user behavior adaptation rules Within the baseline of scenario-based power, user behavior adaptation rules are invoked, and user riding behavior fingerprints and short-term power needs are combined to make personalized adjustments to the baseline parameters (such as adjusting the median assist ratio for users with steady power delivery and increasing the torque rate for users with frequent start-stop cycles), so that the power parameters match the individual user's habits.

[0081] 3. Top-level constraints: Battery safety regulations as a safety net The power parameters, after being adjusted according to the scenario baseline and user rules, are verified by battery safety constraints (such as cell voltage, remaining power, and charging / discharging power limits). If the parameters exceed the battery safety range, they are adjusted to within the safety threshold (such as the upper limit of compression assist ratio when the battery is low). Finally, an adaptive power adjustment command is generated that fits the scenario and user while ensuring battery safety.

[0082] The integration logic of the three presents a priority of "scenario foundation → user optimization → safety safeguard", which not only meets the experience needs of scenarios and users, but also ensures the safe operation of the battery.

[0083] Optionally, step 3 includes: Step 31: Match the scenario type in the scenario-intention prediction result with the scenario-based power baseline in the personalized power adjustment rule base to determine the range of basic power parameters for riding; Step 32: Based on the range of basic cycling power parameters, call the user behavior adaptation rules to adjust the power demand of the user in the scenario-intent prediction results to obtain the adjusted cycling power parameters by adjusting the assist ratio and torque change rate parameters. Step 33: Limit the power parameters after riding adjustment according to the battery protection constraint rules to generate an adaptive power adjustment command.

[0084] The technical essence of step 31 is based on the differences in power requirements of different riding scenarios of Ebike. By accurately matching the scenario-based power baseline, it provides a basic parameter framework that fits the characteristics of the scenario for subsequent personalized power adjustment. This is different from the traditional mode of ignoring scenario differences with a single power baseline. It allows the initial range of power adjustment to not only adapt to the core needs of the scenario, but also reserve reasonable space for users to make personalized adjustments.

[0085] Preferably, the specific implementation process of step 31 is as follows: First, a scenario-based power baseline library is constructed. This baseline library is classified according to common Ebike riding scenarios, including five core scenarios: flat road commuting scenario, uphill scenario, downhill scenario, leisure riding scenario, and fitness riding scenario. Each scenario corresponds to an independent scenario-based power baseline entry. Each entry includes three core power parameter items: scenario label, assist ratio range, torque change rate range, and maximum output power threshold. The parameter settings for scenario-based power baselines are based on the power characteristics of different riding scenarios: For flat road commuting scenarios, to accommodate frequent starts and stops, the assist ratio range should be set to a medium range (e.g., 20%-50%), the torque change rate range to a relatively sensitive range (e.g., 1-1.5 N·m / s), and the maximum output power threshold to a medium level (e.g., 150W); for hill climbing scenarios, to provide continuous power support, the assist ratio range should be set to a high range (e.g., 40%-70%), the torque change rate range to a gradual range (e.g., 0.5-1 N·m / s), and the maximum output power threshold to a high level (e.g., 200W); for downhill scenarios, to reduce power redundancy, the assist ratio range should be set to a low range (e.g., 0%-20%), and the torque change rate range to a relatively sensitive range (e.g., 1-1.5 N·m / s), and the maximum output power threshold to a high level (e.g., 200W). The torque change rate range is set to a low-speed range (e.g., 0.3-0.8 N·m / s), and the maximum output power threshold is set to a low level (e.g., 100W). For leisure cycling scenarios, comfort and stability are required, so the assist ratio range is set to a low-to-medium range (e.g., 10%-40%), the torque change rate range is set to a gradual range (e.g., 0.6-1.2 N·m / s), and the maximum output power threshold is set to a medium level (e.g., 120W). For fitness cycling scenarios, space for user active force application is required, so the assist ratio range is set to a low-to-medium range (e.g., 10%-30%), the torque change rate range is set to a medium range (e.g., 0.8-1.3 N·m / s), and the maximum output power threshold is set to a medium level (e.g., 150W). The scenario type field is extracted from the scenario-intent prediction results. This field is precisely matched with the scenario tags in the scenario-based power baseline library. The corresponding scenario-based power baseline entry is retrieved, and the assist ratio range, torque change rate range, and maximum output power threshold in the entry are integrated to generate the basic cycling power parameter range. This range serves as the basic constraint framework for adjusting the user behavior adaptation rules in step 32.

[0086] The technical essence of step 32 is based on the differentiated characteristics of individual riding behavior of Ebike users. Within the basic parameter range of the scenario-based power baseline, the power parameters are finely adjusted through customized user behavior adaptation rules. This is different from the traditional uniform power adjustment mode that ignores user habits. It allows the power output to not only meet the needs of the scenario, but also match the individual characteristics of the user such as the force exertion rhythm and response preferences.

[0087] Preferably, the specific implementation process of step 32 is as follows: First, a user behavior adaptation rule base is constructed. This rule base is divided into two core rules according to the user's cycling behavior characteristics: force exertion habit adaptation rules and cycling rhythm adaptation rules. Each type of rule corresponds to the parameter adjustment logic under different scenarios. The user's power demand information is extracted from the scenario-intent prediction results. This information includes the user's short-term force exertion intensity demand (such as "high force exertion demand" and "stable force exertion demand") and response speed demand (such as "high response demand" and "low response demand"). At the same time, it is associated with the corresponding user's cycling behavior fingerprint features (such as cadence fluctuation amplitude and pedal force output mode). Taking a flat road commuting scenario as an example, if the user's power demand is "high response demand" and the cycling behavior fingerprint shows a cadence fluctuation range of ≥20% (corresponding to an "intermittent power type" user), then the cycling rhythm adaptation rule will be invoked. Within the basic power parameter range for flat road commuting scenarios (assistance ratio 20%-50%, torque change rate 1-1.5 N·m / s), the assist ratio will be adjusted by 10%-15% towards the higher end of the range (e.g., adjusted to 35%-50%), and the torque change rate will be adjusted towards the lower end of the range. For high-value settings, adjust by 10%-20% (e.g., to 1.3-1.5 N·m / s). If the user's power requirement is "stable power delivery" and the cycling behavior fingerprint shows a cadence fluctuation of <10% (corresponding to a "stable power delivery" user), then apply the power delivery habit adaptation rules, adjusting the assist ratio to within ±5% of the midpoint of the range (e.g., to 30%-40%), and the torque change rate to within ±10% of the midpoint of the range (e.g., to 1.1-1.3 N·m / s). After adjusting the parameters according to the matched user behavior adaptation rules, integrate the adjusted assist ratio and torque change rate parameters to generate the adjusted cycling power parameters. These parameters conform to the basic constraints of the scenario-based power baseline while also reflecting the user's individual cycling habits.

[0088] The technical essence of step 33 is to limit the power parameters to a safe boundary through battery protection constraint rules based on scenario adaptation and user personalization, so as to achieve coordinated adaptation between Ebike power output and battery safety status. This is different from the traditional power adjustment mode that only focuses on experience and ignores battery life. It ensures the riding power demand and avoids battery damage due to overcharging, over-discharging or power overload.

[0089] Preferably, the specific implementation process of step 33 is as follows: First, a battery protection constraint rule library is constructed. This rule library contains three core rules: cell voltage protection rules, remaining capacity adaptation rules, and charge / discharge power limitation rules. Each rule corresponds to a key threshold for safe battery operation. The cell voltage protection rules set the highest safe voltage (e.g., 4.2V / cell) and the lowest safe voltage (e.g., 3.2V / cell) for the cell. The remaining capacity adaptation rules divide different power parameter constraint ranges according to the remaining capacity percentage, such as ≥50% remaining capacity as the sufficient range, 20%-50% as the medium range, and <20% as the low capacity range. The charge / discharge power limitation rules set the rated maximum discharge power (e.g., 200W) and the rated maximum charging power (e.g., 50W) of the battery. The battery status data collected in real time by the battery management system is extracted, including the current cell voltage, remaining capacity percentage, and real-time charge / discharge power. This data is used as the basis for determining battery protection constraints.

[0090] Taking the adjusted power parameters for a hill-climbing scenario as an example, if the discharge power corresponding to this parameter is 220W, exceeding the rated maximum discharge power (200W), then the charging and discharging power limiting rule is invoked to reduce the torque change rate by 10%-15% (e.g., from 1N·m / s to 0.85-0.9N·m / s), so that the discharge power falls back to the rated range; if the remaining battery charge is 15% (in the low charge range), and the assist ratio of the adjusted power parameters is 60% (at the high end of the baseline for the hill-climbing scenario), then the remaining charge adaptation rule is invoked to compress the assist ratio to the low to medium range of the baseline range (e.g., adjusted to 40%-50%); if the real-time cell voltage is 3.1V (below the minimum safe voltage), then the cell voltage protection rule is invoked to reduce the assist ratio to the lowest value of the baseline range (e.g., 20%), and at the same time reduce the torque change rate to the lowest value of the baseline range (e.g., 0.5N·m / s).

[0091] After verifying and adjusting the battery protection constraints for each of the power parameters after riding adjustments, the final assist ratio, torque change rate, and charging / discharging parameters are integrated to generate an adaptive power adjustment command. This command not only meets the power needs of the current scenario and the user, but also complies with the constraints of battery safe operation.

[0092] Optionally, step 4 includes: Step 41: Perform hierarchical analysis on the charging and discharging parameters, assist ratio parameters, and torque change rate parameters in the adaptive power adjustment command, and extract the reference values ​​and adjustment thresholds of each parameter to generate a power adjustment parameter analysis set; Step 42: Real-time collection of remaining battery power, cell voltage consistency and energy pool working status, integration to form a battery energy state dataset, and feeding the battery energy state dataset back to the adjustment module to trigger the parameter dynamic calibration logic; Step 43: Based on the battery energy state dataset and the power adjustment parameter analysis set, construct a "demand-supply" adaptation model. When the battery has sufficient remaining power, adjust the assist ratio according to user behavior characteristics to match the power exertion habit. When the battery has low remaining power, optimize the torque change rate while maintaining the basic power. Step 44: Based on the road conditions of the current driving scenario, perform scenario adaptation verification on the adjusted assist ratio and torque change rate, and measure the smoothness of the power output connection in different stages of start-stop, acceleration, and constant speed, so as to achieve dynamic collaborative adaptation between Ebike power output and user behavior, driving scenario and battery status.

[0093] The technical essence of step 41 is to perform layered analysis and boundary extraction of the core parameters in the adaptive power adjustment command, providing a structured parameter benchmark framework for subsequent dynamic calibration combined with battery status. Unlike the traditional parameter analysis mode that only extracts a single value, by clarifying the benchmark value and adjustment threshold of each parameter, it not only preserves the adjustable space of the parameters, but also ensures that the adjustment process conforms to the power demand characteristics of different riding scenarios of Ebike.

[0094] Preferably, the specific implementation process of step 41 is as follows: First, the three core parameters included in the adaptive power adjustment command are clarified: charging and discharging parameters, assist ratio parameters, and torque change rate parameters. Each type of parameter is associated with the current riding scenario type (such as flat road commuting and hill climbing). A hierarchical parsing mechanism is adopted, and parsing is performed sequentially according to parameter type: First, the charging and discharging parameters are parsed. This parameter includes two sub-parameters: the upper limit of discharge power and the upper limit of charging power. The specific value of the upper limit of discharge power is extracted from the adaptive power adjustment command as the baseline value of the charging and discharging parameters. At the same time, its adjustment threshold is determined based on the battery's rated power range (e.g., when the upper limit of discharge power is 200W, the adjustment threshold is set to ±20W). Next, the assist ratio parameter is parsed. The current value of the assist ratio is extracted from the adaptive power adjustment command as the baseline value of the assist ratio parameter. Its adjustment threshold is determined in combination with the power baseline range of the corresponding scenario (e.g., when the assist ratio baseline value is 50% in the climbing scenario, the adjustment threshold is set to ±10%). Finally, the torque change rate parameter is parsed. The current value of the torque change rate is extracted as the baseline value of the torque change rate parameter. Its adjustment threshold is determined in combination with the adjustment range of the user behavior adaptation rules (e.g., when the torque change rate baseline value is 1N·m / s, the adjustment threshold is set to ±0.2N·m / s).

[0095] Taking a flat road commuting scenario as an example, the charging and discharging parameters in the adaptive power adjustment command show an upper limit of 150W for discharge power, a boost ratio parameter of 30%, and a torque change rate parameter of 1.2N·m / s. When analyzing the charging and discharging parameters, 150W is taken as the baseline value, and the adjustment threshold is set to ±15W (based on a 10% fluctuation range of the battery's rated discharge power of 150W). When analyzing the boost ratio parameter, 30% is taken as the baseline value, and the adjustment threshold is set to ±8% (based on a 20%-50% fluctuation range of the power baseline in a flat road scenario). When analyzing the torque change rate parameter, 1.2N·m / s is taken as the baseline value, and the adjustment threshold is set to ±0.3N·m / s (based on a 1-1.5N·m / s fluctuation range of the torque change rate in a flat road scenario). The baseline values ​​and adjustment thresholds of the three types of parameters are integrated according to the structure of "parameter type-baseline value-adjustment threshold" to generate a dynamic adjustment parameter analysis set. This set provides a clear basis for parameter adjustment for the "demand-supply" adaptation model in step 43.

[0096] The technical essence of step 42 is to capture and structure the core operating status parameters of the battery in real time, and trigger a dynamic calibration mechanism through status feedback. This provides real-time basis at the battery level for the accurate adaptation of Ebike's power parameters. Unlike the traditional fixed-period collection mode that ignores scenario differences, this method optimizes the collection strategy by combining the characteristics of riding scenarios, ensuring both data timeliness and the synergy between battery status feedback and power adjustment.

[0097] Preferably, the specific implementation process of step 42 is as follows: First, configure the battery status acquisition module, which includes a remaining power detection unit, a cell voltage monitoring unit, and an energy pool working status sensing unit. Each unit operates according to a scenario-based acquisition frequency: for high-intensity power output scenarios such as climbing and acceleration, the acquisition frequency is set to a higher level (e.g., 50ms / time) to ensure that rapid changes in battery status are captured; for low-intensity scenarios such as flat road constant speed and downhill, the acquisition frequency is set to a medium level (e.g., 100ms / time) to balance data timeliness and energy consumption; for low-power scenarios where the remaining battery power is <20%, the acquisition frequency is automatically increased to a high level (e.g., 40ms / time) to enhance real-time monitoring of battery status. The remaining battery capacity detection unit collects the battery's remaining capacity percentage using coulomb counting and simultaneously records the rate of capacity change (e.g., the rate of capacity increase during charging and the rate of capacity consumption during discharging). The cell voltage monitoring unit collects the real-time voltage values ​​of all individual cells, calculates the maximum, minimum, and voltage difference (the difference between the maximum and minimum values) to characterize cell voltage consistency. The energy pool operating status sensing unit collects battery temperature (e.g., cell surface temperature, battery pack internal temperature) and charging / discharging circuit current, combines the voltage data to calculate real-time charging / discharging power, and determines the current operating mode of the energy pool (charging mode, discharging mode, standby mode). The collected parameters, including remaining battery capacity percentage, rate of capacity change, individual cell voltage values, cell voltage differences, battery temperature, real-time charging / discharging power, and operating mode, are linked and integrated according to the structure of "state type - parameter value - collection timestamp" to generate a battery energy state dataset. The dataset is fed back to the adjustment module via a data transmission link. The adjustment module has a built-in state threshold judgment logic. If the parameters in the dataset exceed the preset safety threshold (such as cell voltage difference ≥ 0.3V, battery temperature ≥ 55℃, and remaining charge < 10%), the parameter dynamic calibration logic is directly triggered. If the parameters are within the safe range, the calibration logic is also triggered synchronously according to the collection cycle to ensure that the power parameters are always adapted to the real-time battery status. Taking the hill climbing scenario as an example, if the remaining charge is 18% (below the 20% low charge threshold), the cell voltage difference is 0.25V (close to the 0.3V safety threshold), and the battery temperature is 52℃ (close to the 55℃ safety threshold), after integrating and generating the battery energy state dataset and feeding it back to the adjustment module, the emergency dynamic calibration logic is triggered, providing an emergency state basis for parameter optimization in the subsequent step 43.

[0098] The technical essence of step 43 is to build a dynamic adaptation mechanism of "user power demand - battery energy supply". By distinguishing the remaining battery power status, the assist ratio and torque change rate are optimized in a targeted manner. This ensures a personalized riding experience for users when the battery energy is sufficient, and balances basic power and battery range when the battery power is low. Unlike the traditional single-dimensional parameter adjustment mode, this achieves three-way coordination of power output, user habits and battery status.

[0099] Preferably, the specific implementation process of step 43 is as follows: First, it is clarified that the core input of the "demand-supply" adaptation model is the battery energy state dataset and the power adjustment parameter analysis set. The model has a built-in dual-branch adjustment logic, corresponding to two scenarios: sufficient battery remaining power and low remaining power. The remaining power percentage is extracted from the battery energy state dataset as the core judgment indicator. A high percentage threshold is set as sufficient state (e.g., ≥50%), and a low percentage threshold is set as low remaining power (e.g., <50%). The assist ratio benchmark value, assist ratio adjustment threshold, torque change rate benchmark value, and torque change rate adjustment threshold are extracted from the power adjustment parameter analysis set as the basic range for parameter adjustment. At the same time, the user's riding behavior fingerprint features are associated, focusing on extracting the power application habits (e.g., steady power application, intermittent power application) and response preferences (e.g., high response, low response) features.

[0100] When the battery is determined to be fully charged, the model triggers a branch that prioritizes adjusting the assist ratio. Based on the user's exertion habits and response preferences, personalized adjustments are made within the assist ratio adjustment threshold range: If the user is an intermittent exertor with a high response preference, the assist ratio baseline value is adjusted to a smaller range towards the higher end of the adjustment threshold (e.g., 5%-10%, for example, when the baseline value is 30% and the adjustment threshold is ±8%, it is adjusted to 35%-38%); if the user is a steady exertor with a low response preference, the assist ratio baseline value is adjusted towards the middle end of the adjustment threshold, maintaining it within a narrow range near the baseline value (e.g., ±3%, for example, when the baseline value is 30%, it is adjusted to 27%-33%). At this time, the torque change rate remains unchanged at the baseline value. Only when the power connection becomes uneven after the assist ratio adjustment is a very small adjustment (e.g., ±5%) is made within the adjustment threshold range to ensure that the power output matches the user's exertion rhythm.

[0101] When the model determines that the remaining battery power is low, it triggers a torque change rate optimization branch. This branch prioritizes ensuring the assist ratio is not lower than the minimum required for the scenario's basic power needs (e.g., a lower minimum for flat roads, example 20%; a higher minimum for uphill scenarios, example 40%). The assist ratio is locked between this minimum and the baseline value to prevent excessively low assist ratios from negatively impacting the riding experience. Simultaneously, based on the battery energy state dataset's data on power consumption rate and cell voltage consistency, the torque change rate is optimized: if the power consumption rate is fast (e.g., reaching a set rate threshold, example ≥2% / min) and cell voltage consistency is good (e.g., voltage difference lower than the set consistency value), the torque change rate is optimized. If the torque variation rate reference value is within a certain range (e.g., 10%-15%, for example, when the reference value is 1.2 N·m / s and the adjustment threshold is ±0.3 N·m / s, adjust it to 1.02-1.08 N·m / s) to reduce the fluctuation of power output and reduce power consumption; if the cell voltage consistency is poor (e.g., the voltage difference reaches the set consistency threshold, for example, ≥0.2 V), adjust the torque variation rate reference value to a low to medium range of the adjustment threshold (e.g., when the reference value is 1.2 N·m / s, adjust it to 1.0-1.1 N·m / s) to avoid further widening of cell voltage differences due to torque fluctuations.

[0102] Taking a flat road commuting scenario as an example, if the battery's remaining charge reaches the sufficient threshold (e.g., 60%), and the user is a high-frequency start-stop user (the start-stop frequency displayed by the riding behavior fingerprint reaches the set frequency threshold, for example, ≥3 times / minute), the power adjustment parameter analysis will set the assist ratio baseline value to a certain intermediate value (e.g., 30%), the assist ratio adjustment threshold to a symmetrical range (e.g., ±8%), and the torque change rate baseline value to a normal value (e.g., 1.2 N·m / s), the adjustment threshold to a symmetrical range (e.g., ±0.3 N·m / s). The model will adjust the assist ratio to a range close to the high end of the adjustment threshold (e.g., 36%-3). If the remaining battery charge is low (e.g., 30%), the minimum basic power in a flat road scenario is relatively low (e.g., 20%), the power consumption rate reaches a relatively fast threshold (e.g., 1.8% / min), and the cell voltage difference is within a good range (e.g., 0.15V), the model locks the assist ratio in the range between the baseline and the minimum value (e.g., 25%-30%), and adjusts the torque change rate to a medium-low range (e.g., 1.05-1.1N·m / s) to achieve a balance between basic power guarantee and low power consumption optimization.

[0103] After the adjustment is completed, the final assist ratio and torque change rate parameters are integrated to generate dynamically calibrated power parameters, providing input for the scenario adaptation verification in step 44.

[0104] The technical essence of step 44 is to build a scenario-based adaptation and verification mechanism for road condition characteristics and power parameters. By distinguishing the road condition characteristics of different driving scenarios, the smoothness of power output connection in the start-stop, acceleration, and constant speed stages is specifically verified. This ensures that the safety requirements of the scenario are met, while also conforming to the user's riding behavior and battery energy status. This is different from the traditional unified power connection verification mode and achieves refined adaptation of power output under multi-dimensional collaboration.

[0105] Preferably, the specific implementation process of step 44 is as follows: First, extract the road condition features of the current driving scenario from the scene perception data. The road condition features include four core dimensions: road surface slope level, road surface smoothness, traffic flow density, and road type identification. Among them, the road surface slope level is divided into three levels according to the slope size: gentle, mild, and steep. The road surface smoothness is divided into three levels according to the degree of bumpiness: smooth, slightly bumpy, and severely bumpy. The traffic flow density is divided into three levels according to the number of vehicles passing through per unit time: low flow, medium flow, and high flow. The road type identification includes types such as non-motorized vehicle lanes, mixed lanes, and buffer zones around schools. Design a scene adaptation verification model. This model has built-in power connection smoothness judgment criteria corresponding to different road condition features. The smoothness judgment criteria take the power parameter change rate threshold and the stage switching response time threshold as core indicators, and the threshold settings are different in different scenarios: for scenarios with high traffic or buffer zones around schools, which have high safety requirements, a lower power parameter change rate threshold and a shorter stage switching response time threshold are set; for scenarios with high requirements for power stability, such as steep slopes or severely bumpy roads, a lower power parameter fluctuation threshold is set.

[0106] The dynamically calibrated assist ratio and torque change rate are input into the scenario adaptation verification model, along with extracted road condition features. The model is verified in three riding phases: start-stop, acceleration, and constant speed. In the start-stop phase, the focus is on verifying the adaptability of the torque change rate to the road surface gradient and traffic flow density. If the scenario is high-flow and the torque change rate exceeds the threshold for that scenario, the torque change rate is adjusted down by a certain percentage within the threshold range (e.g., from a certain value to a lower value in the example, the specific adjustment depends on the threshold) to ensure a smooth and abrupt start-stop process. In the acceleration phase, the rate of increase of the assist ratio is verified. The efficiency ratio is adapted to the road surface smoothness and road type. If the road surface is severely bumpy and the power assist ratio increases too quickly, causing power fluctuations to exceed the threshold, the rate of increase of the power assist ratio is slowed down so that the power assist ratio gradually reaches the target value over a longer period of time (e.g., in the example, the increase is adjusted from a short time to a longer time). During the constant speed phase, the stability of the power assist ratio is verified to adapt to the road surface slope level and traffic flow density. If the power assist ratio fluctuation exceeds the threshold when driving on a gentle slope and at a constant speed, the dynamic adjustment range of the power assist ratio is narrowed to maintain it within a smaller fluctuation range (e.g., in the example, the adjustment is adjusted from a larger fluctuation range to a smaller fluctuation range).

[0107] Taking the buffer zone around a school as an example, the road conditions are characterized by low slope, flat road surface, and high traffic flow. The threshold values ​​for torque change rate during the start-stop phase, the acceleration phase assist ratio growth rate during acceleration, and the constant speed phase assist ratio fluctuation threshold are set to a relatively low value (e.g., 1.0 N·m / s in the example), a relatively low value (e.g., 5% / s in the example), and a relatively small range (e.g., ±2% in the example) for the dynamic adaptation verification model. If the dynamically calibrated torque change rate is higher than the threshold for the start-stop phase and the assist ratio growth rate is higher than the threshold for the acceleration phase, the model will lower the torque change rate to within the threshold range and slow down the assist ratio growth rate to within the threshold range. After verification, the final power output parameters after scenario adaptation are generated. If, during the verification process, it is found that a parameter in a certain stage exceeds the threshold and the adaptation requirements cannot be met after adjustment, the model returns to step 43 to re-execute the parameter optimization of the "demand-supply" adaptation model until the smoothness of the power output in all three stages meets the adaptation requirements of the current road conditions, ultimately achieving dynamic collaborative adaptation of Ebike power output with user behavior, driving scenarios, and battery status.

[0108] This application provides a computer-readable storage medium storing at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the method described in any of the foregoing embodiments.

[0109] like Figure 2 As shown in the embodiment of this application, an electronic device is provided. The electronic device includes a processor and a memory. The memory stores at least one instruction or at least one program. The at least one instruction or the at least one program is loaded and executed by the processor to implement the method described in any of the above embodiments.

[0110] This application provides a computer program product, which includes computer instructions that, when executed by a processor, implement the method described in any of the above embodiments.

Claims

1. A method for predicting Ebike user behavior and personalizing motivation adjustment, characterized in that, Includes the following steps: Step 1: Collect Ebike multi-dimensional operation data, which includes user static physiological data, dynamic cycling behavior data, and scene perception data. The multi-dimensional operation data is processed by feature engineering to generate the user's cycling behavior fingerprint. Step 2: Generate scene-intent prediction results by processing the cycling behavior fingerprint through multi-scale temporal modeling and intent decoding; Step 3: Call the personalized power adjustment rule library, which integrates scenario-based power baseline, user behavior adaptation rules and battery protection constraint rules, and perform dynamic calibration of power parameters on the scenario-intent prediction results to generate adaptive power adjustment instructions; Step 4: Perform power parameter calibration on the adaptive power adjustment command, and simultaneously feed back the battery energy status to the adjustment module to dynamically adjust the assist ratio and torque change rate, so as to coordinate the Ebike power output with user behavior, driving scenario and battery status. Step 1 includes: Step 11: Clean and integrate user static physiological data, dynamic cycling behavior data and scene perception data, remove abnormal and interfering data and associate them with user identifiers to generate a set of cycling adaptation and compliance features. Step 12: Construct cycling behavior trend features and cycling parameter correlation features for the cycling adaptation compliance feature set. The cycling behavior trend features reflect the changing patterns of cycling parameters, and the cycling parameter correlation features reflect the collaborative relationship between different cycling parameters, so as to generate cycling time-series correlation feature groups. Step 13: Perform weighted adaptive fusion on the cycling time-series associated feature group to dynamically adjust the weight ratio of static physiological features, dynamic behavioral features and scene perception features according to the scene type, and generate cycling behavior fingerprint; Step 2 includes: Step 21: Perform multi-scale convolutional feature extraction on cycling behavior fingerprints. Local and global behavioral features are captured in parallel by using a first-size convolutional kernel and a second-size convolutional kernel. The size of the second-size convolutional kernel is larger than that of the first-size convolutional kernel to generate a multi-scale feature map of cycling. Step 22: Input the cycling multi-scale feature map into a bidirectional temporal network to mine the sequential dependencies of behavioral features in order to generate a cycling temporal correlation feature vector; Step 23: Perform intent decoding on the cycling time-series associated feature vector and generate scene-intent prediction results by combining scene feature matching.

2. The method according to claim 1, characterized in that, Step 11 includes: Step 111: Standardize the format of the user's static physiological data, unify the data measurement units and storage format, in order to generate standardized physiological data for cycling adaptation; Step 112: Perform outlier removal on dynamic cycling behavior data. Based on the behavior rationality rule base, determine and filter parameter values ​​that exceed the normal cycling range to generate regular cycling behavior data. Step 113: Remove redundant information from the scene perception data, retain the scene parameters associated with cycling behavior, and associate and bind the standardized physiological data for cycling adaptation with the regularized behavioral data for cycling to generate a set of cycling adaptation compliance features.

3. The method according to claim 1, characterized in that, Step 12 includes: Step 121: Calculate the rate of change of time-series parameters in the cycling adaptation compliance feature set, capture the increase or decrease trend of parameters over time, and generate cycling behavior trend features; Step 122: Perform correlation analysis on different types of cycling parameters in the cycling adaptation compliance feature set, construct the correlation mapping relationship between parameters, and generate cycling parameter correlation features; Step 123: Integrate the cycling behavior trend features and cycling parameter correlation features, classify and sort them according to data dimensions to generate cycling time-series correlation feature groups.

4. The method according to claim 1, characterized in that, Step 13 includes: Step 131: Construct a scene type recognition model to classify scene perception data to determine the current cycling scene type; Step 132: Set feature weight allocation rules for the scene type, wherein the weight ratio of relevant features is adjusted for different scenes; Step 133: Perform weighted fusion operation on each feature in the cycling time-series associated feature group according to the weight allocation rules to generate a cycling behavior fingerprint that matches the current cycling scenario type.

5. The method according to claim 1, characterized in that, Step 21 includes: Step 211: Use a first-size convolutional kernel to extract local features from the cycling behavior fingerprint to obtain the original local cycling features, and perform association enhancement processing on the original local cycling features based on the correlation of instantaneous cycling parameters to generate a cycling local feature map; Step 212: Use a second-size convolutional kernel to extract global features from the cycling behavior fingerprint to obtain the original global cycling features. The size of the second-size convolutional kernel is larger than that of the first-size convolutional kernel. Perform time-series trend aggregation processing on the original global cycling features to determine the global cycling trend features based on the behavioral change trends within continuous time periods, so as to generate a global cycling feature map. Step 213: Perform channel attention weighted fusion on the local feature map of cycling and the global feature map of cycling to generate a multi-scale feature map of cycling.

6. The method according to claim 1, characterized in that, Step 22 includes: Step 221: Decompose the multi-scale feature map of cycling into cycling time-series feature segments according to the time dimension; Step 222: Input the cycling time-series feature fragments into the forward and reverse paths of the bidirectional temporal network, and extract the influence of historical behavior and the trend of future behavior to obtain the bidirectional cycling path output features. Step 223: Fuse the output features of the bidirectional cycling path to construct a feature representation containing complete temporal dependencies, so as to generate a cycling temporal correlation feature vector.

7. The method according to claim 1, characterized in that, Step 3 includes: Step 31: Match the scenario type in the scenario-intention prediction result with the scenario-based power baseline in the personalized power adjustment rule base to determine the range of basic power parameters for riding; Step 32: Based on the range of basic cycling power parameters, call the user behavior adaptation rules to adjust the assist ratio and torque change rate parameters to obtain the adjusted cycling power parameters based on the user power demand in the scenario-intent prediction results. Step 33: Limit the power parameters after riding adjustment according to the battery protection constraint rules to generate an adaptive power adjustment command.

8. The method according to claim 1, characterized in that, Step 4 includes: Step 41: Perform hierarchical analysis on the charging and discharging parameters, assist ratio parameters, and torque change rate parameters in the adaptive power adjustment command, and extract the reference values ​​and adjustment thresholds of each parameter to generate a power adjustment parameter analysis set; Step 42: Real-time collection of remaining battery power, cell voltage consistency and energy pool working status, integration to form a battery energy state dataset, and feeding the battery energy state dataset back to the adjustment module to trigger the parameter dynamic calibration logic; Step 43: Based on the battery energy state dataset and the power adjustment parameter analysis set, construct a "demand-supply" adaptation model. When the battery has sufficient remaining power, adjust the assist ratio according to user behavior characteristics to match the power exertion habit. When the battery has low remaining power, optimize the torque change rate while maintaining the basic power. Step 44: Based on the road conditions of the current driving scenario, perform scenario adaptation verification on the adjusted assist ratio and torque change rate, and measure the smoothness of the power output connection in different stages of start-stop, acceleration, and constant speed, so as to achieve dynamic collaborative adaptation between Ebike power output and user behavior, driving scenario and battery status.

Images