Automatic opening and closing method of louvered ventilation window based on multi-parameter intelligent sensing system

Abstract

The application discloses a louvered ventilation window automatic opening and closing method based on a multi-parameter intelligent sensing system, through a multi-level intelligent sensing network, indoor and outdoor environment and window body state data are synchronously collected, after data fusion and abnormality elimination processing, a standardized data set is generated; the system utilizes a deep reinforcement learning model to dynamically adjust the decision weight of each environment parameter, and combines an LSTM network to predict the future micro environment; taking indoor comfort, energy efficiency and air health degree as optimization targets, a multi-objective optimization model is constructed, the target opening and angle of the louvered window are calculated through an optimization algorithm, and high-precision execution is realized by a mute driving mechanism; meanwhile, the system integrates anti-pinch protection and state monitoring functions, and continuously updates a user preference file by learning user manual adjustment behavior, so that closed-loop optimization is realized; the application realizes intelligent and adaptive control of the louvered window, and effectively improves environmental comfort, energy saving effect and air health degree.

Description

Technical Field

[0001] This invention belongs to the field of smart home and environmental control technology, and specifically discloses a method for automatically opening and closing louvered ventilation windows based on a multi-parameter intelligent sensing system. Background Technology

[0002] As a common building envelope component, louvered windows not only provide basic functions such as sun shading, ventilation, and privacy protection, but are also gradually becoming a key interface for indoor environmental control and energy management. With the development of smart building and home automation technologies, the automatic control of louvered windows has become an important research direction for improving living comfort and achieving building energy conservation.

[0003] Traditional automatic control methods for venetian blinds are mostly based on setting fixed thresholds for opening / closing or adjusting angles based on single or a few environmental parameters (such as light intensity and indoor temperature). While these methods achieve preliminary automation, they have significant limitations: First, the control strategies are simple and rigid, unable to cope with complex and changing environmental conditions (such as temperature and humidity coupling, and conflicts between light and ventilation requirements); second, they lack the ability to learn and adapt to users' personalized habits; third, they fail to comprehensively consider the synergistic optimization of multiple objectives such as indoor air quality (such as CO2 and PM2.5) and energy efficiency; fourth, the systems typically rely on wired power supply and communication, resulting in poor deployment flexibility and a lack of effective abnormal data identification and fault tolerance mechanisms.

[0004] In recent years, some improved solutions have introduced more sensors and rule-based or simple feedback-based control logic, such as adjusting according to schedules or indoor-outdoor temperature differences. However, these methods still have shortcomings in the depth of parameter fusion, the level of intelligence in decision-making, and the system's adaptability. Specifically, environmental parameter weights are mostly statically preset and cannot be dynamically adjusted according to seasons, weather, or usage scenarios; there is a lack of predictive ability for short-term environmental changes, leading to lag in control response; multi-objective optimization problems are often simplified to single-objective problems or processed through linear weighting, making it difficult to obtain a truly Pareto optimal solution; in addition, there is insufficient consideration for safety (such as anti-pinch), robustness (such as abnormal data handling), and energy autonomy.

[0005] Therefore, it is necessary to invent an automatic opening and closing method for louvered ventilation windows based on a multi-parameter intelligent sensing system to solve the above problems. Summary of the Invention

[0006] To overcome the aforementioned shortcomings of existing technologies, this invention provides an automatic opening and closing method for louvered ventilation windows based on a multi-parameter intelligent sensing system. The system synchronously collects indoor and outdoor environmental data and window status data through a multi-level intelligent sensing network, generating a standardized dataset after data fusion and anomaly removal. The system dynamically adjusts the decision weights of various environmental parameters using a deep reinforcement learning model and combines it with an LSTM network to predict the future microenvironment. A multi-objective optimization model is constructed with indoor comfort, energy efficiency, and air quality as optimization goals. The target opening degree and angle of the louvers are calculated using an optimization algorithm, and executed with high precision by a silent drive mechanism. Simultaneously, the system integrates anti-pinch protection and status monitoring functions, and continuously updates user preference profiles by learning user manual adjustment behaviors, achieving closed-loop optimization and effectively solving the problems mentioned in the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an automatic opening and closing method for louvered ventilation windows based on a multi-parameter intelligent sensing system, specifically including the following steps:

[0008] S1. Construct a multi-level intelligent sensor network to simultaneously collect multi-source sensor data, including indoor and outdoor data and window status data.

[0009] S2. The collected multi-source sensor data is fused and processed. Through cross-verification of optical, electronic and airflow signals, abnormal data is removed and a standardized environmental situation dataset is generated.

[0010] S3. A dynamic parameter weight adaptive system is built based on a deep reinforcement learning model. The decision weight of each environmental parameter is calculated according to time pattern, seasonal factor, user preference, and energy consumption coefficient, and the priority of parameter influence is adjusted in real time under different scenarios.

[0011] S4. Integrate weather forecast API data, use an LSTM network to predict microenvironmental changes in the next 2 hours, and build a multi-objective optimization model by combining the real-time dataset from step S2.

[0012] S5. Using indoor comfort, energy efficiency, and air health as optimization objectives and weather conditions as constraints, calculate the target opening and angle of the louvers through a multi-objective optimization algorithm.

[0013] S6. The target control command is executed through a silent drive mechanism to achieve window blade adjustment with an accuracy of 0.1°. At the same time, anti-pinch protection and status monitoring are achieved through a motor torque sensor and a micro-vibration sensor.

[0014] S7. Collect environmental feedback data and user manual adjustment behavior data after control, and update user preference profiles and weight allocation models through personalized habit learning engine and unsupervised learning algorithm to complete closed-loop learning.

[0015] The technical effects and advantages of this invention are as follows:

[0016] 1. Through a multi-level wireless mesh sensor network, multi-dimensional indoor and outdoor environmental data (such as light, temperature and humidity, air quality, and personnel status) are collected synchronously. Combined with the window's own status monitoring, a comprehensive environmental situation awareness is formed. By adopting cross-validation of optical, electronic, and airflow signals and an abnormal data removal mechanism, the data quality and system reliability are significantly improved, providing accurate input for intelligent decision-making.

[0017] 2. The dynamic parameter weighting system based on deep reinforcement learning can automatically adjust the decision priority of different environmental parameters according to time, season, user preferences and energy consumption requirements, so that the system can adapt to diverse scenarios (such as summer sunshade, winter heat preservation, and pollution purification); it integrates LSTM network and weather forecast to predict micro-environmental changes in the next 2 hours, so that the venetian blind adjustment is forward-looking and can adapt to weather changes in advance.

[0018] 3. With comfort, energy efficiency, and air health as optimization goals, a multi-objective optimization algorithm is used to solve the Pareto optimal solution to achieve a scientific balance among multiple goals; through unsupervised learning, users' manual adjustment habits are continuously learned to establish personalized preference profiles and automatically generate scenario-based strategies (such as sleep mode and meeting mode) to improve user experience and satisfaction. Attached Figure Description

[0019] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0020] Figure 1 This is a diagram showing the overall system architecture and data flow of the present invention.

[0021] Figure 2 This is a flowchart of the data fusion and prediction system of the present invention.

[0022] Figure 3 This is the multi-objective optimization and execution control logic diagram of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This invention provides an automatic opening and closing method for louvered ventilation windows based on a multi-parameter intelligent sensing system. The overall system architecture and data flow are as follows: Figure 1 As shown, the present invention specifically includes the following steps:

[0025] S1. Construct a multi-level intelligent sensor network to simultaneously collect multi-source sensor data, including indoor and outdoor data and window status data.

[0026] Furthermore, in the above technical solution, the multi-level intelligent sensor network adopts a wireless mesh networking method. The outdoor array, indoor array, and window patch sensor all integrate a multi-parameter fusion SoC chip. The outdoor array collects light intensity / spectrum, temperature and humidity, rainwater signal, wind speed and direction, PM2.5 concentration, and CO2 concentration; the indoor array collects personnel occupancy status, temperature and humidity, CO2 concentration, and light distribution; and the window patch sensor collects window leaf vibration status and motor torque data.

[0027] It should be further explained that when constructing a multi-level intelligent sensor network, wireless mesh networking technology is used as the core architecture to achieve full-domain coordination and synchronous data acquisition among outdoor arrays, indoor arrays, and window patch sensors. Specific implementation details are as follows:

[0028] First, the deployment of sensor nodes is planned. Outdoor arrays are prioritized for key locations on the building facade: 3-5 sensor nodes are placed on each of the sunny, shaded, and windward sides to form a three-dimensional monitoring network, ensuring comprehensive capture of changes in the outdoor environment. Nodes are installed 2-3 meters above the ground, avoiding obstructions and strong electromagnetic interference sources (such as air conditioner units and communication base stations), with a spacing of 8-10 meters between nodes to ensure stable relay communication within the Mesh network. Indoor arrays are deployed according to functional areas: one core node is configured in each of the main activity spaces such as the living room, bedroom, and study; auxiliary nodes are added in special areas such as the kitchen and bathroom. Nodes are installed 1.5-2 meters above the ground, avoiding environmental disturbance sources such as air conditioner vents and radiators, while being located near areas with frequent human activity to improve the accuracy of parameters such as occupancy status. Window panel sensors use a modular adhesive design. One vibration sensor is attached to each end of the frame of each window slat, and one torque sensor is fixed to the motor drive. The contact surfaces between the sensors and the window slats are made of non-slip, wear-resistant material to ensure no attenuation of vibration signals.

[0029] All sensing nodes integrate a multi-parameter fusion SoC chip, which must have three core functions: a data preprocessing module that performs mean filtering and noise suppression on the raw sensor data to initially remove high-frequency interference signals; a wireless communication module that supports the Mesh self-organizing network protocol, automatically identifies surrounding nodes and establishes communication links, and automatically switches to the backup link when a node fails to ensure communication continuity, with the communication frequency set at 2.4GHz, the data transmission rate not less than 1Mbps, and the end-to-end latency controlled within 100ms; and a low-power control module that adopts an intermittent working mode, automatically entering sleep mode during non-collection periods, with a sleep current of less than 10μA, significantly reducing energy consumption.

[0030] The data acquisition synchronization mechanism is achieved through time calibration and dynamic periodic control: all nodes automatically connect to the network gateway after power-on, and perform clock calibration every hour through the NTP time synchronization protocol to ensure that the timestamp error of the data collected by each node does not exceed 5ms; the acquisition period is dynamically adjusted according to the parameter characteristics. For parameters with slow changes such as temperature and humidity, CO2 concentration, and PM2.5 concentration, the acquisition period is set to 30 seconds / time; for parameters with rapid changes such as light intensity, wind speed and wind direction, the acquisition period is shortened to 10 seconds / time; for sudden parameters such as rainwater signals and motor torque data, trigger-based acquisition is adopted. When the signal change amplitude is detected to exceed the preset threshold (such as a sudden change in rainwater signal intensity ≥50%), acquisition is immediately started and data is uploaded first.

[0031] The sensor selection strictly matches the monitoring requirements: the outdoor array's light intensity / spectral sensor has a measurement range of 0-200,000 lux and an accuracy of ±5%, supporting segmented acquisition of visible and near-infrared spectra; the temperature and humidity sensor covers a measurement range of -40℃ to 85℃ and 0-100%RH, with a temperature accuracy of ±0.3℃ and a humidity accuracy of ±2%RH; the wind speed and direction sensor has a wind speed measurement range of 0-30m / s and an accuracy of ±0.2m / s, and a wind direction measurement range of 0-360° and an accuracy of ±5°; the PM2.5 and CO2 sensors meet the measurement requirements of 0-1000μg / m³ (accuracy ±10μg / m³) and 0-5000ppm (accuracy ±50ppm), respectively. The indoor array of personnel occupancy sensors adopts a dual-mode fusion design of infrared + millimeter-wave radar, with a detection distance of 0-5m. It can accurately identify the presence and movement of personnel and is not affected by light intensity. The light distribution sensor integrates four acquisition probes facing different directions to capture the light intensity in the four directions of indoor space, achieving all-round monitoring of light distribution.

[0032] S2. The collected multi-source sensor data is fused and processed. Through cross-verification of optical, electronic and airflow signals, abnormal data is removed and a standardized environmental situation dataset is generated.

[0033] Furthermore, in the above technical solution, the abnormal data removal method in step S2 is to set the normal fluctuation range and physical correlation constraints of each sensor parameter. When a single point data exceeds the threshold range or has a logical conflict with at least two of the related signals in optical, electronic, and airflow signals, it is determined to be an abnormal value and smooth interpolation replacement is performed.

[0034] It should be further explained that the core of step S2 is to transform the raw data collected by multiple sensors into a high-quality, standardized environmental situation dataset. The specific implementation details are as follows:

[0035] First, data preprocessing is performed, including format standardization and noise suppression of raw data from multiple sensors. All raw data collected by sensors must be standardized according to a preset protocol: timestamps are uniformly formatted in UTC, accurate to milliseconds; numerical units are strictly aligned (e.g., temperature in °C, illumination in lux, concentration in μg / m³ or ppm); and sensor IDs are bound to the type of collected parameters to avoid data confusion. For high-frequency noise in the raw data, a two-stage filtering process is employed: the first stage uses the mean filtering algorithm built into the multi-parameter fusion SoC chip to calculate a moving average of five consecutive acquisitions of single-point sensor data, initially filtering out random interference; the second stage, for parameters with drastic dynamic changes such as wind speed and vibration, additionally uses a Kalman filter algorithm to construct state and observation equations based on historical parameter trends, further reducing the impact of noise on data accuracy. Simultaneously, for potential data packet loss or missing field issues during transmission, the retransmission mechanism of the wireless mesh network is used to supplement missing data. If data is not acquired after three retransmissions, it is marked as temporarily missing and will be supplemented during the subsequent fusion stage.

[0036] Subsequently, multi-signal cross-validation was conducted to establish a correlation verification model for three types of signals: optical, electronic, and airflow, thereby achieving cross-verification of data validity. Optical signals include light intensity / spectral data and rainwater signals (the shading characteristics of rainwater on light); electronic signals include temperature and humidity, CO2 concentration, PM2.5 concentration, motor torque, and personnel occupancy status data; airflow signals include wind speed and direction, and indoor and outdoor air pressure difference (derived from temperature, humidity, and atmospheric pressure sensors). The verification logic is designed based on the physical correlation between environmental parameters. For example, when the optical signal detects a light intensity ≥10000 lux (sunny day characteristic), the outdoor temperature in the electronic signal should meet the constraint of "deviation from the historical average of the same period ≤3℃", and the wind speed and direction in the airflow signal should be consistent with the general trend of the local weather forecast. If there is a logical conflict among the three (e.g., strong light but significantly low temperature and no rainfall signal), the relevant data is deemed to have validity issues. Similarly, when the airflow signal detects a wind speed ≥5m / s, the electronic signal (vibration state) of the window patch sensor should exhibit a correlation characteristic of "vibration amplitude ≥0.1mm". If the vibration signal has no response and there are no other obstructions, the wind speed or vibration sensor data may be abnormal. During cross-validation, each parameter must be correlated with at least two different types of signals. Only when all correlation verifications pass can the subsequent processing stage begin.

[0037] Abnormal data removal employs a dual standard of "threshold judgment + logical conflict," combining parameter characteristics to set precise judgment rules. First, dynamic normal fluctuation ranges are set for each sensor parameter: temperature fluctuation thresholds are adjusted seasonally (indoor-outdoor temperature difference ≤10℃ in summer, ≤15℃ in winter, single parameter fluctuation ≤5℃ within 10 minutes); humidity fluctuation thresholds are set at ≤15%RH within 10 minutes; PM2.5 concentration single-time mutation ≤200μg / m³, CO2 concentration increase ≤500ppm within 1 hour in unoccupied conditions; wind speed instantaneous mutation ≤3m / s, vibration amplitude ≤0.5mm (no external force intervention scenario). When a single data point exceeds the above corresponding thresholds, it is directly marked as a suspected anomaly. Simultaneously, based on cross-validation results, if a parameter data has an irreconcilable logical conflict with at least two of the optical, electronic, and airflow signals (e.g., a rain signal indicating "rain" but outdoor humidity <80%RH, and no significant decrease in light intensity), it is judged as a confirmed anomaly. For suspected abnormal data, a "continuous observation" mechanism is adopted: if the data collected in the next two collections returns to the normal range and the associated signals are verified, the original data is retained and the fluctuation is recorded; if the data collected in the next three collections is suspected abnormal or directly determined to be confirmed abnormal, the abnormal data replacement process is initiated. For parameters that change slowly (such as temperature, humidity, and CO2 concentration), linear interpolation is used to fill in the missing values ​​based on the changing trends of the three valid data points before and after the abnormal data; for parameters that change rapidly (such as light intensity and wind speed), cubic spline interpolation is used to accurately fit the parameter change curve and ensure that the replaced data conforms to the laws of environmental physical change.

[0038] Finally, data fusion and standardized dataset generation are performed. A weighted fusion algorithm is used to integrate the preprocessed and cross-validated multi-source data: fusion weights are assigned based on the sensor's accuracy level and deployment location reliability (e.g., outdoor core area sensors are weighted at 0.8, auxiliary area sensors at 0.2; indoor densely populated areas sensors are weighted at 0.7, non-densely populated areas at 0.3), and the final representation value of each environmental parameter is obtained by weighted summation. Simultaneously, spatiotemporal alignment is performed: temporally, based on the timestamps after NTP synchronization, all parameter data are divided into time slices of 10 seconds, with corresponding associations between multi-source data within each time slice; spatially, for parameter data from different indoor areas and different outdoor locations, Kriging interpolation is used to generate a global environmental parameter distribution matrix, realizing the transformation from discrete point data to continuous situational data. The standardized dataset is stored in JSON format. Each data record contains eight core fields: "sensor ID, acquisition timestamp, parameter type, raw data, preprocessed data, fused data, anomaly marker (normal / suspected anomaly / confirmed anomaly), and processing method". The dataset is archived daily, and a data quality report (including indicators such as the percentage of abnormal data, data integrity, and fusion accuracy) is generated to provide high-quality data input for the subsequent weight calculation in step S3 and model building in step S4.

[0039] S3. A dynamic parameter weight adaptive system is built based on a deep reinforcement learning model. The decision weight of each environmental parameter is calculated according to time pattern, seasonal factor, user preference, and energy consumption coefficient, and the priority of parameter influence is adjusted in real time under different scenarios.

[0040] Furthermore, in the above technical solution, the weight allocation rules of the dynamic parameter weight adaptive system include: summer daytime temperature weight > light intensity weight > other parameters; winter nighttime heat preservation weight > ventilation weight > other parameters; polluted weather air purification weight > all other parameters, with the weight value range being 0.1-0.8.

[0041] It should be further explained that the core of step S3 is to build a dynamic parameter weight adaptive system that combines scene adaptability and personalization based on a deep reinforcement learning model. Through precise quantification and model iteration of the four core influencing factors, the decision weights of each environmental parameter are optimized and adjusted in real time. The specific implementation details are as follows:

[0042] First, the architecture of the deep reinforcement learning model is designed, employing an Actor-Critic dual-network structure to build the core model, ensuring the scientific nature of weight allocation and dynamic response capability. The state space is constructed based on the standardized environmental situation dataset generated in step S2, integrating quantitative data from four core input factors: the environmental parameter dimension includes real-time values ​​of key indicators such as indoor and outdoor temperature and humidity, light intensity, PM2.5 concentration, CO2 concentration, wind speed, and wind direction; the time and season dimension includes quantified time period identifiers, seasonal codes, and solar term characteristics; the user preference dimension includes personalized coefficients extracted from user preference profiles (such as temperature sensitivity and ventilation preference); and the energy consumption dimension includes real-time building energy consumption data and energy consumption coefficients corresponding to equipment operating status. The action space is defined as the decision weight allocation combination for each environmental parameter. The weight of each parameter is strictly limited to the range of 0.1-0.8, and the sum of all parameter weights is normalized to 1 to ensure the rationality of weight allocation. The reward function is designed around the system optimization goal, taking into account three core indicators: indoor comfort compliance rate (such as the percentage of time the temperature is maintained at 22-26℃ and humidity is maintained at 40%-60%), energy consumption reduction rate (compared with the energy consumption of traditional control methods in the same period in history), and air health compliance rate (the percentage of time the PM2.5 < 35μg / m³ and CO2 < 1000ppm). At the same time, a compliance penalty item for weight allocation is introduced (such as deducting rewards when violating the scenario priority rules), forming a reward mechanism of "positive incentive + reverse constraint" to guide the model to output the optimal weight combination.

[0043] Subsequently, the four input factors were quantified to provide standardized input for the model. The time pattern was divided into four time periods according to the circadian rhythm and life scenarios: early morning (0:00-6:00), morning (6:00-12:00), afternoon (12:00-18:00), and night (18:00-24:00). One-hot encoding was used to convert these into vector data that the model could recognize, while incorporating the activity intensity coefficient of people within the time period (e.g., low activity intensity in the early morning and high activity intensity in the afternoon). In addition to the basic encoding according to spring, summer, autumn, and winter, the seasonal factors were further refined into seasonal feature sub-items (e.g., summer is divided into hot summer, early summer, and late summer, and winter is divided into severe cold, early winter, and late winter). Based on the historical environmental parameter patterns of the same period, each sub-item was assigned a corresponding seasonal weight coefficient (e.g., the temperature correlation coefficient for hot summer is 0.9, and the insulation correlation coefficient for early winter is 0.8). User preference factors are based on personalized profiles built using S7, transforming users' manual adjustment behaviors into quantitative indicators. For example, if a user frequently increases the opening of window slats at 25℃, it is judged as high temperature sensitivity, and the corresponding temperature-weighted preference coefficient is set to 1.2; if a user frequently closes the blinds at a light intensity of 5000 lux, it is judged as high light sensitivity, and the corresponding light-weighted preference coefficient is set to 1.1. The preference coefficient ranges from 0.8 to 1.5, with higher coefficients indicating greater user attention to the parameter. The energy consumption coefficient is calculated through the building energy consumption monitoring module, combining the current operating power and duration of air conditioning, heating, and other equipment, as well as the differences between indoor and outdoor environments (such as indoor and outdoor temperature and humidity differences), using a linear weighting method. The value ranges from 0 to 1, with higher values ​​indicating greater current energy consumption pressure, requiring the model to place greater emphasis on energy efficiency targets.

[0044] The core process of weight calculation consists of four steps: "initial allocation - rule constraint - model optimization - output confirmation".

[0045] The first step is to generate an initial weight allocation scheme based on the four quantified factors through the Actor network model. This scheme initially reflects the influence of each factor on the weight (e.g., when the energy consumption coefficient is high, the weight of energy efficiency-related parameters is initially increased).

[0046] The second step involves introducing preset scenario rules for constraint correction, strictly adhering to the priorities specified in the document: During summer daytime (6:00-18:00), a mandatory weighting of temperature (0.6-0.8) > light intensity (0.4-0.6) > other parameters (0.1-0.3) is applied; during winter nighttime (18:00-6:00), a mandatory weighting of heat preservation (i.e., the weighting of indoor and outdoor temperature difference, 0.6-0.8) > ventilation (0.3-0.5) > other parameters (0.1-0.2) is applied; during polluted weather (PM2.5 > 75μg / m³ or CO2 > 1500ppm), a mandatory weighting of air purification (0.7-0.8) > all other parameters (0.1-0.2) is applied. If the initial allocation scheme violates the above priorities, the weight range is automatically adjusted to ensure rule compliance.

[0047] The third step is to input the weight scheme with the constraint correction into the Critic network, and evaluate the fit of the weight scheme with the three optimization objectives by combining the reward function value under the current environment. If the reward value is lower than the preset threshold (e.g., 80 points out of 100), the Actor network parameters are adjusted by gradient descent, the weight scheme is regenerated, and the iteration continues until the reward value reaches the target.

[0048] The fourth step is to confirm that the weight scheme that meets the criteria is the final decision weight, and to store it synchronously in the weight allocation model library for use as parameter input for subsequent multi-objective optimization models.

[0049] The dynamic adjustment mechanism, centered on "real-time monitoring - trigger judgment - rapid update," ensures the weights respond instantly to changes in the scenario. The system monitors input factors every 30 seconds, focusing on three types of trigger events: 1) sudden changes in environmental parameters (e.g., PM2.5 concentration increasing by ≥50μg / m³ within 10 minutes, or sudden changes in outdoor temperature by ≥5℃); 2) seasonal changes (e.g., transitioning from morning to afternoon, or seasonal transition periods); and 3) updates to user preferences (three or more manual adjustment actions reported by S7). When a trigger event is detected, the system immediately initiates a weight recalculation process, skipping the regular 5-minute update cycle to ensure the generation and application of new weights within 100ms. In the absence of trigger events, weight fine-tuning is performed every 5 minutes, optimizing the accuracy of weight allocation based on the latest environmental data and energy consumption changes. Simultaneously, a smooth weight transition mechanism is established to prevent drastic fluctuations in control strategies caused by sudden weight changes: if the difference between the new and old weights exceeds 0.2, the weights are gradually increased / decreased by 0.05 every 10 seconds to ensure system stability.

[0050] Model training and iteration employ a combined approach of "offline pre-training + online fine-tuning." In the offline pre-training phase, historical environmental data, building energy consumption data, and simulated user preference data from the past 12 months are collected to construct a training dataset containing over 100,000 samples. The model is iteratively trained until the reward function converges (reward value fluctuation ≤3% over 500 consecutive iterations), forming the initial weight allocation model. In the online fine-tuning phase, each time the system receives environmental feedback data from S7 and user manual adjustment data, key features (such as the combination of environmental parameters during adjustment and the preference tendency corresponding to the adjustment magnitude) are extracted and used as incremental training samples to fine-tune the model. Each fine-tuning iteration lasts 20 rounds, updating the model's network parameters and weight allocation rules to ensure the model continuously adapts to environmental changes and evolving user habits. Furthermore, model performance is evaluated monthly by comparing the deviation between the actual control effect after weight application and the theoretical optimization target (e.g., comfort achievement rate deviation ≤5%, energy consumption reduction rate deviation ≤8%). If the deviation exceeds the threshold, the combined optimization of offline pre-training and online fine-tuning is repeated to ensure the model's long-term effectiveness.

[0051] S4. Integrate weather forecast API data, use an LSTM network to predict microenvironmental changes in the next 2 hours, and build a multi-objective optimization model by combining the real-time dataset from step S2.

[0052] Furthermore, in the above technical solution, the input features of the LSTM network include historical environmental data, weather forecast data, seasonal features, and time period features, and the predicted output includes the temperature change range, light intensity change, and pollutant concentration trend for the next 2 hours.

[0053] It should be further explained that the core of step S4 is to complete the seamless integration of the entire process: "weather forecast API data integration - multi-source data preprocessing - LSTM network prediction - multi-objective optimization model construction." This injects forward-looking information into the optimization model through time-series prediction, while simultaneously achieving deep fusion of real-time and predicted data. The data fusion and prediction system process is as follows: Figure 2 As shown, the specific implementation details are as follows:

[0054] First, the system integrates and adapts weather forecast API data. It prioritizes real-time API interfaces provided by authoritative meteorological platforms (such as the National Meteorological Information Center API and mainstream meteorological service platform APIs) to ensure data accuracy and timeliness. API data acquisition parameters must precisely match the microenvironmental forecast requirements. Core acquisition content includes 10-minute forecasts for the next two hours: temperature (accuracy ±0.5℃), humidity (accuracy ±3%RH), light intensity level (0-5, corresponding to the 0-200,000 lux range), precipitation probability (0%-100%), wind speed and direction (wind speed accuracy ±0.3m / s, wind direction accuracy ±10°), and pollutant concentration levels (PM2.5, CO2, according to national standards). The data call frequency is set to once every 15 minutes. Each call simultaneously retrieves historical prediction deviation data (such as the difference between the predicted and actual values ​​in the previous hour) for subsequent prediction error correction. If the API call fails (e.g., network interruption, interface response timeout), a backup plan is immediately activated: temporary prediction data is generated using the three most recent valid API data and the real-time environmental data from step S2, through exponential smoothing. The call is then retried every 5 minutes until normal operation is restored. After the API data is accessed, it needs to be standardized: the meteorological platform's graded data (such as light intensity level, pollution level) is converted into continuous values ​​(e.g., light intensity level 3 corresponds to 100,000-150,000 lux, taking the median value of 125,000 lux). The timestamp is kept consistent with the standardized dataset from step S2 in UTC millisecond format, and the field names are consistent with the sensor data fields (e.g., "outdoor temperature" corresponds to "outdoor_temperature") to ensure unambiguous data fusion.

[0055] Subsequently, preprocessing of the multi-source input data is performed to provide high-quality training and prediction data for the LSTM network. The input features are integrated into four modules by category, all of which require quantization, normalization, and sequence construction processing: First, historical environmental data, which extracts the standardized dataset of the past 24 hours from step S2, samples it at 10-minute time granularity, forming a continuous sequence of 144 time steps, covering core parameters such as indoor and outdoor temperature and humidity, light intensity, PM2.5 concentration, CO2 concentration, wind speed and direction; Second, weather forecast data, which is the 10-minute forecast data for the next 2 hours after adaptation processing, forming a sequence of 12 time steps; Third, seasonal features, which use binary encoding to convert spring, summer, autumn, winter and sub-seasonal items (such as hot summer and cold winter) into vectors, while incorporating the average environmental parameter deviation values ​​corresponding to the season (such as average temperature deviation in summer and average humidity deviation in winter); Fourth, time period features, which use unique thermal encoding to convert the four time periods of early morning, morning, afternoon and night into vectors, and combine the human activity intensity coefficient (0.2-0.9) and historical energy consumption feature values ​​within the time period to enrich the representation of the time period dimension. All feature data were mapped to the [0,1] interval using the Min-Max normalization method. The normalization parameters (maximum and minimum values) were determined based on historical data statistics from the past 6 months and updated monthly. For missing values ​​in the data sequence, linear interpolation was used to complete them, ensuring the continuity of the sequence. The final input sequence was constructed with a length of 156 time steps (24 hours of historical data + 2 hours of predicted data), and each time step contained 23 feature dimensions (covering all quantitative indicators of the four major modules).

[0056] The construction and training of the LSTM network follows an "offline pre-training + online fine-tuning" model to ensure prediction accuracy and scene adaptability. The network architecture is designed as a deep structure of "input layer - hidden layer - dropout layer - output layer": the number of neurons in the input layer is consistent with the dimension of the input features (23); there are 3 hidden layers, each with 128 neurons, using ReLU activation function, and layer normalization mechanism is added to each layer to accelerate model convergence and improve stability; the dropout layer has a dropout rate of 0.2, randomly dropping some neurons to avoid model overfitting; the output layer has 3 neurons, corresponding to the three prediction targets for the next 2 hours (temperature change, light intensity change, and pollutant concentration trend), using Linear activation function, and the output is a continuous numerical value. In the offline pre-training phase of model training, historical environmental data, API forecast data, and actual environmental change data from the past 12 months are collected to construct a training set containing over 50,000 samples (samples consist of input sequences and corresponding actual output sequences). The Adam optimizer is used, with the learning rate initially set to 0.001 and adjusted using a learning rate decay strategy (decreasing by 10% every 100 iterations). The mean squared error (MSE) loss function is used. Training iterations stop when the loss value fluctuation is ≤2% for 300 consecutive iterations, and the optimal pre-trained model is saved. In the online fine-tuning phase, the system extracts the latest real-time data and API update data from step S2 every hour to generate 10 incremental training samples for fine-tuning the model (20 iterations per time). Simultaneously, the prediction error is calculated (evaluated using MAE and RMSE). If the error exceeds a preset threshold (MAE > 0.05), the number of fine-tuning samples is increased (to 20) and the learning rate is adjusted (temporarily increased to 0.0005) to ensure that the model adapts to environmental changes in real time.

[0057] The execution and error correction process for microenvironment prediction must ensure the reliability of the prediction results. During prediction, the preprocessed input sequence is fed into a trained LSTM network, which outputs three prediction results for the next two hours, every 10 minutes: temperature change (in °C, representing the difference from the current temperature), light intensity change (in lux, representing the difference from the current light intensity), and pollutant concentration trend (in μg / m³ or ppm, representing the difference from the current concentration). After the prediction results are output, the error correction mechanism is immediately activated: combining the historical prediction deviation of the API data with the recent prediction error of the LSTM model, the correction coefficient for each prediction time step is calculated (e.g., if the historical API deviation is -0.3 °C and the recent model error is 0.2 °C, then the correction coefficient is -0.1), and the correction coefficient is added to the prediction result of the corresponding time step; at the same time, the latest real-time data in step S2 is compared with the prediction value at the start of the prediction. If the deviation is >10%, the weight allocation of the input sequence is readjusted (increasing the weight of real-time data to 0.7), and the prediction is re-executed. The final generated prediction data needs to be converted into a format consistent with the standardized dataset in step S2, and correlated according to time steps to form a complete environmental situation sequence of "real-time data + prediction data for the next 2 hours", providing input with both real-time and forward-looking characteristics for the multi-objective optimization model.

[0058] The construction of the multi-objective optimization model is based on "real-time + prediction" dual-dimensional data, and the mathematical expression of the optimization objective and constraints is clearly defined. The input data of the model is divided into two parts: one is the latest standardized environmental situation dataset generated in step S2 (reflecting the current environmental state), and the other is the LSTM prediction data for the next 2 hours after error correction (reflecting the future trend of environmental change). Data fusion is achieved through spatiotemporal alignment (corresponding in 10-minute time steps in time and matching indoor and outdoor areas in space) to form the global environmental state input of the model. The mathematical expression of the optimization objectives closely revolves around three core aspects: indoor comfort, energy efficiency, and air health. The indoor comfort objective function uses the percentage of time spent meeting the standards for temperature (22-26℃), humidity (40%-60%), and light intensity (3000-10000 lux) as core indicators, and uses a Gaussian function to quantify the comfort score (1.0 for meeting the standard, and a lower score for deviations). The energy efficiency objective function uses the building energy consumption reduction rate as the core, and establishes a mapping relationship between energy consumption and louver opening / angle based on the difference between indoor and outdoor environments (temperature, humidity, and light) and the operating power of equipment such as air conditioning and heating, with the goal of minimizing energy consumption. The air health objective function uses the percentage of time spent meeting the standards for PM2.5 concentration (<35μg / m³) and CO2 concentration (<1000ppm) as the core, and uses a piecewise function for quantification (1.0 for meeting the standard, 0.5 for light pollution, and 0.1 for heavy pollution). The constraints are explicitly defined as weather conditions: when the API predicts a precipitation probability of ≥50% in the next 2 hours, the maximum opening of the venetian blinds is limited to ≤30%; when the predicted wind speed is ≥8m / s, the maximum opening is limited to ≤20%; when the predicted pollution level is severe (PM2.5 > 115μg / m³ or CO2 > 2000ppm), the opening is limited to ≤10%, while prioritizing the adjustment weights of air purification-related parameters. The model output is the feasible solution space for venetian blind opening (0%-100%) and angle (0°-90°), providing a solution basis for the optimization algorithm in step S5.

[0059] S5. Using indoor comfort, energy efficiency, and air health as optimization objectives and weather conditions as constraints, calculate the target opening and angle of the louvers through a multi-objective optimization algorithm.

[0060] Furthermore, in the above technical solution, the multi-objective optimization algorithm adopts a non-dominated sorting genetic algorithm with an elite strategy to search for Pareto optimal solutions for the three objective functions of indoor comfort, energy efficiency, and air health, and selects the final execution solution from the optimal solution set according to the dynamic weights obtained in step S3.

[0061] It should be further explained that the core of step S5 is a non-dominated sorting genetic algorithm (NSGA-II) with an elitist strategy. Under weather constraints, it performs synergistic optimization on three objectives: indoor comfort, energy efficiency, and air health, ultimately solving for the optimal opening and angle of the blinds. The multi-objective optimization and execution control logic is as follows: Figure 3 As shown, the specific implementation details are as follows:

[0062] First, mathematical modeling and quantitative calculations of the three major optimization objectives were completed to ensure that the objective functions could be directly used for algorithm solutions. The indoor comfort objective function (f1) is constructed using a multi-dimensional weighted Gaussian function, comprehensively considering the compliance of the three core indicators: temperature, humidity, and light intensity. Temperature comfort is optimally within the range of 22-26℃, and the comfort score decreases according to a Gaussian curve when deviating from this range (e.g., a score of 0.8 at 21℃ or 27℃, and 0.3 at 18℃ or 30℃). Humidity comfort is optimally within the range of 40%-60%RH, and decreases linearly when deviating from this range (e.g., a score of 0.6 at 30%RH or 70%RH). Light comfort is optimally within the range of 3000-10000 lux, and the score drops below 0.2 when it is too low (<1000 lux) or too high (>20000 lux). The weight ratio of each indicator is determined by the dynamic weight in step S3 (e.g., daytime temperature in summer has a weight of 0.6, light intensity 0.3, and humidity 0.1). The weighted sum is used to obtain the final value of f1, which ranges from [0,1]. The closer the value is to 1, the higher the comfort level.

[0063] The energy efficiency objective function (f2) focuses on minimizing building energy consumption and establishes a mapping relationship between energy consumption and the opening and angle of the louvers. Based on the real-time indoor and outdoor temperature and humidity difference and light intensity from step S2, and the predicted data for the next 2 hours from step S4, combined with the rated power and operating status of equipment such as air conditioners and heaters, an energy consumption calculation model is constructed. When the indoor and outdoor temperature difference is small (summer < 5℃, winter < 8℃), increasing the louver opening can reduce the equipment operating time, and energy consumption decreases exponentially with the increase of opening. When the temperature difference is large, excessive opening will lead to an increase in energy consumption, forming a U-shaped energy consumption curve. f2 uses the energy consumption reduction rate for quantification, which is the ratio of energy consumption under the current control strategy to energy consumption under the "louver fixed closed" mode, with a value range of [0,1]. The closer to 1, the higher the energy efficiency.

[0064] The air health objective function (f3) employs a piecewise weighted function, focusing on the compliance status of PM2.5 and CO2 concentrations: PM2.5 concentration < 35 μg / m³ scores 1.0, 35-75 μg / m³ scores 0.5, and > 75 μg / m³ scores 0.1; CO2 concentration < 1000 ppm scores 1.0, 1000-1500 ppm scores 0.6, and > 1500 ppm scores 0.2. Based on the dynamic weights determined in step S3, the proportions of the two indicators are determined (e.g., PM2.5 weight 0.7, CO2 weight 0.3 during polluted weather), and the weighted sum is used to obtain the final value of f3, ranging from [0,1]. The closer the value is to 1, the higher the air health level. All three objective functions require Min-Max normalization to eliminate dimensional differences and ensure the fairness of the algorithm's solution.

[0065] The constraints are then quantified and transformed into mathematical constraints that the algorithm can recognize. Hard constraints include three categories: first, precipitation constraints, where if step S4 predicts a precipitation probability of ≥50% for the next 2 hours, the louver opening is limited to [0%, 30%] to prevent rainwater infiltration; second, wind speed constraints, where if the predicted wind speed is ≥8m / s, the opening is limited to [0%, 20%] to prevent damage to the louvers; and third, pollution constraints, where if the predicted PM2.5 >115μg / m³ or CO2 >2000ppm, the opening is limited to [0%, 10%] to prioritize clean indoor air. Soft constraints include mechanical structure constraints: opening ∈ [0%, 100%], angle ∈ [0°, 90°] (0° for fully closed, 90° for fully open), and the angle adjustment must match the opening (e.g., if the opening is <20%, the angle is limited to 30°-60° to avoid excessively low ventilation efficiency). All constraints are incorporated into the algorithm's solution process using the penalty function method. Solutions that violate the constraints will be assigned extremely low fitness values ​​to ensure they are not included in the optimal solution set.

[0066] The parameter configuration of the NSGA-II algorithm needs to balance solution accuracy and real-time performance: the population size is set to 80 individuals, with each individual corresponding to a set of solution vectors (opening degree, angle). The opening degree is discretized with a step size of 1% (0%-100%), and the angle is discretized with a step size of 0.1° (0°-90°) to ensure the refinement of the solution; the number of iterations is set to 150 generations to ensure that the algorithm converges to a stable Pareto front while controlling the single solution time to ≤500ms, meeting the real-time control requirements of the system; the crossover probability is set to 0.85, using a simulated binary crossover operator to perform crossover operations on the opening degree and angle genes of individuals respectively, retaining superior genes; the mutation probability is set to 0.03, using a polynomial mutation operator to perform small-amplitude random perturbations on gene values ​​to maintain population diversity and avoid the algorithm getting trapped in local optima.

[0067] The algorithm execution flow follows the core logic of NSGA-II, consisting of seven steps: initialization, non-dominated sorting, crowding calculation, selection, crossover, mutation, and iterative convergence. Step 1: Population initialization. Within the solution space allowed by the constraints, 80 valid individuals are randomly generated to ensure the initial population covers different combinations of opening and angle. Step 2: Non-dominated sorting. For each individual in the population, the number of times it is dominated (i.e., the number of other individuals superior to that individual) is calculated, and domination levels are assigned. Level 1 represents non-dominated solutions (Pareto optimal solution candidates), Level 2 represents solutions dominated by Level 1, and so on. Step 3: Crowding calculation. For each individual within a domination level, the distance between it and its neighboring individuals in the target space is calculated. A larger distance indicates lower crowding and better individual diversity, which is used to prioritize retaining individuals with high diversity during subsequent selection.

[0068] The fourth step, selection, employs a "roulette wheel selection + elite retention" strategy. Parent individuals are selected from the current population and the previous generation's elite population (level 1 non-dominated solutions) based on fitness values ​​(combining dominance level and crowding) to ensure the inheritance of superior genes. The fifth step, crossover, pairs the selected parent individuals and crossovers the opening and angle genes according to the crossover probability to generate offspring individuals. For example, crossing parent A (50% opening, 45° angle) with parent B (30% opening, 60° angle) may generate offspring C (42% opening, 52° angle), while ensuring the offspring meet the constraints. The sixth step, mutation, randomly perturbs the genes of the offspring individuals. For example, 50% opening mutates to 51%, and 45° angle mutates to 44.7°. After mutation, constraint compliance must be verified; if violations occur, mutation is repeated.

[0069] Step 7: Iterative convergence: Merge the offspring population with the previous generation of elite population, repeat the non-dominated sorting and crowding calculation, and select the new generation population until the number of iterations reaches 150 generations, or the change of the Pareto front is ≤0.5% for 20 consecutive generations (i.e., the maximum fluctuation of each objective function value in the optimal solution set is ≤0.005). At this time, the algorithm is considered to have converged, and the non-dominated solution set of level 1 is the final Pareto optimal solution set.

[0070] Finally, the optimal solution is selected and determined: All individuals in the Pareto optimal solution set are extracted, and combined with the dynamic parameter weights obtained in step S3, the comprehensive evaluation value of each individual is calculated—Comprehensive evaluation value = f1×w1 + f2×w2 + f3×w3, where w1, w2, and w3 are the dynamic weights of indoor comfort, energy efficiency, and air health, respectively (the sum of the three is 1). For example, during summer daytime, w1=0.5, w2=0.3, w3=0.2; during winter nighttime, w1=0.4, w2=0.5, w3=0.1; and during polluted weather, w1=0.2, w2=0.1, w3=0.7. The individual with the highest comprehensive evaluation value is selected, and its corresponding opening and angle become the target control parameters for the blinds, which are synchronously output to the drive mechanism in step S6 for execution. If multiple optimal solutions have the same comprehensive evaluation value, the solution with the smallest change in opening and angle is selected first to reduce frequent adjustments to the drive mechanism and improve system stability.

[0071] S6. The target control command is executed through a silent drive mechanism to achieve window blade adjustment with an accuracy of 0.1°. At the same time, anti-pinch protection and status monitoring are achieved through a motor torque sensor and a micro-vibration sensor.

[0072] Furthermore, in the above technical solution, the drive mechanism adopts a magnetic levitation motor with an operating noise of <25dB, is equipped with a power failure mechanical holding mechanism, and the self-powered system includes a flexible solar thin film integrated into the window frame and a wind-induced vibration energy harvesting device, which is combined with a micro supercapacitor to achieve energy autonomy.

[0073] Furthermore, in the above technical solution, the anti-pinch protection monitors load changes in real time through a motor torque sensor. When the torque exceeds a preset threshold of 0.5 N·m, the window leaf is immediately adjusted in the opposite direction at an adjustment speed of ≤0.5° / s.

[0074] It should be further explained that the core of step S6 is to achieve high-precision execution of the target opening degree and angle of the louvers through a silent drive mechanism, while simultaneously relying on dual sensor linkage to complete anti-pinch protection and full-state monitoring. Combined with the energy adaptation strategy of the self-powered system, the execution process is ensured to be accurate, safe, low-noise, and stable. The specific implementation details are as follows:

[0075] First, the system receives and parses the target control commands. The system control gateway encapsulates the target parameters (opening: 0%-100%, angle: 0°-90°) output in step S5 into standardized control commands, and sends them to the control module of the drive mechanism via a wireless mesh network using an encrypted transmission protocol (AES-128), with the transmission delay controlled within 50ms. After receiving the command, the control module immediately initiates the command verification process: First, it verifies the validity of the parameters (the opening must be in the range of 0%-100%, the angle must be in the range of 0°-90°, and it must meet the matching constraints of the opening and angle, such as the angle must be in the range of 30°-60° when the opening is <20%). If the parameters are invalid, it returns "invalid command" feedback and refuses to execute. If the parameters are valid, it converts the opening and angle values ​​into electrical signal commands that the drive mechanism can recognize—the opening corresponds to the number of rotations of the motor (converted according to the window blade transmission ratio, 1 rotation corresponds to 5% opening), and the angle corresponds to the precise rotation angle of the motor (converted with 0.1° as the minimum step unit). At the same time, it combines the actual position data of the current window blade (provided in real time by the encoder) to plan the shortest adjustment path and avoid invalid reciprocating actions.

[0076] The high-precision execution of the drive mechanism is based on a magnetic levitation motor, coupled with a closed-loop servo control system to achieve an adjustment accuracy of 0.1°. The motor selection must match the window leaf specifications: when the weight of a single window leaf is ≤5kg, a 50W miniature magnetic levitation motor is selected; when the weight is 5-10kg, an 80W motor is selected to ensure that the output torque meets the adjustment requirements and there is no redundant energy consumption. The motor control employs a three-loop servo strategy: position-speed-torque. The position loop uses an absolute encoder (1024 lines resolution) to acquire the actual position of the window leaf in real time, providing feedback for every 0.1° adjustment. The position signal is compared with the target position to calculate the deviation, and automatic fine-tuning is initiated when the deviation exceeds 0.2°. The speed loop dynamically adjusts the rotational speed based on the adjustment range: 1° / s for an adjustment range >10°, 0.5° / s for an adjustment range of 5°-10°, and 0.3° / s for an adjustment range <5°, balancing adjustment efficiency and accuracy. The torque loop monitors the output torque in real time using a motor torque sensor, ensuring the torque remains stable within the 0.1-0.4 N·m range (when anti-pinch protection is not triggered) to avoid overload damage. For noise control, the motor uses magnetic levitation bearings to eliminate mechanical friction noise. The outer shell is wrapped with a 5mm thick silicone sound insulation sleeve, and the transmission gears are made of POM engineering plastic. Lubrication reduces meshing noise, ensuring that operating noise is strictly controlled below 25dB, meeting indoor quiet operation requirements. The coordinated adjustment logic of opening and angle is as follows: First, drive the motor to move the window leaf along the guide rail to the position corresponding to the target opening (the angle remains in the initial state during the translation process). Then, use the motor's rotation angle adjustment function to fine-tune the window leaf to the target angle. Set a 500ms buffer pause when switching between translation and rotation to avoid accuracy deviation caused by motion impact.

[0077] The anti-pinch protection mechanism employs a dual-protection scheme of "torque monitoring + vibration-assisted judgment" to ensure the safety of personnel and objects. The sampling frequency of the motor torque sensor is set to 50Hz to capture load changes in real time during the adjustment process. The preset anti-pinch threshold is 0.5N・m (which cannot be changed arbitrarily and has passed safety testing and verification). When the torque value exceeds 0.5 N·m after three consecutive samplings, the control module immediately triggers the anti-pinch response: it cuts off the adjustment drive signal immediately, stopping the motor's current action; after a 100ms delay, it starts reverse adjustment, with the reverse adjustment speed strictly controlled within 0.5° / s to avoid secondary damage caused by the reverse action; after the reverse adjustment reaches 3°, it pauses and continuously monitors the torque change. If the torque recovers to below the threshold (<0.3 N·m), it is determined that the foreign object clamping has been released, and after a 2-second delay, it resumes executing the target command along the original path (or finely adjusts the path according to the foreign object's position); if the torque is still higher than the threshold after the reverse adjustment, it is determined that the clamping is still present, and all adjustment actions are immediately stopped. An "anti-pinch alarm" notification is sent to the user terminal through the gateway, and the current window leaf position remains unchanged until the user manually resets or removes the foreign object. To avoid false triggering, a micro-vibration sensor is introduced to assist in the judgment: when the torque exceeds the threshold, the vibration amplitude and frequency of the micro-vibration sensor are detected simultaneously. If the vibration amplitude is <0.1mm and the frequency does not fluctuate significantly (determined to be mechanical jamming rather than foreign object clamping), the anti-pinch reverse action is not activated, but a "mechanical fault" alarm is sent to prompt the user to perform maintenance.

[0078] The status monitoring covers three dimensions: drive mechanism, window blades, and operating environment, enabling full-process fault early warning. Motor status monitoring is accomplished through integrated temperature and current sensors: the temperature sensor monitors the motor coil temperature in real time, with a normal operating temperature of <70℃. If the temperature exceeds 75℃, a "motor overheating" alarm is triggered, and the motor power is automatically reduced and the adjustment speed is slowed down. The current sensor monitors the motor operating current. If the current fluctuation exceeds ±20% and lasts for 3 seconds, it is determined to be a motor abnormality (such as coil short circuit or bearing jamming), and operation is immediately stopped and an alarm is triggered. Window blade status monitoring relies on an absolute encoder and a micro-vibration sensor: the encoder provides real-time feedback on the actual opening and angle of the window blades, compares it with the target value to calculate the deviation, and if the deviation exceeds 0.3° and lasts for 5 seconds, automatic calibration is initiated (the drive motor is fine-tuned to the target position). If the deviation cannot be eliminated after 3 calibrations, it is determined to be a "positioning abnormality." The micro-vibration sensor monitors the vibration amplitude (normally <0.3mm) and frequency during the window blade adjustment process. When the vibration abnormally increases (>0.5mm), it is determined to be a loose window blade or guide rail deformation, and a "window blade status abnormality" alarm is sent. The operating environment monitoring focuses on the temperature and humidity around the drive mechanism (collected by temperature and humidity sensors integrated into the drive module). When the humidity exceeds 85%RH, moisture protection is activated (the control module automatically cuts off unnecessary circuits, retaining only the monitoring function) to prevent moisture from causing motor short circuits or circuit failures. All monitoring data is uploaded to the system gateway every 2 seconds to form a status monitoring log, providing a basis for troubleshooting and maintenance.

[0079] The energy adaptation and mechanical protection of the drive mechanism must align with the system's energy management strategy. The voltage status of the micro supercapacitor in the self-powered system is the core judgment criterion: when the capacitor voltage is ≥3.6V (sufficient energy), the drive mechanism executes the normal control mode, completing the adjustment according to the preset speed and accuracy, while continuously replenishing the capacitor; when the voltage is between 2.8V and 3.6V (insufficient energy), it switches to a low-power execution mode, reducing the motor speed (≤0.3° / s) and the sensor sampling frequency (torque sensor sampling frequency reduced to 20Hz), prioritizing adjustment accuracy and safety functions; when the voltage is <2.8V (critical energy), unnecessary adjustment actions are suspended, only responding to safety-related commands (such as anti-pinch protection, emergency shutdown), while simultaneously initiating energy recovery (accelerating energy replenishment through a wind-induced vibration energy harvesting device) until the voltage rises back to above 2.8V. The power-off mechanical holding mechanism employs an electromagnetic locking design. Under normal power supply, the lock unlocks without affecting window slat adjustment; in the event of a power outage or energy criticality, the lock automatically pops out and engages with the window slat drive gear to prevent wind or external forces from causing the window slats to sway or shift, ensuring structural safety. Furthermore, the drive mechanism's housing features an IP65 waterproof and dustproof design, suitable for both indoor and outdoor installation environments, preventing rainwater and dust from affecting the operation of internal components.

[0080] Finally, the execution result feedback process ensures closed-loop connection: after the drive mechanism completes the target adjustment, it feeds back the "execution status (success / failure), actual opening, actual angle, and monitoring data summary" to the system gateway via the wireless Mesh network within 100ms. The gateway compares the feedback data with the target parameters in step S5, generates an execution effect evaluation (such as accuracy compliance rate, whether abnormal protection is triggered), and pushes it to step S7 simultaneously. As a core component of the environmental feedback data, it supports the updating of user preference profiles and weight allocation models.

[0081] S7. Collect environmental feedback data and user manual adjustment behavior data after control, and update user preference profiles and weight allocation models through personalized habit learning engine and unsupervised learning algorithm to complete closed-loop learning.

[0082] Furthermore, in the above technical solution, the personalized habit learning engine identifies the time, environmental conditions, and adjustment range of user manual adjustments, establishes exclusive preference profiles for family members, and automatically generates scenario-based control strategies such as work mode, sleep mode, and guest mode.

[0083] Furthermore, in the above technical solution, the method also includes a system energy management strategy. The self-powered system monitors the voltage status of the micro supercapacitor, executes the normal control mode when energy is sufficient, and switches to a low-power sensing and basic maintenance mode when energy is insufficient, prioritizing the regulation of safety and health-related parameters.

[0084] It should be further explained that the core of step S7 is to construct a closed-loop learning system of "data collection-analysis-optimization-application" through full-dimensional data collection, refined preprocessing, personalized preference mining, and incremental model updates. This continuously iterates the user preference profile and weight allocation model, making the system control strategy more aligned with user habits. Specific implementation details are as follows:

[0085] First, comprehensive data collection is conducted, covering two core types: environmental feedback data and user manual adjustment behavior data, to ensure data integrity and relevance. Environmental feedback data collection relies on the multi-level intelligent sensor network in step S1. After the drive mechanism completes the target adjustment, key indoor and outdoor parameters are continuously collected for 3 minutes at a frequency of 10 seconds per collection. These parameters include indoor temperature and humidity, light intensity, PM2.5 concentration, CO2 concentration, and occupancy status, as well as outdoor temperature and humidity, wind speed and direction, and pollution level. At the same time, the actual opening degree and angle of the window slats and the motor operating status (torque, temperature, current) are recorded, forming a complete data chain of "adjusted environment - equipment status". User manual adjustment behavior data adopts a "trigger-based + full recording" mode. When a user manually intervenes in adjusting the blinds via physical buttons, mobile app, voice commands, etc., the system immediately initiates data collection: recording the adjustment trigger timestamp, trigger method, operator identifier (matched by device login account or behavioral characteristics, such as frequently used adjustment periods and intensity habits), system automatic control parameters before adjustment (opening degree, angle), target parameters after adjustment (opening degree, angle), real-time environmental data before and after adjustment (consistent with the dimensions of environmental feedback data collection), and environmental change trend after adjustment (continuously collecting data for 5 minutes). All collected data is stored in JSON format, with fields including "data type, trigger scenario, timestamp, environmental parameter set, device status set, adjustment action set, and operator identifier," ensuring consistency with the standardized dataset format of steps S2 and S3, facilitating data retrieval across steps.

[0086] The data preprocessing stage focuses on "denoising, alignment, and filtering" to provide a high-quality data foundation for preference mining and model updates. First, data denoising is performed: for high-frequency fluctuations in environmental feedback data (such as sudden changes in illumination or minor sensor errors), a mean filtering algorithm (5-times moving average) consistent with step S2 is used for smoothing. For outliers in user manual adjustment data (such as accidental adjustments followed by immediate resets, or invalid commands exceeding the mechanical range), a "time continuity check + logical rationality judgment" is used to eliminate them—if the adjustment action duration is less than 1 second, the original state is restored within 3 seconds after adjustment, or the adjustment parameter exceeds the range of 0%-100% (opening) or 0°-90° (angle), it is judged as a misoperation, marked, and excluded from subsequent analysis. Second, spatiotemporal alignment is performed: the environmental feedback data and user manual adjustment data are precisely matched by timestamps to ensure a one-to-one correspondence between "adjustment action and environmental change." For example, if a user manually adjusts the window flap at 14:30:25, environmental data from one minute before and after that time point is synchronously correlated to clarify the causal relationship between the adjustment behavior and the environmental state. Finally, data filtering is performed: valid adjustment behavior data (non-misoperation, non-repetitive adjustment) is extracted and classified by "scenario dimension" (such as work hours, sleep hours, polluted weather, sunny days, etc.). Repeated adjustment behaviors of the same user in similar scenarios are selected as the core samples for preference mining to ensure the representativeness of the samples.

[0087] Personalized preference mining is achieved through the collaboration of a personalized habit learning engine and unsupervised learning algorithms to accurately extract implicit preferences behind user behavior. First, scene feature clustering is performed. The K-means clustering algorithm is used to divide effective regulatory behavior data into scenes. Clustering features include time dimensions (time period, season, weekday / weekend), environmental dimensions (indoor / outdoor temperature and humidity, light intensity, pollution level, number of people), and equipment status dimensions (opening / angle before adjustment, motor operating mode). The number of clusters is dynamically set according to the size of the household (4-6 clusters for households of 2-4 people). This ultimately forms typical scene clusters such as work mode, sleep mode, guest mode, polluted day mode, and sunny day mode, with each cluster corresponding to a type of high-frequency user activity scene. Then, the core preference features are extracted: For each scene cluster, the mapping relationship between "environmental parameter combination - adjustment action" is analyzed by association rule mining algorithm. For example, in the "sleep mode" cluster, if the user repeatedly adjusts the opening from 30% to 10% and the angle from 45° to 30° after 23:00, when the indoor temperature is 24℃ and the light intensity is <500 lux, the core preference feature "night + low light + suitable temperature → low opening + small angle" is extracted. The user's sensitivity to each parameter is identified by density peak clustering. For example, if the user frequently fine-tunes the angle when the temperature is 25-26℃, it is judged as high temperature sensitivity. If the user quickly closes the window leaf when the light intensity is >10000 lux, it is judged as high light sensitivity. The sensitivity is quantified into a coefficient of 0.8-1.5 (the higher the coefficient, the stronger the sensitivity). At the same time, a family member exclusive preference profile is established, assigning a unique identifier to each user and recording their preference characteristics, sensitivity coefficient, adjustment habits (such as large-scale adjustment or small-scale fine-tuning) and forbidden parameter ranges (such as light intensity >8000 lux and opening degree >60%) in each scene cluster. The profile is stored according to scene categories and supports dynamic supplementation and the elimination of outdated preferences (such as summer scene preferences automatically going into hibernation in winter and being reactivated and updated the following summer).

[0088] The incremental model update focuses on the iterative optimization of the weight allocation model, adopting an "unsupervised learning + incremental training" mode to ensure that the model can respond to changes in user preferences in real time. First, the filtered effective adjustment data is transformed into training samples that the model can recognize: taking the scene feature vector (quantized time, environment, and device state parameters) as input and the parameter weight requirements corresponding to the user's adjustment action (such as the temperature parameter having a more significant impact after the user's adjustment, corresponding to the temperature weight preference value) as output, an incremental training dataset is constructed. For the dynamic parameter weight adaptive system in step S3, the online gradient descent algorithm in unsupervised learning is used for model fine-tuning: incremental training samples are input into the model in batches (every batch consists of 5 valid adjustment data points), with each fine-tuning iteration lasting 20 rounds and a learning rate of 0.0005 to avoid sudden weight changes. During fine-tuning, the focus is on optimizing the mapping relationship between scenarios and parameter weights. For example, if a user increases the heat preservation weight multiple times in sleep mode, the model automatically increases the base coefficient of the heat preservation weight for winter nights (from the original 0.6-0.8 to 0.7-0.85). At the same time, the preference coefficients of each parameter weight are corrected based on the user's sensitivity coefficient (the preference coefficient of highly sensitive parameters is multiplied by 1.2-1.3). If a user develops a new adjustment habit (such as turning on the ventilation mode at a high level for the first time on a smoggy day and continuing for more than 3 times), the system identifies this new scenario through density clustering, automatically adds "smoggy day ventilation mode" to the preference profile, and initializes the weight allocation rule for this scenario based on the association rule mining results. Subsequent optimization is carried out using more adjustment data.

[0089] A closed-loop verification and iteration mechanism ensures the effectiveness and stability of model updates. After a model update, the system prioritizes using the new weight allocation rules to generate control commands in the corresponding scenario, while monitoring user feedback at a frequency of 1 minute per instance: if the user does not manually adjust again, and the comfort compliance rate of the adjusted environmental parameters (calculated according to the comfort objective function in step S5) is ≥90%, the model update is deemed effective, and the new weight configuration is retained; if the user manually corrects again within 30 minutes, or the comfort compliance rate is <80%, a second fine-tuning is initiated—extracting the data from this correction behavior, analyzing the reasons for model deviation (such as not considering the user's preference for a specific lighting angle), adjusting the weight coefficients of relevant parameters, and regenerating the control strategy. In addition, the preference profile and model are integrated and optimized monthly: outdated preferences (such as temporary sunshade preferences in summer that are not reused in winter), similar scenarios (such as merging "weekday evening" and "weekend evening" preferences into "nighttime leisure mode"), and the overall model fit is evaluated (if the frequency of manual adjustment by users decreases by ≥30% compared to the previous month, it is considered to be well-fitted; if the decrease is less than 10%, the offline training sample size is increased and re-optimized) to ensure that the system maintains accurate fit to user preferences in the long term and forms a closed-loop learning effect of continuous iteration.

[0090] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic opening and closing method for louvered ventilation windows based on a multi-parameter intelligent sensing system, characterized in that, Includes the following steps: S1. Construct a multi-level intelligent sensor network to simultaneously collect multi-source sensor data, including indoor and outdoor data and window status data. S2. The collected multi-source sensor data is fused and processed. Through cross-verification of optical, electronic and airflow signals, abnormal data is removed and a standardized environmental situation dataset is generated. S3. A dynamic parameter weight adaptive system is built based on a deep reinforcement learning model. The decision weight of each environmental parameter is calculated according to time pattern, seasonal factor, user preference, and energy consumption coefficient, and the priority of parameter influence is adjusted in real time under different scenarios. S4. Integrate weather forecast API data, use an LSTM network to predict microenvironmental changes in the next 2 hours, and build a multi-objective optimization model by combining the real-time dataset from step S2. S5. Using indoor comfort, energy efficiency, and air health as optimization objectives and weather conditions as constraints, calculate the target opening and angle of the louvers through a multi-objective optimization algorithm. S6. The target control command is executed through a silent drive mechanism to achieve window blade adjustment with an accuracy of 0.1°. At the same time, anti-pinch protection and status monitoring are achieved through a motor torque sensor and a micro-vibration sensor. S7. Collect environmental feedback data and user manual adjustment behavior data after control, and update user preference profiles and weight allocation models through personalized habit learning engine and unsupervised learning algorithm to complete closed-loop learning.

2. The automatic opening and closing method for a louvered ventilation window based on a multi-parameter intelligent sensing system according to claim 1, characterized in that: The multi-level intelligent sensor network adopts a wireless mesh networking method. The outdoor array, indoor array, and window patch sensors all integrate multi-parameter fusion SoC chips. The outdoor array collects light intensity / spectrum, temperature and humidity, rainwater signal, wind speed and direction, PM2.5 concentration, and CO2 concentration; the indoor array collects personnel occupancy status, temperature and humidity, CO2 concentration, and light distribution; and the window patch sensors collect window leaf vibration status and motor torque data.

3. The automatic opening and closing method for a louvered ventilation window based on a multi-parameter intelligent sensing system according to claim 1, characterized in that: The weight allocation rules of the dynamic parameter weight adaptive system include: summer daytime temperature weight > light intensity weight > other parameters; winter nighttime heat preservation weight > ventilation weight > other parameters; polluted weather air purification weight > all other parameters, with the weight value ranging from 0.1 to 0.

8.

4. The automatic opening and closing method for a louvered ventilation window based on a multi-parameter intelligent sensing system according to claim 1, characterized in that: The input features of the LSTM network include historical environmental data, weather forecast data, seasonal features, and time period features. The predicted output includes the temperature change range, light intensity change, and pollutant concentration trend for the next 2 hours.

5. The automatic opening and closing method for a louvered ventilation window based on a multi-parameter intelligent sensing system according to claim 1, characterized in that: The personalized habit learning engine identifies the time, environmental conditions, and adjustment range of user manual adjustments, establishes exclusive preference profiles for family members, and automatically generates scenario-based control strategies such as work mode, sleep mode, and visitor mode.

6. The automatic opening and closing method for a louvered ventilation window based on a multi-parameter intelligent sensing system according to claim 1, characterized in that: The drive mechanism uses a magnetic levitation motor with an operating noise of <25dB. It is equipped with a mechanical holding mechanism in case of power failure. The self-powered system includes a flexible solar film integrated into the window frame and a wind-induced vibration energy harvesting device, which is combined with a micro supercapacitor to achieve energy autonomy.

7. The automatic opening and closing method for a louvered ventilation window based on a multi-parameter intelligent sensing system according to claim 1, characterized in that: The anti-pinch protection monitors load changes in real time through a motor torque sensor. When the torque exceeds a preset threshold of 0.5 N·m, the window blades are immediately adjusted in the opposite direction at a speed of ≤0.5° / s.

8. The automatic opening and closing method for a louvered ventilation window based on a multi-parameter intelligent sensing system according to claim 1, characterized in that: The abnormal data removal method in step S2 is to set the normal fluctuation range and physical correlation constraints of each sensor parameter. When a single point data exceeds the threshold range or has a logical conflict with at least two of the related signals in optical, electronic, and airflow signals, it is determined to be an abnormal value and smooth interpolation replacement is performed.

9. The automatic opening and closing method for a louvered ventilation window based on a multi-parameter intelligent sensing system according to claim 1, characterized in that: The multi-objective optimization algorithm employs a non-dominated sorting genetic algorithm with an elitist strategy to search for Pareto optimal solutions for the three objective functions of indoor comfort, energy efficiency, and air health, and selects the final execution solution from the optimal solution set based on the dynamic weights obtained in step S3.

10. The automatic opening and closing method for a louvered ventilation window based on a multi-parameter intelligent sensing system according to claim 1, characterized in that: The method also includes a system energy management strategy. The self-powered system monitors the voltage status of the micro supercapacitor, executes a normal control mode when energy is sufficient, and switches to a low-power sensing and basic maintenance mode when energy is insufficient, prioritizing the regulation of safety and health-related parameters.

Images