Dynamic environment adaptive sunroof sunshade curtain intelligent control method and system based on multi-sensor fusion

Abstract

The application discloses a kind of dynamic environment adaptive sunroof sunshade curtain intelligent control method and system based on multi-sensor fusion, first output parameter set is generated based on real-time collected data as first level real-time environment driving response, and the bottom line output parameter corresponding to the safety bottom line, the comfortable bottom line and the energy-saving bottom line is output;Using bidirectional LSTM network combined with attention mechanism, with historical user behavior data, real-time environmental data and real-time vehicle state data as input, generate the second output parameter set corresponding to the predicted user preference as the second level habit learning prediction response;Determine whether the second output parameter set is lower than at least one of the bottom line output parameter;If yes, then use the first output parameter set to dynamically adjust the second output parameter set and generate sunshade curtain control instruction;The opening and closing parameter control of the sunshade curtain is realized by the sub-area sunshade curtain execution mechanism according to the sunshade curtain control instruction.

Description

Technical Field

[0001] This invention relates to the field of automotive intelligent control technology, specifically to a multimodal sunroof sunshade control system and method that combines in-vehicle environmental parameters, user behavior habits, and dynamic weather prediction, which is suitable for improving in-vehicle comfort and energy-saving effects. Background Technology

[0002] With the increasing demand for intelligent and comfortable vehicles, sunroof sunshades, as important components for regulating interior light and temperature, are evolving from traditional manual or single-parameter triggering to intelligent, multi-parameter fusion control. However, current technologies often rely on single parameters (such as light intensity, user sleep status, or preset time) for sunshade control, which has the following limitations: Chinese Patent No. CN118596800A discloses a multimedia module that links with weather data, but fails to integrate user preference data, resulting in insufficient personalized experience; Chinese Patent No. CN112277594B discloses a reliance on image recognition of passenger sleep status, which can achieve a certain degree of automation, but may be subject to user resistance and legal restrictions due to privacy issues; Chinese Patent No. CN113427981A discloses a method of adjusting the sunshade using an optical path envelope model, which is innovative in optical control, but fails to combine dynamic environmental parameters (such as air quality and humidity) and real-time user behavior habits, making it difficult to adapt to complex and changing driving environments. Summary of the Invention

[0003] One of the purposes of this application is to address at least one deficiency in the prior art and to propose an improved intelligent control method and system for sunroof shading curtains.

[0004] Therefore, some embodiments of this application provide a dynamic environmentally adaptive intelligent control method for sunroof sunshades based on multi-sensor fusion, which includes the following steps: S1 collects in-vehicle environmental parameters, user behavior data, vehicle status data, and future weather forecast data; S2, based on the real-time collected data, generates a first set of output parameters as the first-level real-time environment-driven response, and outputs the corresponding bottom-line output parameters for safety bottom line, comfort bottom line and energy saving bottom line; using a bidirectional LSTM network combined with an attention mechanism, taking historical user behavior data, real-time environment data and real-time vehicle status data as input, generates a second set of output parameters corresponding to the predicted user preferences as the second-level habit learning prediction response; S3, determine whether the second output parameter set is lower than at least one of the baseline output parameters; if not, the third output parameter set is the second output parameter set; if so, the second output parameter set is dynamically adjusted using the first output parameter set through fuzzy logic rules, thereby generating the third output parameter set; S4, Generate sunshade control instructions based on the third set of output parameters; S5, the regional sunshade curtain actuator controls the opening and closing parameters of the sunshade curtain according to the sunshade curtain control command.

[0005] In some embodiments, the method further includes: initiating a second-level response when it is detected that the time interval between two consecutive manual adjustments by the user is less than a predetermined value and the user makes three consecutive manual adjustments, specifically including: the sunshade control command generation unit converts the user's actual manual intervention parameters into smooth gradient execution commands and issues them, while recording the operation context and passing it to the data preprocessing unit for feature encoding; the preprocessed data is sent to the adaptive learning unit to perform online incremental updates to the core prediction model, while updating the relevant weight parameters of the dynamic priority calculation unit; the updated model regenerates the second output parameter set, and after adjustment by the fuzzy conflict resolution unit, the sunshade control command generation unit outputs optimized control commands, realizing a closed loop from manual learning to automatic decision-making.

[0006] In some embodiments, the method further includes: the baseline output parameters include safety baseline output parameters, comfort baseline output parameters, and energy-saving baseline output parameters; when one output parameter in the second output parameter set is lower than at least one of the baseline output parameters, the second output parameter set is adjusted according to the priority order of the safety baseline output parameters, comfort baseline output parameters, and energy-saving baseline output parameters to generate a third output parameter set.

[0007] In some embodiments, the method further includes a privacy protection step: the user can disable the behavior sensor and switch to a pure environmental parameter control mode.

[0008] In some embodiments, the second set of output parameters for generating the corresponding predicted user preferences includes preprocessing and feature encoding of the collected data to construct temporal features; the bidirectional LSTM network is a two-layer structure with 128 neurons in each layer, combined with an attention mechanism, and the bidirectional LSTM model combined with the attention mechanism is used to predict personalized user preferences.

[0009] In some embodiments, the zoned sunshade curtain actuator, in accordance with the sunshade curtain control command, controls the opening and closing parameters of the sunshade curtain, including driving the zoned sunshade curtain actuator to achieve gradient opening and closing according to the control command.

[0010] Other embodiments of this application provide a dynamic environmentally adaptive intelligent control system for sunroof sunshades based on multi-sensor fusion, comprising: a multimodal sensing module for real-time acquisition of in-vehicle environmental parameters, user behavior data, vehicle status data, and future weather forecast data; a data processing and control module, including a dynamic priority calculation unit, an adaptive learning unit, and a sunshade control command generation unit, wherein the dynamic priority calculation unit is configured to generate a first set of output parameters as a first-level response based on the real-time acquired data, and simultaneously output baseline output parameters corresponding to safety baseline, comfort baseline, and energy-saving baseline; the adaptive learning unit employs a bidirectional LSTM network combined with an attention mechanism based on historical data... Historical user behavior data, real-time environmental data, and real-time vehicle status data are used as inputs to generate a second set of output parameters corresponding to predicted user preferences. The data processing and control module further includes a fuzzy conflict resolution unit. When one output parameter in the second set is lower than at least one of the baseline output parameters, the fuzzy conflict resolution unit dynamically adjusts the second set of output parameters using fuzzy logic rules to generate a third set of output parameters. A sunshade control command generation unit generates sunshade control commands based on the third set of output parameters. A zoned sunshade actuator controls the opening and closing parameters of the sunshade according to the sunshade control commands.

[0011] In some embodiments, the adaptive learning unit includes an online learning unit. When the user continuously manually intervenes to trigger the second-level response, the sunshade control command generation unit feeds back the actual control parameters and context to the data preprocessing unit. After encoding, the data is sent to the adaptive learning unit to fine-tune the parameters of the core prediction model of the bidirectional LSTM network combined with the attention mechanism and update the model version. At the same time, the priority weight parameters of the dynamic priority calculation unit are updated. The updated model regenerates the second output parameter set, which is adjusted by the fuzzy conflict resolution unit and then output by the sunshade control command generation unit to form an optimized control command, thus forming a closed-loop optimization mechanism.

[0012] In some embodiments, the multimodal sensing module includes: an environmental sensor group for collecting data on light intensity, outdoor temperature, ultraviolet index, PM2.5 concentration, and humidity; a user behavior sensor that uses a non-contact infrared sensor to detect passenger location, activity status, and historical operation records; a vehicle networking module for acquiring data on future precipitation probability, cloud thickness, sunrise and sunset times, and real-time weather conditions; and a vehicle status sensor for acquiring data on vehicle speed, steering wheel angle, air conditioning set temperature, navigation route, and vehicle orientation. The above data serves as input parameters for the adaptive learning unit.

[0013] In some embodiments, the output parameters include on / off status, percentage of opening in different zones, sunroof and side window zones, adjustment speed, and notification push.

[0014] In some embodiments, the decision execution section further includes a regional gradient control function, used to convert the sunshade control command into a control command with a smooth gradient execution action having physical and dynamic partitions.

[0015] In some embodiments, the zoned sunshade actuator divides the sunroof into three independent zones: front, middle, and rear, and the side windows into four zones: left front, right front, left rear, and right rear; and dynamically identifies the sun-facing side and the shaded side based on the steering wheel angle and the illumination angle, thereby achieving gradient percentage adjustment and smooth speed control.

[0016] The present invention relates to a dynamic environmentally adaptive intelligent control method and system for sunroof sunshades based on multi-sensor fusion. Through the organic combination of a multi-parameter fusion decision mechanism and a two-level phased response mechanism, it achieves precise, personalized, smooth, and continuously adaptive evolution of in-vehicle sunshade control.

[0017] The specific beneficial effects are as follows: This invention integrates environmental parameters, user behavior data, vehicle status, and weather forecast information through multimodal sensors. It combines dynamic priority calculation, bidirectional LSTM network and attention mechanism for personalized prediction, fuzzy logic conflict resolution, and online learning closed-loop mechanism to achieve coordinated optimization of real-time environment-driven response and user habit learning while ensuring safety and comfort. Through gradient control and smooth speed curve by combining regional physical partitioning and dynamic partitioning, it achieves impact-free and noise-free fine execution.

[0018] The intelligent control method and system for dynamic environmentally adaptive sunroof sunshades proposed in this invention optimizes the control logic through multi-sensor fusion and machine learning algorithms, significantly improving the comfort of the in-vehicle environment and the system's adaptive capabilities.

[0019] In addition, the present invention also provides a privacy protection mode, which allows users to disable the behavior sensor and switch to a pure environmental parameter control mode, effectively solving the privacy controversy of image acquisition in the prior art. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the intelligent control method for a dynamic environmentally adaptive sunshade based on multi-sensor fusion according to the present invention. Figure 2 This is a flowchart illustrating the user preference conflict handling and fuzzy logic decision-making process of the present invention. Figure 3This is a system architecture diagram of the intelligent control system for a dynamic environmentally adaptive sunshade curtain based on multi-sensor fusion according to the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to explain the present invention and are not intended to limit the present invention.

[0022] like Figure 3 As shown, some embodiments of this application provide a smart control system 100 for a sunroof shading curtain, which includes a multimodal sensing module 10, a data processing and control module 20, and an actuator 30.

[0023] The multimodal sensing module 10 includes an environmental sensor group 11, a user behavior sensor 12, a vehicle networking module 13, and a vehicle status sensor. Specifically, the environmental sensor group 11 is used to collect real-time data on in-vehicle temperature, humidity, light intensity, PM2.5 concentration, and ultraviolet intensity; the user behavior sensor 12 uses a non-contact infrared sensor to detect passenger position, activity status, and historical operation records; the vehicle networking module 13 is used to acquire future precipitation probability, cloud thickness, sunrise and sunset times, and real-time weather data; the vehicle status sensor is used to acquire vehicle speed, steering wheel angle, air conditioning set temperature, navigation route, and vehicle orientation data. The above data serves as input parameters for the adaptive learning unit.

[0024] The data processing and control module 20 includes a data preprocessing unit 23, a dynamic priority calculation unit 21, an adaptive learning unit 22, a fuzzy conflict resolution unit 24, and a control command generation unit 25. All input parameters are first preprocessed and feature-encoded (including user operation record encoding, time-series feature construction, and environmental and vehicle data normalization processing) by the data preprocessing unit 23, and then enter the dynamic priority calculation unit 21 and the adaptive learning unit 22.

[0025] The dynamic priority calculation unit 21 generates the opening and closing priority of the sunshade curtain based on environmental parameters, user historical operation data and weather forecast, such as "sun protection priority" or "ventilation priority".

[0026] The adaptive learning unit 22 uses a bidirectional LSTM network combined with an attention mechanism to predict user preferences based on historical user behavior data. Specifically, the adaptive learning unit 22 includes a core prediction model and a decision execution part.

[0027] The core prediction model takes historical user behavior data, real-time environmental data, and real-time vehicle status data as input and directly outputs predicted user preferences. This model employs a bidirectional LSTM network (two layers, 128 neurons per layer) combined with an attention mechanism, achieving 92% accuracy on the test set.

[0028] The core prediction model of the adaptive learning unit employs a bidirectional long short-term memory network (Bi-directional LSTM) combined with an attention mechanism to predict user preferences. The input features of this core prediction model include three types of multimodal data: user behavior data, environmental data, and vehicle status data. Specifically, the input features are 48-dimensional, with 18 dimensions for user behavior data, including historical sunshade operation records such as opening / closing time, position percentage, operation frequency encoding, driving mode preferences such as sport / comfort / eco-friendly modes, seat position adjustment records, voice command semantic encoding, and manual intervention records; 15 dimensions for environmental data, including real-time light intensity (normalized), outdoor temperature, UV index, PM2.5 concentration, humidity, and weather forecast data such as the probability of precipitation in the next 24 hours, cloud thickness, sunrise and sunset times, and real-time weather status one-hot encoding; and 15 dimensions for vehicle status data, including current vehicle speed, steering wheel angle, air conditioning set temperature, navigation route information, and vehicle orientation.

[0029] During the feature engineering phase, user operation records are temporally encoded, transforming continuous manual operations into fixed-length time-series sequences. Environmental and vehicle state data are normalized, for example, using Min-Max normalization or Z-score standardization, and a sliding window-based temporal feature is constructed to capture short-term and long-term dependencies. All input features undergo unified encoding and normalization by the data preprocessing unit 23 before entering the model.

[0030] The core prediction model employs a two-layer bidirectional LSTM network structure, with each layer containing 128 neurons. The bidirectional LSTM can simultaneously capture forward and backward dependencies in time series, effectively modeling the temporal evolution of user operation habits. An attention mechanism is introduced at the output of the bidirectional LSTM. Attention weights are generated by calculating the similarity between the hidden state at each time step and the current query vector, and then a weighted sum is applied to the hidden states. This allows the model to adaptively focus on historical operation segments that are more important to the current prediction task, improving its ability to capture personalized user preferences.

[0031] The model training employed a cross-entropy loss function with L2 regularization to prevent overfitting. The Adam optimizer was selected, with an initial learning rate of 0.001 and a learning rate decay strategy (decreasing to 0.9 times the original rate every 10 epochs). Training hyperparameters were set as follows: batch size of 32, epochs of 50, and early stopping to avoid overfitting. During training, data augmentation was performed using a sliding time window to improve the model's generalization ability to user behavior across different time scales. Validation on the test set showed that the model achieved a 92% accuracy rate in predicting user preferences.

[0032] To adapt to long-term changes in user habits, seasonal differences, and regional variations, the core prediction model also integrates an online learning unit. During vehicle operation, the system collects new user operation data, environmental change data, and vehicle status data every 5 minutes. This new data is then merged with historical data to incrementally update the model. Online learning employs mini-batch gradient descent for parameter fine-tuning and includes a model version management mechanism to ensure a balance between predictive performance and real-time accuracy in the updated model, while avoiding catastrophic forgetting issues.

[0033] The adaptive learning unit 22 of this invention employs a combination of offline training and online learning to achieve continuous model optimization. In the initial stage, the system uses historical multimodal data (user behavior data, environmental data, and vehicle status data, totaling 48 dimensions) to train the bidirectional LSTM+attention mechanism model offline. Training uses a cross-entropy loss function combined with L2 regularization, with Adam as the optimizer, and hyperparameters set to batch size=32 and epoch=50. Overfitting is prevented through an early stopping mechanism and a learning rate decay strategy, achieving a test set accuracy of 92%. To adapt to long-term changes in user habits, seasonal differences, and regional variations, the system further integrates an online learning unit. During vehicle operation, new user operation data, environmental change data, and vehicle status data are collected every 5 minutes. After merging the new data with historical data, incremental parameter fine-tuning of the model is performed using mini-batch gradient descent, and a model version management mechanism is used to avoid catastrophic forgetting. The updated model is immediately used for subsequent user preference prediction, forming a continuously evolving closed-loop learning mechanism.

[0034] The data processing and control module further includes a fuzzy conflict resolution unit 24. This unit 24, when a conflict arises between user preferences and real-time collected environmental data, applies fuzzy logic decision-making to dynamically adjust the model output weights when a conflict occurs between the user preferences predicted by the model and real-time environmental parameters. In this invention, fuzzy logic decision-making refers to the process where, when a conflict arises between the second set of output parameters (i.e., the user preferences predicted by the model) output by the adaptive learning unit and the baseline output parameters (safety baseline, comfort baseline, energy-saving baseline) generated by the dynamic priority calculation unit, the system dynamically adjusts the model output weights using fuzzy logic methods to generate a third set of output parameters that balances real-time environmental constraints and user-specific preferences. This fuzzy logic decision-making prioritizes safety, followed by comfort, and then energy saving, specifically as follows... Figure 2 As shown, this involves employing a Mamdani-type fuzzy inference method. This method fuzzifies three input variables—environmental deviation, model prediction confidence, and safety margin—and performs inference based on a pre-defined fuzzy rule base. Finally, it uses the centroid method for defuzzification, outputting environmental weight adjustment factors and habit weight adjustment factors to correct the original model's prediction weights. This fuzzy logic decision-making mechanism proactively reduces reliance on users' historical habits under extreme environmental conditions, prioritizing in-vehicle safety and comfort. Simultaneously, under more moderate environmental conditions, it appropriately respects users' personalized preferences, achieving a dynamic balance between model prediction and real-time environmental constraints.

[0035] The fuzzy conflict resolution unit 24 employs fuzzy logic to resolve conflicts. Its input variables include three fuzzy variables: environmental deviation, model prediction confidence, and safety margin. Environmental deviation is defined as the degree of deviation of current real-time environmental parameters (light intensity, temperature, UV index, etc.) from preset comfort / safety thresholds, and is divided into three fuzzy levels: "low," "medium," and "high." Model prediction confidence is defined as the credibility of the bidirectional LSTM model output, also divided into three fuzzy levels: "low," "medium," and "high." Safety margin is defined as the degree of proximity of current environmental parameters to a safety baseline, also divided into three fuzzy levels: "low," "medium," and "high." The output variables are the environmental weight adjustment factor and the habit weight adjustment factor, both using three fuzzy levels: "reduce," "maintain," and "increase."

[0036] The membership functions for all input variables are triangular membership functions, while the output variables use trapezoidal membership functions. The parameters of the triangular membership functions are calibrated based on historical data statistics and expert experience to ensure good discriminative power under different environmental conditions. The fuzzy rule base is designed based on the principle of prioritizing safety over comfort, and contains 27 fuzzy rules. The rule design follows this basic logic: when the environmental deviation is "high" and the safety margin is "low," regardless of the model prediction confidence, the environmental weight is prioritized to increase, and the habit weight is decreased; when the environmental deviation is "low" and the model prediction confidence is "high," the habit weight is maintained or appropriately increased. Typical rule examples include: IF Environmental Deviation = High AND Safety Margin = Low THEN Environmental Weight Adjustment Factor = Increase AND Habitual Weight Adjustment Factor = Decrease IF Environmental Deviation = Medium AND Model Prediction Confidence = High THEN Environmental Weight Adjustment Factor = Maintain AND Habitual Weight Adjustment Factor = Maintain IF Environmental Deviation = Low AND Model Prediction Confidence = High AND Safety Margin = High THEN Habitual Weight Adjustment Factor = Increase The fuzzy inference process employs the Mamdani-type fuzzy inference method. First, the input variables are fuzzified to obtain the membership degrees of each input variable to different fuzzy sets. Then, rule activation and inference are performed based on the fuzzy rule base. The minimum operation (Min) is used to synthesize rule premises, and the maximum operation (Max) is used to aggregate rule conclusions. Finally, the centroid method is used for defuzzification to obtain the clear values ​​of the environmental weight adjustment factor and the habit weight adjustment factor.

[0037] The specific algorithm for weight adjustment is as follows: The system corrects the output weights of the original model based on the defuzzified adjustment factor. Let the environmental weights predicted by the original model be... Habitual weighting The adjusted weights are then calculated using the following formula:

[0038] in, and These are the weight adjustment factors obtained after deblurring. and These are adjustment coefficients (0.4 and 0.3 in this embodiment, respectively). The adjusted weights are used to recalculate the third set of output parameters, thereby achieving a dynamic balance between the model prediction results and real-time environmental constraints.

[0039] Through the aforementioned fuzzy conflict resolution mechanism, the system can proactively reduce its reliance on users' historical habits under extreme environmental conditions (such as strong sunlight at noon in summer or sudden rainstorms), prioritizing the safety and comfort of the in-vehicle environment. At the same time, when environmental conditions are relatively mild, it appropriately respects users' personalized preferences.

[0040] The result after the fuzzy conflict resolution unit 24 makes a decision is sent to the sunshade control command generation unit 25. The sunshade control command generation unit 25, as the final link in the system decision-making chain, is responsible for converting the third output parameter set into specific sunshade execution control commands, and forming a close connection with the second-level habitual learning response. Specifically, this unit receives the third output parameter set (including adjusted opening / closing priority, regional percentage, etc.) from the fuzzy conflict resolution unit, or the actual control parameters when the user manually intervenes to trigger the second-level response, and maps them into executable drive commands for the actuator 30, including: binary signals of the open / closed state, regional opening percentage (step size 5%), sunroof front, middle, and rear zones and side window four-zone identifiers, speed adjustment parameters (dynamically adjusted from 0.1 to 10 cm / s), and notification push content; simultaneously, it calculates dynamic zoning (prioritizing protection on the sun-facing side) based on vehicle status data, generates a smooth gradient control curve (linear acceleration / deceleration over 2 seconds), ensuring a shock-free and noise-free execution process; and sends the commands to the motor controllers of each zone via the vehicle network. Regarding the association with the second-level response, when continuous manual adjustments by the user trigger habit learning updates are detected, the sunshade control command generation unit 25 immediately converts the control parameters after the user's actual intervention into execution commands and issues them, achieving smooth physical execution of manual operations. Simultaneously, it fully records the operation context (time, location, area, environmental state) and feeds it back to the adaptive learning unit 22 as incremental training samples to update the core prediction model. At the same time, it adjusts the relevant weights of the dynamic priority calculation unit, so that the updated model generates a second set of output parameters in subsequent cycles that better match the user's real-time preferences. Finally, after fuzzy conflict resolution and processing by the command generation unit, it outputs a better automatic control command, achieving seamless integration between the learning loop and the execution module. Some embodiments of the present invention also design a two-level phased response mechanism to achieve an organic combination of real-time environmental response and user habit learning, specifically including a first-level real-time environment-driven response and a second-level habit learning response.

[0041] The first-level response is a real-time environment-driven response, triggered when the environmental parameters collected in real-time by the environmental sensor array exceed preset safety or comfort thresholds. Specifically, the system immediately initiates the first-level response when the light intensity exceeds 8000 lx, the outdoor temperature exceeds 28°C, the UV index exceeds level 5, or the probability of rainfall in the next hour exceeds 60%. The judgment logic of the first-level response prioritizes safety and comfort thresholds: the system first obtains the threshold output parameters (including safety, comfort, and energy-saving threshold output parameters) corresponding to the current environmental parameters from the dynamic priority calculation unit. Then, it determines whether the second set of output parameters (i.e., the user preference predicted by the model) output by the adaptive learning unit is lower than any threshold output parameter. If it is lower, the model prediction result is dynamically adjusted by the fuzzy conflict resolution unit to generate a third set of output parameters; if it is not lower, the model prediction result is directly used as the third set of output parameters. The specific execution actions of the first-level response include: generating a sunshade control command based on the adjusted third set of output parameters, driving the regional sunshade actuator to open and close the corresponding area in a gradient manner (step size of 5%), and simultaneously pushing the response result to the user interface.

[0042] The second-level response is the habit-learning response, triggered by continuous manual intervention by the user. Specifically, when the system detects that the time interval Δt between two consecutive manual adjustments is less than 1 second and the user makes three consecutive manual adjustments, the second-level response is initiated. The logic for the second-level response is as follows: the system uses the user's continuous manual intervention records (including operation time, percentage of open / closed position, and operation area) as new training samples, inputting them into the core prediction model of the adaptive learning unit for online learning and updating, in order to strengthen or correct the model's ability to predict the user's personalized preferences. The specific execution actions of the second-level response are completed collaboratively by the core modules: the user's manual intervention is first received by the sunshade control command generation unit 25 and converted into a gradient execution command with a smooth speed curve, which is then sent to the regional sunshade execution mechanism 30 to ensure that the physical execution process is shock-free and compatible with the current environmental constraints; at the same time, the unit 25 transmits the complete context of the manual operation (operation parameters, environmental parameters at the time of triggering, vehicle status data, user identity, etc.) to the data preprocessing unit 23 for temporal feature encoding and normalization; the preprocessed data is sent as incremental samples to the adaptive learning unit 22 to perform online parameter fine-tuning of the core prediction model of the bidirectional LSTM network combined with the attention mechanism (using mini-batch gradient descent, combined with model version management to avoid catastrophic forgetting), and the updated model is immediately used for subsequent predictions; in addition, the system synchronously updates the priority weight parameters related to the user in the dynamic priority calculation unit 21. After the update is completed, the system can use the new model to regenerate the second output parameter set for the current state. After necessary adjustments by the fuzzy conflict resolution unit 24, the sunshade control command generation unit 25 outputs the optimized control command, thereby realizing closed-loop optimization from manual learning to automatic decision-making.

[0043] The first-level response and the second-level response interact as follows: The first-level response primarily handles immediate conflicts between real-time environmental parameters and model predictions, offering a fast response time and prioritizing safety and user comfort; the second-level response is initiated after continuous manual intervention by the user, using an online learning mechanism to feed the user's new habits back to the core prediction model, enabling continuous model optimization. When working together, the system first executes the first-level response to meet current environmental constraints. If the user subsequently performs continuous manual intervention, the second-level response is triggered to update the model, thereby gradually improving the consistency between model predictions and the user's actual preferences while ensuring immediate response.

[0044] The actuator 30 is a zone-controlled sunshade that supports gradient opening and closing. The decision execution part also includes a zone-based gradient control function, which is used to convert the sunshade control command into a control command with smooth gradient execution action with physical and dynamic partitions. The physical partitions divide the sunroof into three independent zones: front, middle, and rear, and the side windows into four zones: left front, right front, left rear, and right rear. The dynamic partitions automatically identify the sun-facing and shaded sides based on the steering wheel angle and the light angle, realizing gradient percentage adjustment (5% step) and smooth speed control (opening / closing time of 2 seconds).

[0045] like Figure 1 As shown, the multi-parameter fusion decision-making process of this invention begins with the multi-modal data acquisition of the multi-modal sensing module 10. The environmental sensor group 11 collects real-time data on light intensity (unit: lx, through multiple sensors on the roof and side windows), outdoor temperature (resolution 0.1℃), ultraviolet index, PM2.5 concentration, etc. The user behavior sensor 12 obtains passenger location, activity status, and historical sunshade operation records (opening / closing time, location percentage, operation frequency) through a non-contact infrared sensor. The vehicle networking module 13 obtains the probability of precipitation in the next 24 hours, cloud thickness, sunrise and sunset times, and real-time weather conditions (sunny, rainy, snowy, foggy, etc.). The vehicle networking module 13 also obtains vehicle status data, including current vehicle speed (km / h), steering wheel angle (degrees, used to determine changes in light angle), air conditioning set temperature (℃), navigation route information, and vehicle orientation (determined by GPS and gyroscope).

[0046] All input parameters, after being processed by the data preprocessing unit 23 and the feature encoding unit of the data processing and control module 20, enter the dynamic priority calculation unit 21 and the adaptive learning unit 22. Data preprocessing may include: encoding user operation records, for example, "on" = 100, "off" = 010, "gradient adjustment" = 001; constructing temporal features by combining these with time data; and normalizing environmental and vehicle data.

[0047] Taking a specific sampling moment as an example: the user behavior code is 100 (on), location is 0.5, driving mode is 1 (comfort), and voice command is 3 (too sunny); the environmental data includes a normalized light intensity of 0.75 (corresponding to 8000 lx), a temperature of 1.2 (corresponding to 32℃), UV index of 0.6, and weather condition of {1,0,0,0} (sunny); the vehicle data includes a speed of 0.4 (80 km / h) and a steering wheel angle of 0.1 (2°). These parameters are collectively input into the model.

[0048] This invention proposes an adaptive control system for vehicle sunshades based on a hybrid decision-making approach combining deep learning and fuzzy rules. The core prediction model (hereinafter referred to as the "core model") uses a bidirectional LSTM network combined with an attention mechanism as the adaptive learning unit. This model takes user behavior data (historical operations, driving mode preferences, voice commands, manual intervention, etc.), environmental data (light intensity, temperature, UV index, weather forecast, etc.), and vehicle status data (vehicle speed, steering wheel angle, navigation information, etc.) as multimodal inputs, and directly outputs sunshade control commands (open / closed status, percentage of opening in different areas, sunroof / side window partitioning, speed adjustment, notification push, etc.).

[0049] Specifically, the complete list of model input parameters includes: User behavior data includes: historical sunshade operation records (opening / closing time, position percentage, operation frequency), driving mode preferences (sport / comfort / eco-friendly mode selection records), seat position adjustment records, user identification records (e.g., via fingerprint or facial recognition), voice command records ("too sunny", "too dark"), and manual intervention records (whether the automatic closing was revoked within 10 seconds).

[0050] Environmental data, such as real-time light intensity, outdoor temperature, UV index, and weather forecast data (probability of precipitation in the next 24 hours, cloud thickness, sunrise and sunset times, and real-time weather conditions).

[0051] Vehicle status data, such as current speed, steering wheel angle, air conditioning set temperature, navigation route information, and vehicle orientation.

[0052] The output parameters are sunshade control commands, including open / close status (binary), opening percentage of each zone (0%~100%, step size 5%), sunroof zones (front, middle, rear), side window zones (left front, right front, left rear, right rear), adjustment speed (0.1~10 cm / s, dynamic smooth adjustment), and whether to push notifications (such as "adjusted according to habit").

[0053] In terms of training details, we adopted the cross-entropy loss function with L2 regularization, the optimizer was Adam, the hyperparameters were set as batch size = 32, epochs = 50, initial learning rate = 0.001, and data augmentation was performed by sliding time window.

[0054] Building upon the aforementioned model predictions, this invention further constructs a multi-level decision-making and execution architecture. Through an online learning mechanism, the core model continuously adapts to long-term changes in user habits and seasonal and regional differences. In this online learning mechanism, the model updates in real time during vehicle operation, collecting new data (including new user actions and environmental changes) every 5 minutes to adapt to long-term changes such as seasons and regions. For example, in winter, users tend to retain some light transmission; the model automatically adjusts its control strategy to reduce reliance on glare reduction.

[0055] A fuzzy logic rule base is used to dynamically adjust the weights and resolve conflicts in the model output. Specifically, when predicted user preferences conflict with real-time environmental parameters, the system automatically increases the weight of environmental parameters and decreases the weight of historical habits to prioritize safety. For example, when predicted user preferences conflict with real-time environmental parameters, the system makes a dynamic decision using the fuzzy logic rule base. Taking a summer midday period with light intensity >8000 lx and temperature >32℃ as an example: before adjustment, the model tends to follow the user's past operating habits (light intensity weight 0.6, temperature weight 0.5, historical operation weight 0.9); after adjustment, the system detects that the environmental parameters exceed the safety / comfort threshold, applies an environmental bias, increases the light intensity weight to 0.8 (an increase of 33%), increases the temperature weight to 0.7 (an increase of 40%), and simultaneously decreases the user's historical operation weight to 0.6 (a decrease of 33%). The physical meaning is that in extreme sunlight conditions, the model actively reduces its reliance on the user's past habit of "liking to open the sunroof," forcibly triggering the sunshade to close, preventing overheating inside the car and aging of the interior.

[0056] In cases of conflicting habits (such as a sleeping habit requiring full closure versus a ventilation habit requiring partial closure), the system prioritizes safety (sun protection / rain protection) first, followed by comfort (glare control), and finally energy efficiency. Driver identity is personalized via fingerprint or voice recognition, with rear passenger habits serving only as a secondary consideration.

[0057] In some embodiments of this invention, high-level decisions are ultimately translated into precise, shock-free physical execution through regional gradient control (physical zoning + dynamic zoning + smooth speed curve). Regional gradient opening is achieved through physical and dynamic zoning. Physical zoning divides the sunroof into three independent zones: front, middle, and rear. Side windows are divided into four zones based on seating position: left front, right front, left rear, and right rear. Each zone is equipped with an independent motor and light sensor. Dynamic zoning automatically identifies the sun-facing side based on steering wheel angle and light angle. For example, when the vehicle turns right, the right-side window automatically enters "strong light protection mode," closing the sunshade 80%, while the left side remains at its default setting.

[0058] Gradient control employs percentage adjustment (5% step) and smooth speed control: when activated, the speed linearly increases from 0 to 5 cm / s (over 2 seconds), and when deactivated, it linearly decreases from 5 cm / s to 0 (over 2 seconds), effectively avoiding mechanical shock. Performance verification shows that compared to the traditional fully on / fully off mode, the average interior temperature in summer is reduced by 3-5℃, and air conditioning energy consumption is reduced by 15-20%. User subjective evaluations show that the glare rate for rear passengers decreased from 40% to 12%, and driver distraction was reduced by 30%.

[0059] The architecture of the embodiment, incorporating all the aforementioned features, achieves a combination of personalized prediction through models, safety rule correction through fuzzy logic, rapid response and learning loop through phased control, and smooth physical execution through gradient control. This ensures a 92% prediction accuracy while meeting the real-time, safety, and resource constraints required by automotive embedded systems. Simulation tests show that this solution, through multi-sensor fusion and adaptive learning, achieves precise and personalized sunshade control, improving system response speed by 30% and reducing user manual intervention frequency by 45%, demonstrating both innovation and commercial viability. This invention possesses significant technical advantages in dynamic priority algorithms, regional gradient control, machine learning-based personalized prediction, and fuzzy conflict handling.

[0060] In summary, such as Figure 1 As shown, an embodiment of the present invention provides a method for intelligent control of a dynamic environmentally adaptive sunshade based on multi-sensor fusion, comprising the following steps: Step S1 involves the multimodal sensing module 10 collecting in-vehicle environmental parameters, user behavior data, vehicle status data, and future weather forecast data in real time. Specifically, this includes the environmental sensor group 11 collecting light intensity, temperature, UV index, PM2.5 concentration, and humidity; the user behavior sensor 12 using a non-contact infrared sensor to detect passenger position, activity status, and historical operation records; the vehicle networking module 13 acquiring weather forecast data; and the vehicle status sensor acquiring information such as vehicle speed and steering wheel angle.

[0061] In step S2, the collected data is first preprocessed and feature-encoded by the data preprocessing unit 23 to construct temporal features. The processed data then enters the dynamic priority calculation unit 21 and the adaptive learning unit 22. The dynamic priority calculation unit 21 generates a first set of output parameters based on real-time environmental parameters and outputs baseline output parameters corresponding to safety, comfort, and energy efficiency limits. The core prediction model of the adaptive learning unit 22 uses a bidirectional LSTM network combined with an attention mechanism, taking historical user behavior data, real-time environmental data, and real-time vehicle status data as input to generate a second set of output parameters corresponding to the predicted user preferences, serving as the second-level habit learning prediction response.

[0062] In step S3, the adaptive learning unit 22 determines whether the second output parameter set is lower than at least one of the baseline output parameters. If not, the third output parameter set is the second output parameter set; if so, the fuzzy conflict resolution unit dynamically adjusts the second output parameter set using the first output parameter set according to the order of highest priority for safety baseline output parameters, followed by comfort baseline output parameters, and then energy-saving baseline output parameters, through fuzzy logic rules, thereby generating the third output parameter set.

[0063] In step S4, the sunshade control command generation unit 25 generates sunshade control commands based on the third output parameter set. Specifically, this includes parameter mapping to motor drive signals for each zone, dynamic zone-differentiated command generation, smooth speed curve planning (linear transition lasting 2 seconds), safety verification, and user notification push. When the second-level habit learning response is triggered by user manual intervention, the unit also converts the actual control parameters into execution commands and records them as feedback to the adaptive learning unit for online model updates.

[0064] Step S5: The zoned sunshade curtain actuator 30 drives the zoned controllable sunshade curtain to achieve gradient opening and closing control according to the control command, including smooth execution of physical and dynamic zones.

[0065] During the operation of S1 to S5 above, when it is detected that the user makes three consecutive manual adjustments and the time interval between two adjacent manual adjustments is less than a predetermined value, the adaptive learning unit 22 starts the habit learning update mechanism of the second-level response, and uses the user's manual intervention operation record as a new training sample to input into the core prediction model for online learning and update.

[0066] In step S2, the data preprocessing unit 23 performs preprocessing and feature encoding on the collected data to construct temporal features. Specifically, this includes: temporal encoding of user operation records, converting continuous manual operation behaviors into fixed-length temporal sequences; normalizing environmental and vehicle state data, and constructing temporal features based on a sliding window. The preprocessed data enters the core prediction model of the adaptive learning unit 22. This model employs a two-layer bidirectional LSTM network structure, with each layer containing 128 neurons, and combines an attention mechanism. It takes historical user behavior data, real-time environmental data, and real-time vehicle state data as input to generate a second set of output parameters corresponding to the predicted user preferences.

[0067] The second-level response is the habit learning response, triggered by the following condition: when the time interval between two consecutive manual adjustments by the user is less than a predetermined value, and the user makes three consecutive manual adjustments, the adaptive learning unit initiates the recording process of historical user behavior data. At this time, the system uses the user's operation record for this manual intervention (including operation time, percentage of opening / closing position, and operation area) as a new training sample, inputting it into the core prediction model for online learning and updating, in order to strengthen or correct the model's ability to predict the user's personalized preferences, and simultaneously update the user's historical operation record database.

[0068] In step S3, when one of the output parameters in the second set of output parameters from the adaptive learning unit 22 is lower than at least one of the baseline output parameters, the fuzzy conflict resolution unit dynamically adjusts the second set of output parameters in the order of highest priority for safety baseline output parameters, followed by comfort baseline output parameters, and then energy-saving baseline output parameters, thereby generating a third set of output parameters. Specifically, the system prioritizes the safety baseline; when environmental parameters deviate significantly from the safety threshold, the weight of environmental factors is significantly increased, while the weight of historical habits is decreased. Once the safety baseline is met, the comfort baseline and energy-saving baseline are considered, thus achieving priority coordination among multiple objectives.

[0069] In step S5, the zoned sunshade actuator drives the zoned controllable sunshade to achieve gradient opening and closing according to the sunshade control command. Specifically, this includes: dividing the sunroof into three independent zones (front, middle, and rear) and the side windows into four zones (left front, right front, left rear, and right rear); dynamically identifying the sun-facing and shaded sides based on the steering wheel angle and the lighting angle; adjusting the gradient percentage in 5% increments and using smooth speed control (opening / closing takes 2 seconds) to achieve precise, shock-free execution.

[0070] The method of the present invention also includes a privacy protection step: the user can disable the user behavior sensor 12 through the human-computer interaction interface, and the system switches to a pure environmental parameter control mode. In this mode, the adaptive learning unit 22 makes decisions based solely on the data collected by the environmental sensor group 11 and the vehicle networking module 13, and no longer uses user behavior data, thereby effectively protecting user privacy.

Claims

1. A method for intelligent control of a dynamic environmentally adaptive sunshade curtain based on multi-sensor fusion, characterized in that, Includes the following steps: S1 collects in-vehicle environmental parameters, user behavior data, vehicle status data, and future weather forecast data; S2, based on the real-time collected data, generates a first set of output parameters as the first-level real-time environment-driven response, and outputs the corresponding bottom-line output parameters for safety bottom line, comfort bottom line and energy saving bottom line; using a bidirectional LSTM network combined with an attention mechanism, taking historical user behavior data, real-time environment data and real-time vehicle status data as input, generates a second set of output parameters corresponding to the predicted user preferences as the second-level habit learning prediction response; S3, determine whether the second set of output parameters is lower than at least one of the baseline output parameters; if not, then the third set of output parameters is the second set of output parameters; If so, then the second output parameter set is dynamically adjusted using the first output parameter set through fuzzy logic rules, thereby generating the third output parameter set; S4, Generate sunshade control instructions based on the third set of output parameters; S5, the regional sunshade curtain actuator controls the opening and closing parameters of the sunshade curtain according to the sunshade curtain control command.

2. The method according to claim 1, characterized in that, When it is detected that the time interval between two consecutive manual adjustments by the user is less than a predetermined value and three consecutive manual adjustments are made, the second-level response is initiated. Specifically, the sunshade control command generation unit converts the user's actual manual intervention parameters into smooth gradient execution commands and issues them, while recording the operation context and passing it to the data preprocessing unit for feature encoding; the preprocessed data is sent to the adaptive learning unit to perform online incremental updates to the core prediction model, while updating the relevant weight parameters of the dynamic priority calculation unit; the updated model regenerates the second output parameter set, which is then adjusted by the fuzzy conflict resolution unit, and the sunshade control command generation unit outputs the optimized control commands, realizing a closed loop from manual learning to automatic decision-making.

3. The method according to claim 1, characterized in that, The baseline output parameters include safety baseline output parameters, comfort baseline output parameters, and energy-saving baseline output parameters. When one of the output parameters in the second set of output parameters is lower than at least one of the baseline output parameters, the second set of output parameters is adjusted according to the priority order of safety baseline output parameters, comfort baseline output parameters, and energy-saving baseline output parameters to generate a third set of output parameters.

4. The method according to claim 1, characterized in that, The method also includes a privacy protection step: users can disable the behavior sensors and switch to a pure environmental parameter control mode.

5. The method according to claim 1, characterized in that, The second set of output parameters for generating the corresponding predicted user preferences includes preprocessing and feature encoding of the collected data to construct temporal features; the bidirectional LSTM network is a two-layer structure with 128 neurons in each layer, combined with an attention mechanism, and predicts personalized user preferences through the bidirectional LSTM model combined with the attention mechanism.

6. The method according to claim 1, characterized in that, The zoned sunshade curtain actuator controls the opening and closing parameters of the sunshade curtain according to the sunshade curtain control command, including driving the zoned sunshade curtain actuator to achieve gradient opening and closing according to the control command.

7. A dynamic environmentally adaptive intelligent control system for sunroof shading curtains based on multi-sensor fusion, characterized in that, include: The multimodal sensing module is used to collect in-vehicle environmental parameters, user behavior data, vehicle status data, and future weather forecast data in real time. The data processing and control module includes a dynamic priority calculation unit, an adaptive learning unit, and a sunshade control command generation unit. The dynamic priority calculation unit is configured to generate a first set of output parameters as a first-level response based on the real-time collected data, and simultaneously output baseline output parameters corresponding to safety baseline, comfort baseline, and energy-saving baseline. The adaptive learning unit uses a bidirectional LSTM network combined with an attention mechanism to generate a second set of output parameters corresponding to predicted user preferences based on historical user behavior data, real-time environmental data, and real-time vehicle status data as input. The data processing and control module further includes a fuzzy conflict resolution unit. When one of the output parameters in the second output parameter set is lower than at least one of the bottom line output parameters, the fuzzy conflict resolution unit dynamically adjusts the second output parameter set using the first output parameter set through fuzzy logic rules to generate a third output parameter set. The sunshade control command generation unit generates sunshade control commands based on the third set of output parameters. The zoned sunshade curtain actuator controls the opening and closing parameters of the sunshade curtain according to the sunshade curtain control command.

8. The system according to claim 7, characterized in that, The adaptive learning unit includes an online learning unit. When the user continuously manually intervenes to trigger the second-level response, the sunshade control command generation unit feeds back the actual control parameters and context to the data preprocessing unit. After encoding, the data is sent to the adaptive learning unit to fine-tune the parameters of the core prediction model of the bidirectional LSTM network combined with the attention mechanism and update the model version. At the same time, the priority weight parameters of the dynamic priority calculation unit are updated. The updated model regenerates the second set of output parameters, which are then adjusted by the fuzzy conflict resolution unit and optimized by the sunshade control command generation unit, forming a closed-loop optimization mechanism.

9. The system according to claim 8, characterized in that, The multimodal sensing module includes: An environmental sensor array is used to collect data on light intensity, outdoor temperature, UV index, PM2.5 concentration, and humidity. The user behavior sensor uses a non-contact infrared sensor to detect passenger location, activity status, and historical operation records. The vehicle-to-everything (V2X) module is used to acquire data on future precipitation probability, cloud thickness, sunrise and sunset times, and real-time weather conditions. Vehicle status sensors are used to acquire data such as vehicle speed, steering wheel angle, air conditioning set temperature, navigation route, and vehicle orientation. The above data serves as the input parameters for the adaptive learning unit.

10. The system according to claim 7, characterized in that, The output parameters include on / off status, percentage of opening in different areas, sunroof and side window partitioning, adjustment speed, and notification push.

11. The system according to claim 7, characterized in that, The decision execution section also includes a regional gradient control function, which is used to convert the sunshade control command into a control command with a smooth gradient execution action with physical and dynamic partitions.

12. The system according to claim 7, characterized in that, The zoned sunshade actuator divides the sunroof into three independent zones: front, middle, and rear, and the side windows into four zones: left front, right front, left rear, and right rear. It also dynamically identifies the sun-facing and shaded sides based on the steering wheel angle and the lighting angle, enabling gradient percentage adjustment and smooth speed control.

Images