Intelligent air conditioner personalized temperature control system and method based on UWB radar sign perception

Abstract

According to the intelligent air conditioner personalized temperature control system and method based on UWB radar sign perception, vital sign signals such as the respiratory rate, the heart rate and the body movement of a human body are collected in a non-contact mode through a UWB radar sensor array, parameters such as the environment temperature and humidity and the air flow speed are synchronously collected, multi-dimensional labeling processing is carried out, and the temperature control accuracy is improved. And constructing a multi-modal human body thermal comfort perception data set. Designing a radar signal preprocessing and sign feature extraction module based on a signal decomposition algorithm and a deep learning model, and respectively extracting vital sign features and human body activity features; and meanwhile, environment parameter feature vectors are constructed. And a personalized thermal comfort identification model based on an attention mechanism and multi-task learning is constructed, and multi-source information fusion and thermal comfort state prediction are realized. A temperature control strategy is optimized through a deep reinforcement learning framework, and a self-adaptive control mechanism based on real-time physical sign feedback is established. A UWB radar sensing system and an air conditioner control system are integrated, and intelligent temperature control decision based on the physiological state of the human body is achieved.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically relating to a personalized temperature control system and method for intelligent air conditioning based on UWB radar signature perception. Background Technology

[0002] In the field of modern built environment and human health, indoor thermal comfort has become an important standard for measuring quality of life. Traditional air conditioning systems generally use fixed temperature thresholds or simple human body sensing mechanisms, which cannot detect changes in the actual physiological state of the human body, resulting in a significant difference between environmental control and individual needs. This mechanized temperature control mode not only wastes energy but may also affect human health and work efficiency due to discomfort from extreme temperatures. As people's demands for high-quality living environments continue to increase, the development of intelligent environmental control systems that can non-invasively sense human body status, integrate multi-dimensional environmental information, and achieve personalized dynamic adjustment has become the core direction of building intelligence and healthy building development.

[0003] However, despite the progress made in smart air conditioning technology in recent years, existing solutions still face multiple technical bottlenecks in achieving truly personalized thermal comfort control, specifically in the following aspects: 1) The means of sensing vital signs are limited and highly intrusive. Current mainstream methods mainly rely on visual sensors such as cameras for personnel detection and posture recognition. This not only poses a risk of privacy leakage, but also results in significant performance degradation in low-light or occluded scenarios, limiting the applicable scenarios of the system.

[0004] 2) Insufficient depth of physiological parameter perception. Existing methods can only acquire surface information such as human posture and position, and cannot accurately monitor deep physiological parameters such as heart rate and respiratory rate in a non-contact manner, making it difficult to accurately quantify the true thermal comfort state of the human body.

[0005] 3) Limited personalization adaptability. Existing technologies are mainly based on preset comfort air conditioning parameters or general human metabolic rate models, lacking fine-grained perception and adaptive adjustment capabilities for individual physiological differences, and thus failing to achieve truly personalized temperature control.

[0006] 4) Incomplete environmental perception dimensions. Existing solutions mainly focus on basic environmental parameters such as temperature and humidity, without fully considering the comprehensive impact of microenvironmental factors such as air velocity and radiation temperature on human thermal comfort, resulting in insufficient temperature control accuracy.

[0007] 5) Weak intelligent decision-making capabilities. Existing methods mostly adopt simple rule-based control strategies, lacking intelligent decision-making mechanisms based on multi-source information fusion, and are unable to achieve dynamic optimization control based on real-time physiological state feedback. Summary of the Invention

[0008] The purpose of this invention is to provide a smart air conditioning personalized temperature control system and method based on UWB radar vital sign perception. By integrating non-contact human vital signs, spatial location and environmental parameters to construct a multimodal thermal comfort perception system, and combining deep learning and reinforcement learning mechanism based on user manual temperature adjustment behavior, the personalized temperature control strategy can be adaptively optimized, significantly improving the comfort experience while synergistically improving energy utilization efficiency.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A personalized temperature control method for smart air conditioners based on UWB radar signature perception includes the following steps: Step 1: Construct a personalized thermal comfort recognition model based on deep learning; The deep learning-based personalized thermal comfort recognition model includes: An improved UWB radar signal preprocessing and vital sign feature extraction module with temporal convolution and feature enhancement: This module is used to preprocess the raw UWB radar signal and extract high-quality vital sign features. It includes an adaptive filtering and baseline drift correction unit, a signal enhancement algorithm unit, an improved empirical mode decomposition (EMD) algorithm unit, a physiological feature temporal convolutional network with residual attention module, a point cloud feature aggregation algorithm unit, and a long short-term memory (LSTM) network with spatial attention gating mechanism. A multi-source information fusion and personalized temperature control model based on cross-modal attention and meta-learning: This model is used to achieve effective fusion of multi-source information and accurate identification of personalized thermal comfort, including improved cross-modal attention mechanisms, improved deep neural networks, multi-task learning frameworks, and meta-learning mechanisms. An adaptive temperature control strategy based on deep reinforcement learning and feedback optimization: This strategy enables autonomous learning and continuous optimization of temperature control policies. It includes a deep reinforcement learning temperature control decision network, a state space, an action space, an improved policy gradient algorithm, advantage function estimation, and a reward function. The state space... Includes environmental parameters, human vital signs, and historical temperature control records; the action space Includes temperature setting, fan speed adjustment, and mode adjustment; Fusion model: used to organically integrate various modules to form a complete end-to-end processing flow; And a reinforcement learning module: used for computational and optimization tasks related to reinforcement learning; Step 2: Deploy an ultra-wideband (UWB) radar sensor array and environmental monitoring sensors to collect multi-source human vital signs signals, location information, and environmental parameter data. Extract basic vital signs and basic human activity features from the human vital signs signals and location information using signal processing algorithms. Perform multi-dimensional annotation on the multi-source data to obtain annotation information. Combine prior physiological knowledge with the feature results extracted by the signal processing algorithms to associate and integrate human vital signs signals, location information, environmental parameters, and annotation information to construct a multimodal human thermal comfort perception dataset, and divide it into training, validation, and test sets. Step 3: Train a personalized thermal comfort recognition model based on deep learning; Step 3.1: UWB radar signal preprocessing and vital sign feature extraction module based on improved temporal convolution and feature enhancement: The original UWB radar echo signal is preprocessed through adaptive filtering and baseline drift correction to effectively suppress environmental noise, multipath interference, and DC components. A signal enhancement algorithm is used to improve the signal-to-noise ratio of vital sign signals. The improved Empirical Mode Decomposition (EMD) algorithm is used to separate respiratory, heartbeat, and body movement components, and extract temporal and frequency domain features from them. The extracted features are input into a physiological feature temporal convolutional network with a residual attention module to obtain a deep feature representation of vital signs. Based on the point cloud feature aggregation algorithm, human activity trajectory, activity amplitude, and frequency features are extracted and input into a Long Short-Term Memory (LSTM) network with a spatial attention gating mechanism to obtain a human activity feature representation and achieve activity state classification. At the same time, environmental parameters such as temperature, humidity, and wind speed are collected and preprocessed from environmental sensors in real time to construct an environmental parameter feature vector. Step 3.2: Multi-source information fusion and personalized temperature control model based on cross-modal attention and meta-learning: Vital signs, activity features, and environmental features are fused at the feature level. An improved cross-modal attention mechanism is introduced to calculate feature association weights, achieving adaptive weighted fusion to generate fused features. An improved deep neural network is constructed based on the fused features to simultaneously predict human thermal comfort level and recommended temperature setpoint under a multi-task learning framework. By introducing a meta-learning mechanism, the model can quickly adapt to different users, perform fine-tuning based on a small amount of individual data, and output personalized thermal comfort level and recommended temperature. Step 3.3 Adaptive Temperature Control Strategy Based on Deep Reinforcement Learning and Feedback Optimization: A deep reinforcement learning-based temperature control decision network with human thermal comfort as the core reward function is designed. A state space is defined, consisting of environmental parameters, vital signs, and historical control records, and includes an action space for temperature setting, wind speed adjustment, and mode switching. An improved policy gradient algorithm is used to optimize the temperature control strategy, and a dominant function estimation is introduced to reduce gradient variance. A feedback correction mechanism is established by analyzing user manual adjustment behavior to dynamically adjust the reward function, achieving continuous self-optimization of the temperature control strategy. Step 4: Use the validation set to evaluate the performance of the deep learning-based personalized thermal comfort recognition model and optimize the parameters; Step 5: Integrate the UWB radar sensor array and environmental monitoring sensors into the air conditioning operating environment. The central control unit interacts with the air conditioning system to exchange data and control commands. Deploy the trained, deep learning-based personalized thermal comfort recognition model into the temperature control system. Simultaneously, deploy an intelligent temperature control system based on edge computing devices to collect and process vital signs and environmental data in real time. Generate feature inputs according to the preprocessing and feature extraction process, and sequentially call the fusion model and reinforcement learning module to generate temperature control commands. At the same time, use the user's manual adjustment of the air conditioner as a reward signal for reinforcement learning, continuously updating the personalized temperature control model and control strategy to achieve iterative optimization of temperature control performance. This ensures thermal comfort while improving energy efficiency and adapting to individual usage habits.

[0010] Furthermore, step 3.1 includes the following process: Step 3.1.1: First, an adaptive noise cancellation algorithm is used to remove environmental noise and multipath interference. Then, a baseline drift correction algorithm is used to eliminate the DC component. The calculation formula is as follows: in This represents the radar signal after the DC component has been removed. Represents the original radar signal. This represents the baseline drift component estimated using the sliding window mean method; Next, a signal enhancement algorithm is used to improve the signal-to-noise ratio of the vital signs signal. By calculating the instantaneous amplitude envelope of the signal and adjusting the gain, the characteristic performance of respiratory and heartbeat signals is enhanced. The continuous radar signal is segmented using a sliding window segmentation technique. The window size is set according to the frequency characteristics of the vital signs signal to ensure that each window contains a complete physiological signal cycle. The segmented signal output after processing maintains the temporal continuity. Step 3.1.2: Input the preprocessed radar signal into the improved empirical mode decomposition algorithm to separate the respiratory signal component. Heartbeat signal components and body motion disturbance components The time-domain features of the separated respiratory and heartbeat signals are extracted, and the frequency-domain features are extracted by performing Fourier transform on the respiratory and heartbeat signals. The extracted time-domain and frequency-domain features are then input into a physiological feature temporal convolutional network to output a deep feature representation of vital signs. ; respiratory frequency component and heart rate frequency component These represent the frequency values ​​with the highest energy in the spectrum; the energy distribution function of the breathing band. Energy distribution function of heart rate band This is obtained by calculating the energy percentage of different frequency bands; respiratory rate The calculation formula is as follows: in Indicates respiratory rate, Indicates time window The number of respiratory cycles within the body; Heart rate The calculation formula is as follows: in Indicates heart rate, Indicates time window The number of heartbeat cycles within a day; respiratory variability coefficient The calculation formula is as follows: in Represents the respiratory variability coefficient. The standard deviation of the respiratory cycle. This represents the average value of the respiratory cycle; Coefficient of variation of heart rate The calculation formula is as follows: in This represents the coefficient of variation in heart rate. The standard deviation of the heartbeat cycle is represented by the standard deviation of the heartbeat cycle. This represents the average value of a heartbeat cycle; Step 3.1.3: Based on UWB radar point cloud data, extract the human motion trajectory using a three-dimensional coordinate matching algorithm. ,in , , They are respectively The coordinates of the human body in three-dimensional space at any given time; Calculate the range of activity The formula is as follows: in Indicates the activity amplitude at time t. , , They are respectively The three-dimensional coordinates of the human body at any given moment; Activity frequency The calculation formula is as follows: in Indicates activity frequency, Indicates time window Peak activity count within the period; The motion trajectory, amplitude, and frequency characteristics are input into the long short-term memory network of key human activity regions, which outputs a representation of human activity characteristics. An improved behavior recognition algorithm is used to classify and identify human activity states, dividing them into three categories: stillness, light activity, and moderate activity. Step 3.1.4: Real-time collection of environmental parameters such as temperature, humidity, and wind speed from deployed environmental monitoring sensors. The collected environmental parameter data is first filtered to remove noise using a moving average, then normalized to map parameters with different dimensions to the same numerical scale. The processed environmental parameters are then concatenated into an environmental parameter feature vector. .

[0011] Furthermore, step 3.2 includes the following process: Step 3.2.1: Represent the deep features of vital signs extracted in Step 3.1. Human activity characteristics representation Environmental parameter characteristics Perform feature-level fusion: An improved cross-modal attention mechanism is used to adaptively weight different feature dimensions. First, the correlation matrix between each feature is calculated. The formula is as follows: in This represents the association weight between the i-th feature and the j-th feature. Representation of features and Cosine similarity; generating feature weight vectors based on the association matrix. ,in , , The weighting coefficients for vital signs, human activity characteristics, and environmental parameters are respectively, satisfying the following conditions. The fusion features are obtained by weighted summation. The formula is as follows: in Indicates multimodal fusion features; Step 3.2.2: Construct a thermal comfort recognition model based on an improved deep neural network, using multimodal feature fusion. As input, feature mapping relationships are constructed through multiple fully connected layers and nonlinear activation functions (ReLU) to establish a mapping from vital signs to human thermal comfort state; The thermal comfort recognition model based on the improved deep neural network adopts a multi-task learning framework, simultaneously optimizing thermal comfort level prediction and recommended temperature setpoint prediction. The thermal comfort level prediction task outputs a probability distribution of human comfort status. ,in , , These represent the probabilities of thermal comfort, neutrality, and cold discomfort, respectively; the recommended temperature setpoint prediction task outputs a recommended temperature. The calculation formula is as follows: in This indicates the recommended temperature setting. This represents the weight matrix of the temperature prediction layer. This represents the output features of the hidden layer of the network. This represents the bias term of the temperature prediction layer; A meta-learning mechanism is introduced, which learns general initialization parameters by meta-training on multiple user datasets. When facing new users, only a small amount of labeled data is needed to quickly fine-tune the model to adapt to the thermal comfort characteristics of new users and output personalized thermal comfort levels and recommended temperatures.

[0012] Furthermore, step 3.3 includes: constructing a deep reinforcement learning temperature control decision network, wherein the network uses human thermal comfort as the core reward function, and the reward value... The calculation formula is as follows: in These correspond to bonus values ​​for different thermal comfort levels; An improved strategy gradient algorithm is used to optimize the temperature control strategy. The strategy update formula is as follows: in This indicates the updated policy parameters. Indicates the current policy parameters. Indicates the learning rate. Indicates the state Take action below The probability, Represents the dominance function; Establish a user feedback mechanism, and calculate the feedback correction coefficient by analyzing implicit user behavior signals and explicit setting adjustments. The formula is as follows: in This represents the feedback correction coefficient. This indicates that the user can manually adjust the number of times. This indicates the number of temperature control cycles during the total operating time, using... The reward function is dynamically adjusted to continuously optimize the temperature control strategy.

[0013] This invention also provides an intelligent air conditioning personalized temperature control system based on UWB radar signature perception, comprising: Human-Environment Status Sensing Module: This module includes an ultra-wideband (UWB) radar sensor array deployed in key monitoring areas of the air conditioning system and environmental monitoring sensors deployed in different areas of the room. The UWB radar sensor array is used for non-contact synchronous acquisition of raw echo signals containing vital signs and spatial location. Vital signs include respiratory rate, heart rate, and body motion frequency, while spatial location includes the real-time coordinates (x, y, z) of the human body in a three-dimensional coordinate system. The environmental monitoring sensors include a temperature sensor, a humidity sensor, and a wind speed sensor, which respectively collect environmental parameters such as ambient temperature, relative humidity, and air velocity. All sensors ensure the consistency of data acquisition time through a unified clock synchronization mechanism. Data acquisition and preprocessing module: used to receive and parse UWB radar signals and environmental sensor data; then execute preprocessing algorithms such as adaptive filtering and baseline drift correction to suppress noise; and initially extract basic time-domain and frequency-domain characteristics and basic human activity characteristics. The intelligent thermal comfort recognition module is used to identify the user's thermal comfort status in real time and generate a recommended temperature based on preprocessed multimodal data. Temperature control decision execution module: used to convert the decisions of the thermal comfort intelligent recognition module into specific air conditioning control commands; Feedback optimization module: used to improve the model and strategy online using actual operating data; the feedback optimization module continuously monitors two signals: one is the predicted state output by the thermal comfort intelligent recognition module, and the other is the user's manual adjustment operation of the air conditioner; the user's manual operation is regarded as a reward signal, and the strategy parameters of the temperature control decision network are dynamically updated by improving the strategy gradient algorithm, thereby realizing the iterative optimization of the temperature control strategy.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention uses a UWB radar sensor array to non-contactly collect human vital signs and location information. Combined with adaptive filtering and an improved empirical mode decomposition algorithm, it accurately extracts key features such as respiration and heart rate, providing a non-invasive and reliable physiological basis for thermal comfort recognition. Furthermore, by utilizing an improved cross-modal attention mechanism, the aforementioned vital signs are adaptively and weightedly fused with human activity characteristics and environmental parameters to construct a multimodal perception system that dynamically reflects the true thermal sensation of the human body, significantly improving the accuracy of state recognition. Based on this, by combining a meta-learning mechanism with deep reinforcement learning that uses user manual operation as feedback signals, the system can not only quickly adapt to new users with limited data but also continuously learn from long-term user behavior and iteratively optimize temperature control strategies. Ultimately, while achieving a highly personalized comfort experience, precise on-demand adjustment avoids energy waste, achieving a synergistic optimization of comfort and energy efficiency.

[0015] This invention employs multiple techniques during feature extraction and model training to ensure system stability across various scenarios. In the radar signal processing stage, multi-level vital signs are extracted using signal decomposition algorithms and deep learning models, effectively addressing individual differences and measurement noise. During model training, the introduction of a multi-task learning framework and meta-learning mechanism enables the model to adapt to different user characteristics and environmental conditions. Furthermore, by establishing a user feedback mechanism, the system can continuously optimize its temperature control strategy based on implicit behavioral signals and explicit settings, further enhancing its adaptability and robustness in practical deployments.

[0016] This invention utilizes UWB radar technology to achieve unobtrusive collection of vital signs such as respiration and heart rate, effectively solving the privacy leakage problems inherent in traditional visual sensors. Existing technologies rely on visual sensors such as cameras and thermal imagers, which have limited effectiveness at night or in obstructed environments and raise significant privacy concerns. UWB radar, however, can operate normally in complete darkness and with slight obstruction, achieving accurate measurement of human physiological parameters through radio frequency signal analysis. This not only protects user privacy but also expands the system's applicable scenarios. Furthermore, the deep learning-based vital sign feature extraction module can automatically learn effective feature representations from the raw radar signals, further improving the accuracy of vital sign monitoring.

[0017] This invention achieves precise temperature control based on individual physiological differences through multimodal feature fusion technology and a personalized thermal comfort recognition model. Existing technologies mainly rely on preset general comfort models or simple human posture recognition, which struggle to accurately reflect an individual's real-time thermal comfort needs. This invention, however, constructs a personalized thermal comfort recognition model by integrating vital sign characteristics, human activity characteristics, and environmental parameters, accurately quantifying the human body's thermal comfort state. The introduction of a meta-learning mechanism enables the model to quickly adapt to new users, significantly improving the system's applicability across different user groups.

[0018] This invention presents an adaptive temperature control strategy optimization mechanism based on deep reinforcement learning, which can dynamically adjust air conditioning operating parameters according to real-time physiological feedback. Existing technologies mostly employ simple rule-based control strategies, which cannot cope with complex and changing environmental conditions and individual needs. The reinforcement learning decision network constructed in this invention takes human thermal comfort as the optimization objective, comprehensively considering environmental parameters, human physiological characteristics, and historical temperature control records. Through a policy gradient algorithm, it continuously optimizes the temperature control strategy, achieving truly intelligent adaptive control. This human-machine collaborative optimization mechanism ensures that the temperature control system meets personalized comfort needs while also optimizing energy efficiency. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be described in detail below with reference to specific embodiments.

[0020] The intelligent air conditioning personalized temperature control method based on UWB radar sign perception described in this embodiment includes the following steps: Step 1: Construct a personalized thermal comfort recognition model based on deep learning The deep learning-based personalized thermal comfort recognition model includes a UWB radar signal preprocessing and vital sign feature extraction module with improved temporal convolution and feature enhancement, a multi-source information fusion and personalized temperature control model with cross-modal attention and meta-learning, an adaptive temperature control strategy with deep reinforcement learning and feedback optimization, a fusion model, and a reinforcement learning module.

[0021] The improved temporal convolution and feature enhancement UWB radar signal preprocessing and vital sign feature extraction module is responsible for preprocessing the original UWB radar signal and extracting high-quality vital sign features. It includes an adaptive filtering and baseline drift correction unit, a signal enhancement algorithm unit, an improved empirical mode decomposition (EMD) algorithm unit, a physiological feature temporal convolutional network with residual attention module, a point cloud feature aggregation algorithm unit, and a long short-term memory network (LSTM) with spatial attention gating mechanism.

[0022] The adaptive filtering and baseline drift correction unit effectively suppresses environmental noise and multipath interference by constructing a reference noise channel and dynamically adjusting the filter coefficients; it also uses the sliding window mean method to estimate and eliminate baseline drift components.

[0023] The signal enhancement algorithm unit improves the signal-to-noise ratio of vital sign-related components by calculating the instantaneous amplitude envelope of the signal and adjusting the gain.

[0024] The improved Empirical Mode Decomposition (EMD) algorithm unit introduces an adaptive noise-assisted analysis module on the basis of traditional Empirical Mode Decomposition. By adding Gaussian white noise of a specific intensity, it suppresses mode aliasing and separates the respiratory, heartbeat, and body motion components.

[0025] Physiological feature temporal convolutional networks with residual attention modules add residual attention modules to traditional temporal convolutional networks. By learning feature weights, the contribution of key physiological features is enhanced, and deep feature representations of vital signs are obtained.

[0026] The point cloud feature aggregation algorithm unit extracts human motion trajectory, activity amplitude, and frequency features based on UWB radar point cloud data through a three-dimensional coordinate matching algorithm.

[0027] The Long Short-Term Memory (LSTM) network with spatial attention gating introduces spatial attention gating mechanism on the basis of LSTM, dynamically adjusting the weights of features at different spatial locations to enhance the capture of features in key areas of human activity.

[0028] A multi-source information fusion and personalized temperature control model based on cross-modal attention and meta-learning is used to achieve effective fusion of multi-source information and accurate identification of personalized thermal comfort. This includes improving the cross-modal attention mechanism, improving deep neural networks, multi-task learning frameworks, and meta-learning mechanisms.

[0029] An improved cross-modal attention mechanism is used to calculate the correlation weights between vital signs, human activity features, and environmental parameters, enabling adaptive weighted fusion of different feature dimensions.

[0030] An improved deep neural network is constructed by taking multimodal fusion features as input and building feature mapping relationships through multiple fully connected layers and nonlinear activation functions.

[0031] The multi-task learning framework simultaneously optimizes two tasks: predicting thermal comfort levels and predicting recommended temperature setpoints.

[0032] The meta-learning mechanism enables the model to quickly adapt to different users.

[0033] The adaptive temperature control strategy based on deep reinforcement learning and feedback optimization enables autonomous learning and continuous optimization of the temperature control strategy. It includes a deep reinforcement learning temperature control decision network, state space, action space, improved policy gradient algorithm, advantage function estimation, and reward function.

[0034] The deep reinforcement learning-based temperature control decision network uses human thermal comfort as the core reward function to construct a dedicated temperature control decision network.

[0035] The state space contains a multidimensional representation of environmental parameters, human physical characteristics, and historical control records.

[0036] The action space includes a complete set of control commands for temperature setting, fan speed adjustment, and mode switching.

[0037] The improved policy gradient algorithm introduces an advantage function estimation module on the basis of the traditional policy gradient, thereby reducing the gradient estimation variance.

[0038] Advantage function estimation is used to accurately assess the relative value of actions and improve learning efficiency.

[0039] The reward function is designed with differentiated reward values ​​based on thermal comfort levels to guide the direction of strategy optimization.

[0040] The fusion model is responsible for organically integrating the various modules to form a complete end-to-end processing flow, which includes: Establish a complete data flow channel from signal preprocessing to feature extraction and multi-source information fusion; realize a collaborative working mechanism between deep learning models and reinforcement learning models; and construct a unified model architecture to ensure data compatibility and processing consistency among sub-modules.

[0041] The reinforcement learning module is responsible for the computational and optimization tasks related to reinforcement learning. Specifically, this includes: implementing forward propagation and gradient backpropagation of the policy network; managing the data storage and sampling of the experience replay buffer; performing policy evaluation and value function updates; and receiving and parsing user feedback signals, transforming them into reward or supervision signals that can be used for training. Step 2: Deploy an ultra-wideband (UWB) radar sensor array and environmental monitoring sensors to collect multi-source human vital signs signals, location information, and environmental parameter data. Use signal processing algorithms to extract basic vital sign features and basic human activity features from the human vital signs signals and location information, respectively. Perform multi-dimensional annotation on the multi-source data to obtain annotation information. Combine prior physiological knowledge with the feature results extracted by the signal processing algorithms to integrate the human vital signs signals, location information, environmental parameters, and annotation information to construct a multimodal human thermal comfort perception dataset, which is then divided into training, validation, and test sets. Tests were conducted in simulated laboratory environments and real-world residential scenarios (bedrooms and living rooms), covering typical conditions such as static rest, light activity (e.g., reading), and moderate activity (e.g., housework), as well as different environmental conditions including high temperature and humidity, and low temperature and dryness. Data acquisition was performed simultaneously using a UWB radar sensor array (sampling rate 120Hz) and multiple types of environmental sensors (temperature accuracy ±0.1℃, humidity accuracy ±2% RH, wind speed accuracy ±0.1m / s). The collected multi-source data underwent multi-dimensional annotation processing, with annotations including human comfort status (thermal comfort, neutral, cold discomfort), activity intensity level (rest, light activity, moderate activity), and category attributes of environmental parameters. By combining professional physiological monitoring equipment (respiratory rate accuracy ±1 breath / minute, heart rate accuracy ±1 breath / minute) with manual annotation, the true thermal comfort values ​​(thermal comfort, neutral, cold discomfort) of 50 test subjects (aged 20-60 years, half male and half female) were determined. A multimodal dataset containing 12,000 samples was constructed and divided into training set, validation set and test set in a 7:2:1 ratio.

[0042] Step 3: Train a deep learning-based personalized thermal comfort recognition model Step 3.1: UWB radar signal preprocessing and vital sign feature extraction module based on improved temporal convolution and feature enhancement: The original UWB radar echo signal is preprocessed through adaptive filtering and baseline drift correction to effectively suppress environmental noise, multipath interference, and DC components. A signal enhancement algorithm is used to improve the signal-to-noise ratio of vital sign signals. The improved Empirical Mode Decomposition (EMD) algorithm is used to separate respiratory, heartbeat, and body movement components, and extract temporal and frequency domain features from them. The extracted features are input into a physiological feature temporal convolutional network with a residual attention module to obtain a deep feature representation of vital signs. Based on the point cloud feature aggregation algorithm, human activity trajectory, activity amplitude, and frequency features are extracted and input into a Long Short-Term Memory (LSTM) network with a spatial attention gating mechanism to obtain a human activity feature representation and achieve activity state classification. At the same time, environmental parameters such as temperature, humidity, and wind speed are collected and preprocessed from environmental sensors in real time to construct an environmental parameter feature vector. Step 3.1.1, Radar signal preprocessing based on adaptive filtering: The raw echo signal acquired by the UWB radar is preprocessed. First, an adaptive noise cancellation algorithm is used to remove environmental noise and multipath interference. This algorithm minimizes the output noise energy by constructing a reference noise channel and dynamically adjusting the filter coefficients. Then, a baseline drift correction algorithm is used to eliminate the DC component. The calculation formula is as follows: in This represents the radar signal after the DC component has been removed. Represents the original radar signal. This represents the baseline drift component estimated using the sliding window mean method.

[0043] Next, a signal enhancement algorithm is used to improve the signal-to-noise ratio of the vital signs signals. By calculating the instantaneous amplitude envelope of the signal and adjusting the gain, the characteristic features of respiratory and heartbeat signals are enhanced. A sliding window segmentation technique is used to segment the continuous radar signal. The window size is set according to the frequency characteristics of the vital signs signals to ensure that each window contains a complete physiological signal cycle. The segmented signals output after processing maintain temporal continuity, providing a stable input for subsequent feature extraction.

[0044] Step 3.1.2: Vital Sign Feature Extraction Based on Improved Empirical Mode Decomposition (EMD): The preprocessed radar signal is input into the improved Empirical Mode Decomposition (EMD) algorithm. This algorithm introduces an adaptive noise-assisted analysis module on the basis of traditional EMD, which adds Gaussian white noise of a specific intensity to suppress mode aliasing and separate the respiratory signal components. Heartbeat signal components and body motion disturbance components .

[0045] Temporal features were extracted from the separated respiratory and heartbeat signals: respiratory rate The calculation formula is as follows: in Indicates respiratory rate, Indicates time window The number of respiratory cycles within the body; Heart rate The calculation formula is as follows: in Indicates heart rate, Indicates time window The number of heartbeat cycles within a day; respiratory variability coefficient The calculation formula is as follows: in Represents the respiratory variability coefficient. The standard deviation of the respiratory cycle. The mean value of the respiratory cycle; the coefficient of variation of heart rate. The calculation formula is as follows: in This represents the coefficient of variation in heart rate. The standard deviation of the heartbeat cycle is represented by the standard deviation of the heartbeat cycle. This represents the average value of the heartbeat cycle.

[0046] Fourier transform was performed on the respiratory and heartbeat signals to extract frequency domain features, including the dominant frequency component. (Breathing) and (Heartbeat) represents the frequency value with the highest energy in the spectrum; spectral energy distribution and The energy proportions of different frequency bands are calculated. The extracted temporal and frequency domain features are input into a physiological feature temporal convolutional network. This network adds a residual attention module to the traditional temporal convolutional network, which learns feature weights to enhance the contribution of key physiological features and outputs a deep feature representation of vital signs. .

[0047] Step 3.1.3: Human activity feature extraction based on point cloud feature aggregation: Based on UWB radar point cloud data, human motion trajectory is extracted using a three-dimensional coordinate matching algorithm. ,in , , They are respectively The coordinates of the human body in three-dimensional space at any given time. Calculate the range of motion. The formula is as follows: in Indicates the activity amplitude at time t. , , They are respectively The three-dimensional coordinates of the human body at any given time. Activity frequency. The calculation formula is as follows: in Indicates activity frequency, Indicates time window The peak number of activities within the region is determined. Motion trajectory, activity amplitude, and frequency features are input into a Long Short-Term Memory (LSTM) network for key human activity regions. This network introduces a spatial attention gating mechanism on top of LSTM to dynamically adjust the weights of features at different spatial locations, enhancing feature capture of key human activity regions and outputting a representation of human activity features. An improved behavior recognition algorithm is used to classify and identify human activity states. This algorithm combines activity feature representation. Based on preset activity intensity thresholds, human activity states are divided into three categories: rest, light activity, and moderate activity.

[0048] Step 3.1.4: Real-time collection of environmental parameters such as temperature, humidity, and wind speed from deployed environmental monitoring sensors. The collected environmental parameter data is first filtered to remove noise using a moving average, then normalized to map parameters with different dimensions to the same numerical scale. The processed environmental parameters are then concatenated into an environmental parameter feature vector. .

[0049] Step 3.2: Multi-source information fusion and personalized temperature control model based on cross-modal attention and meta-learning: Vital signs, activity, and environmental features are fused at the feature level. An improved cross-modal attention mechanism is introduced to calculate feature association weights, achieving adaptive weighted fusion to generate fused features. An improved deep neural network is constructed based on the fused features, simultaneously predicting human thermal comfort levels and recommended temperature settings within a multi-task learning framework. A meta-learning mechanism is introduced to enable rapid adaptation of the model to different users, fine-tuning based on a small amount of individual data, and outputting personalized thermal comfort levels and recommended temperatures. Step 3.2.1: Multimodal feature fusion based on weighted attention mechanism: representing deep features of vital signs Human activity characteristics representation Environmental parameter characteristics Feature-level fusion is performed on feature vectors (including ambient temperature, relative humidity, and air velocity). An improved cross-modal attention mechanism is used to adaptively weight different feature dimensions. First, the correlation matrix between each feature is calculated. The formula is as follows: in This represents the association weight between the i-th feature and the j-th feature. Representation of features and Cosine similarity. Feature weight vectors are generated based on the correlation matrix. ,in , , The weighting coefficients for vital signs, human activity characteristics, and environmental parameters are respectively, satisfying the following conditions. The fusion features are obtained by weighted summation. The formula is as follows: in This indicates a multimodal fusion feature, which highlights the contribution of key features that have a significant impact on thermal comfort perception.

[0050] Step 3.2.2 Personalized Thermal Comfort Recognition Model Based on Meta-Learning Enhancement: Construct a thermal comfort recognition model based on an improved deep neural network, which fuses features from multiple modalities. Using the input as input, a feature mapping relationship is constructed through multiple fully connected layers and a nonlinear activation function (ReLU) to establish a mapping from vital signs to the human body's thermal comfort state.

[0051] The model employs a multi-task learning framework, simultaneously optimizing two tasks: thermal comfort level prediction and recommended temperature setpoint prediction. The thermal comfort level prediction task outputs a probability distribution of human comfort levels. ,in , , These represent the probabilities of thermal comfort, neutrality, and cold discomfort, respectively; the recommended temperature setpoint is used to predict the output temperature for the task. The calculation formula is as follows: in This indicates the recommended temperature setting. This represents the weight matrix of the temperature prediction layer. This represents the output features of the hidden layer of the network. This represents the bias term of the temperature prediction layer.

[0052] A meta-learning mechanism (Model-Agnostic Meta-Learning, MAML) is introduced to enable the model to quickly adapt to different users. By performing meta-training on multiple user datasets, a common initialization parameter is learned. When facing a new user, only a small amount of labeled data is needed to quickly fine-tune the model to adapt to the thermal comfort characteristics of the new user and output a personalized thermal comfort level and recommended temperature.

[0053] Step 3.3: Adaptive Temperature Control Strategy Based on Deep Reinforcement Learning and Feedback Optimization: A deep reinforcement learning-based temperature control decision network with human thermal comfort as the core reward function is designed. A state space is defined, consisting of environmental parameters, vital signs, and historical control records, and an action space including temperature setting, wind speed adjustment, and mode switching. An improved policy gradient algorithm is used to optimize the temperature control strategy, and a dominance function estimation is introduced to reduce gradient variance. A feedback correction mechanism is established by analyzing user manual adjustment behavior to dynamically adjust the reward function, achieving continuous self-optimization of the temperature control strategy. Construct a Deep Reinforcement Learning Temperature Control Network (DRL-TCN) that uses human thermal comfort as the core reward function, with the reward value... The calculation formula is as follows: in These correspond to different reward values ​​for different thermal comfort levels. Since thermal comfort better meets the core needs of human beings for temperature control systems, the network is encouraged to output control strategies that put the human body in a thermal comfort state.

[0054] Design State Space Includes environmental parameters (ambient temperature) relative humidity air velocity ), human physical characteristics (respiratory rate) Heart rate Historical temperature control records (temperature setpoints over the past n moments) Wind speed adjustment value ; Action space Includes temperature setting ( The range is the adjustable temperature range of the air conditioner, and the fan speed adjustment ( It has multiple speed settings and mode switching (cooling, heating, and ventilation).

[0055] An improved policy gradient (IPG) algorithm is used to optimize the temperature control strategy. This algorithm introduces a dominance function estimation module on the basis of the traditional policy gradient algorithm, which reduces the gradient estimation variance. The policy update formula is as follows: in This indicates the updated policy parameters. Indicates the current policy parameters. Indicates the learning rate. Indicates the state Take action below The probability, This represents the dominance function.

[0056] Establish a user feedback mechanism by analyzing implicit user behavior signals (such as the frequency and amplitude of manual temperature adjustments) and explicit setting adjustments (such as directly setting a temperature value) to calculate a feedback correction coefficient. The formula is as follows: in This represents the feedback correction coefficient. This indicates that the user can manually adjust the number of times. This indicates the number of temperature control cycles during the total operating time, using... The reward function is dynamically adjusted to continuously optimize the temperature control strategy.

[0057] Step 4: Use the validation set to evaluate the performance of the deep learning-based personalized thermal comfort recognition model and optimize the parameters; Step 5: Integrate the UWB radar sensor array and environmental monitoring sensors into the air conditioning operating environment. The central control unit interacts with the air conditioning system to exchange data and control commands. Deploy the trained, deep learning-based personalized thermal comfort recognition model into the temperature control system. Simultaneously, deploy an intelligent temperature control system based on edge computing devices to collect and process vital signs and environmental data in real time. Generate feature inputs according to the preprocessing and feature extraction process, and sequentially call the fusion model and reinforcement learning module to generate temperature control commands. At the same time, use the user's manual adjustment of the air conditioner as a reward signal for reinforcement learning, continuously updating the personalized temperature control model and control strategy to achieve iterative optimization of temperature control performance. This ensures thermal comfort while improving energy efficiency and adapting to individual usage habits.

[0058] The intelligent air conditioning personalized temperature control system based on UWB radar sign perception described in this embodiment includes: The system includes a human-environment state perception module, a data acquisition and preprocessing module, a thermal comfort intelligent recognition module, a temperature control decision execution module, and a feedback optimization module. The human-environment state perception module includes an ultra-wideband (UWB) radar sensor array deployed in key air conditioning monitoring areas and environmental monitoring sensors deployed in different areas of the room. The UWB radar sensor array comprises multiple radar nodes, evenly distributed across the ceiling and surrounding walls of the room, used for non-contact synchronous acquisition of raw echo signals containing vital signs and spatial location. Vital signs include respiratory rate, heart rate, and body velocity; spatial location includes the real-time coordinates (x, y, z) of the human body in a three-dimensional coordinate system. The environmental monitoring sensors, including temperature, humidity, and wind speed sensors, collect environmental parameters such as ambient temperature, relative humidity, and airflow velocity. All sensors utilize a unified clock synchronization mechanism to ensure consistent data acquisition timing.

[0059] The data acquisition and preprocessing module is responsible for receiving raw data from the sensing module and performing preprocessing and integration. Its functions include: receiving and parsing UWB radar signals and environmental sensor data; then executing preprocessing algorithms such as adaptive filtering and baseline drift correction to suppress noise; and initially extracting basic time-domain and frequency-domain vital signs and basic human activity characteristics.

[0060] The intelligent thermal comfort recognition module is the core computing unit of the system. Its function is to identify the user's thermal comfort state in real time and generate a recommended temperature based on preprocessed multimodal data. This module first performs deep feature extraction on the preprocessed data through the vital sign feature extraction submodule to obtain a deep feature representation of vital signs (…). ) and representation of human activity characteristics ( Simultaneously, environmental parameters such as temperature, humidity, and wind speed are collected and preprocessed from environmental sensors in real time to construct an environmental parameter feature vector. Then, through the multi-source information fusion submodule, , With environmental characteristics ( The data is integrated and ultimately output as a thermal comfort level prediction and a personalized recommended temperature setpoint. ).

[0061] The temperature control decision execution module is responsible for translating the decisions of the thermal comfort intelligent recognition module into specific air conditioning control commands. Its function is to receive the recommended temperature setpoint (T_rec) output from the thermal comfort intelligent recognition module, and generate specific control signals accordingly to drive the air conditioning actuator to adjust the working mode, set temperature, and fan speed, thereby achieving automatic and adaptive adjustment of the indoor temperature.

[0062] The feedback optimization module is crucial for the system's continuous self-optimization, its role being to improve the model and strategy online using actual operational data. The core of this module is deep reinforcement learning and a feedback optimization mechanism. It continuously monitors two signals: the predicted state output by the intelligent thermal comfort recognition module and the user's manual adjustments to the air conditioner. It treats the user's manual actions as a reward signal and uses this to dynamically update the policy parameters of the temperature control decision network through an improved policy gradient algorithm, thereby achieving iterative optimization of the temperature control strategy.

[0063] Experimental section: 1. Multi-source data acquisition and preprocessing performance The performance of the dataset construction and preprocessing module was verified from three dimensions: data synchronization accuracy, signal preprocessing effect, and feature extraction accuracy.

[0064] Evaluation indicators: Time synchronization error (TSE): Time deviation of multi-source sensor data (unit: ms). The smaller the value, the better the synchronization effect. Signal-to-noise ratio improvement (SNR): The difference in signal-to-noise ratio between the vital signs before and after preprocessing (unit: dB). The larger the value, the better the denoising effect. Vital signs extraction error (FE): The relative error between the extracted respiratory rate and heart rate and the true value (unit: %). The smaller the value, the more accurate the extraction.

[0065] Table 1. Experimental Results of Multi-Source Data Acquisition and Preprocessing Performance Verification 2. Performance of Multimodal Fusion and Personalized Temperature Control Model This study compares the performance of our system with traditional temperature control models (such as PID control and PMV-based models) in terms of thermal comfort prediction accuracy and the rationality of recommended temperatures, while also verifying the personalized adaptation effect of meta-learning.

[0066] Evaluation indicators: Thermal comfort level prediction accuracy (Acc): The matching rate between the model's predicted thermal comfort state and the true value (unit: %). Recommended temperature deviation (T_dev): The difference between the model's output recommended temperature and the tester's preferred temperature (unit: ℃), the smaller the absolute value, the more reasonable the result. Personalization adaptation time (T_adap): The time (in minutes) required for the model to adapt to new users through meta-learning. The smaller the value, the faster the adaptation.

[0067] Table 2. Experimental Results of Performance Verification of Multimodal Fusion and Personalized Temperature Control Model 3. Adaptive temperature control strategy and overall system performance The study verified the comprehensive performance of the adaptive temperature control strategy and system deployment based on deep reinforcement learning and feedback optimization from three dimensions: temperature control response speed, thermal comfort maintenance effect, and energy efficiency optimization, and compared the performance differences under different activity intensities.

[0068] Evaluation indicators: Temperature control response time (T_res): The time (in minutes) for the ambient temperature to adjust from the initial value to the recommended temperature. The smaller the value, the faster the response. Thermal comfort maintenance rate (CR): The percentage of time the test subject is in a thermally comfortable state during system operation (unit: %). The higher the value, the better the effect. Energy efficiency improvement rate (E_save): The percentage reduction in system energy consumption compared to traditional air conditioning fixed mode (unit: %). The higher the value, the more energy-efficient the system.

[0069] Table 3. Experimental Results of Adaptive Temperature Control Strategy and Overall System Performance Verification Experimental results show that this system excels in multi-source data acquisition synchronization and the accuracy of vital sign feature extraction. Multimodal fusion and meta-learning mechanisms significantly improve the accuracy and adaptation speed of personalized temperature control, while the adaptive temperature control strategy achieves a dynamic balance between thermal comfort and energy efficiency. Overall performance surpasses traditional temperature control solutions, providing reliable technical support for intelligent temperature control in residential and office settings. Future development could further expand sensor coverage and enhance system adaptability in scenarios with multiple users simultaneously.

Claims

1. An intelligent air conditioner individual temperature control method based on UWB radar vital sign perception, characterized in that, Includes the following steps: Step 1: Construct a personalized thermal comfort recognition model based on deep learning; The deep learning-based personalized thermal comfort recognition model includes: An improved UWB radar signal preprocessing and vital sign feature extraction module with temporal convolution and feature enhancement: This module is used to preprocess the raw UWB radar signal and extract high-quality vital sign features. It includes an adaptive filtering and baseline drift correction unit, a signal enhancement algorithm unit, an improved empirical mode decomposition (EMD) algorithm unit, a physiological feature temporal convolutional network with residual attention module, a point cloud feature aggregation algorithm unit, and a long short-term memory (LSTM) network with spatial attention gating mechanism. A multi-source information fusion and personalized temperature control model based on cross-modal attention and meta-learning: This model is used to achieve effective fusion of multi-source information and accurate identification of personalized thermal comfort, including improved cross-modal attention mechanisms, improved deep neural networks, multi-task learning frameworks, and meta-learning mechanisms. Deep reinforcement learning and adaptive feedback-optimized temperature control strategy: autonomous learning and continuous optimization for temperature control strategy; including deep reinforcement learning temperature control decision network, state space, action space, improved policy gradient algorithm, advantage function estimation, and reward function; the state space including environmental parameters, human body sign features, and historical temperature control records; the action space including temperature setting, wind speed adjustment, and mode adjustment; Fusion model: used to organically integrate various modules to form a complete end-to-end processing flow; And a reinforcement learning module: used for computational and optimization tasks related to reinforcement learning; Step 2: Deploy an ultra-wideband (UWB) radar sensor array and environmental monitoring sensors to collect multi-source human vital signs signals, location information, and environmental parameter data. Extract basic vital signs and basic human activity features from the human vital signs signals and location information using signal processing algorithms. Perform multi-dimensional annotation on the multi-source data to obtain annotation information. Combine prior physiological knowledge with the feature results extracted by the signal processing algorithms to associate and integrate human vital signs signals, location information, environmental parameters, and annotation information to construct a multimodal human thermal comfort perception dataset, and divide it into training, validation, and test sets. Step 3: Train a personalized thermal comfort recognition model based on deep learning; Step 3.1: UWB radar signal preprocessing and vital sign feature extraction module based on improved temporal convolution and feature enhancement: The original UWB radar echo signal is preprocessed through adaptive filtering and baseline drift correction to effectively suppress environmental noise, multipath interference, and DC components. A signal enhancement algorithm is used to improve the signal-to-noise ratio of vital sign signals. The improved Empirical Mode Decomposition (EMD) algorithm is used to separate respiratory, heartbeat, and body movement components, and extract temporal and frequency domain features from them. The extracted features are input into a physiological feature temporal convolutional network with a residual attention module to obtain a deep feature representation of vital signs. Based on the point cloud feature aggregation algorithm, human activity trajectory, activity amplitude, and frequency features are extracted and input into a Long Short-Term Memory (LSTM) network with a spatial attention gating mechanism to obtain a human activity feature representation and achieve activity state classification. At the same time, environmental parameters such as temperature, humidity, and wind speed are collected and preprocessed from environmental sensors in real time to construct an environmental parameter feature vector. Step 3.2: Multi-source information fusion and personalized temperature control model based on cross-modal attention and meta-learning: Vital signs, activity features, and environmental features are fused at the feature level. An improved cross-modal attention mechanism is introduced to calculate feature association weights, achieving adaptive weighted fusion to generate fused features. An improved deep neural network is constructed based on the fused features to simultaneously predict human thermal comfort level and recommended temperature setpoint under a multi-task learning framework. By introducing a meta-learning mechanism, the model can quickly adapt to different users, perform fine-tuning based on a small amount of individual data, and output personalized thermal comfort level and recommended temperature. Step 3.3: Adaptive Temperature Control Strategy Based on Deep Reinforcement Learning and Feedback Optimization: Design a deep reinforcement learning temperature control decision network with human thermal comfort as the core reward function. Define a state space consisting of environmental parameters, vital signs, and historical control records, and include an action space for temperature setting, wind speed adjustment, and mode switching. Optimize the temperature control strategy using an improved strategy gradient algorithm, and introduce advantage function estimation to reduce gradient variance. Establish a feedback correction mechanism by analyzing user manual adjustment behavior, dynamically adjust the reward function, and achieve continuous self-optimization of the temperature control strategy. Step 4: Use the validation set to evaluate the performance of the deep learning-based personalized thermal comfort recognition model and optimize the parameters; Step 5: Integrate the UWB radar sensor array and environmental monitoring sensors into the air conditioning operating environment. The central control unit interacts with the air conditioning system to exchange data and control commands. Deploy the trained, deep learning-based personalized thermal comfort recognition model into the temperature control system. Simultaneously, deploy an intelligent temperature control system based on edge computing devices to collect and process vital signs and environmental data in real time. Generate feature inputs according to the preprocessing and feature extraction process, and sequentially call the fusion model and reinforcement learning module to generate temperature control commands. At the same time, use the user's manual adjustment of the air conditioner as a reward signal for reinforcement learning, continuously updating the personalized temperature control model and control strategy to achieve iterative optimization of temperature control performance. This ensures thermal comfort while improving energy efficiency and adapting to individual usage habits.

2. The intelligent air conditioner personalized temperature control method based on UWB radar vital sign sensing according to claim 1, characterized in that, The process of step 3.1 includes: Step 3.1.1: First, an adaptive noise cancellation algorithm is used to remove environmental noise and multipath interference. Then, a baseline drift correction algorithm is used to eliminate the DC component. The calculation formula is as follows: wherein denotes the radar signal after elimination of the direct current component, denotes the original radar signal, denotes the baseline drift component estimated by the moving window mean method; Next, a signal enhancement algorithm is used to improve the signal-to-noise ratio of the vital signs signal. By calculating the instantaneous amplitude envelope of the signal and adjusting the gain, the characteristic performance of respiratory and heartbeat signals is enhanced. The continuous radar signal is segmented using a sliding window segmentation technique. The window size is set according to the frequency characteristics of the vital signs signal to ensure that each window contains a complete physiological signal cycle. The segmented signal output after processing maintains the temporal continuity. Step 3.1.2, input the preprocessed radar signal into an improved empirical mode decomposition algorithm to separate out a respiratory signal component , a heartbeat signal component , and a body motion interference component ; extract time domain features from the separated respiratory signal and heartbeat signal, perform Fourier transform on the respiratory signal and the heartbeat signal to extract frequency domain features, input the extracted time domain and frequency domain features into a physiological feature time sequence convolution network to output vital sign deep feature representations ; respiratory frequency component and heart rate frequency component These represent the frequency values ​​with the highest energy in the spectrum; the energy distribution function of the breathing band. Energy distribution function of heart rate band This is obtained by calculating the energy percentage of different frequency bands; respiratory rate The calculation formula is as follows: in Indicates respiratory rate, Indicates time window The number of respiratory cycles within the body; Heart rate The calculation formula is as follows: in Indicates heart rate, Indicates time window The number of heartbeat cycles within a day; respiratory variability coefficient The calculation formula is as follows: in Represents the respiratory variability coefficient. The standard deviation of the respiratory cycle. This represents the average value of the respiratory cycle; Coefficient of variation of heart rate The calculation formula is as follows: in This represents the coefficient of variation in heart rate. The standard deviation of the heartbeat cycle is represented by the standard deviation of the heartbeat cycle. This represents the average value of a heartbeat cycle; Step 3.1.3: Based on UWB radar point cloud data, extract the human motion trajectory using a three-dimensional coordinate matching algorithm. ,in , , They are respectively The coordinates of the human body in three-dimensional space at any given time; Calculate the range of activity The formula is as follows: in This represents the activity level at time t. , , They are respectively The three-dimensional coordinates of the human body at any given moment; Activity frequency The calculation formula is as follows: in Indicates activity frequency, Indicates time window Peak activity count within the period; The motion trajectory, amplitude, and frequency characteristics are input into the long short-term memory network of key human activity regions, which outputs a representation of human activity characteristics. An improved behavior recognition algorithm is used to classify and identify human activity states, dividing them into three categories: stillness, light activity, and moderate activity. Step 3.1.4: Real-time collection of environmental parameters such as temperature, humidity, and wind speed from deployed environmental monitoring sensors. The collected environmental parameter data is first filtered to remove noise using a moving average, then normalized to map parameters with different dimensions to the same numerical scale. The processed environmental parameters are then concatenated into an environmental parameter feature vector. .

3. The intelligent air conditioning personalized temperature control method based on UWB radar characteristic perception as described in claim 2, characterized in that, The process in step 3.2 includes: Step 3.2.1: Represent the deep features of vital signs extracted in Step 3.

1. Human activity characteristics representation Environmental parameter characteristics Perform feature-level fusion: An improved cross-modal attention mechanism is used to adaptively weight different feature dimensions. First, the correlation matrix between each feature is calculated. The formula is as follows: in This represents the association weight between the i-th feature and the j-th feature. Representation of features and Cosine similarity; generating feature weight vectors based on the association matrix. ,in , , The weighting coefficients for vital signs, human activity characteristics, and environmental parameters are respectively, satisfying the following conditions. The fusion features are obtained by weighted summation. The formula is as follows: in Indicates multimodal fusion features; Step 3.2.2: Construct a thermal comfort recognition model based on an improved deep neural network, using multimodal feature fusion. As input, feature mapping relationships are constructed through multiple fully connected layers and nonlinear activation functions (ReLU) to establish a mapping from vital signs to human thermal comfort state; The thermal comfort recognition model based on the improved deep neural network adopts a multi-task learning framework, simultaneously optimizing thermal comfort level prediction and recommended temperature setpoint prediction. The thermal comfort level prediction task outputs a probability distribution of human comfort status. ,in , , These represent the probabilities of thermal comfort, neutrality, and cold discomfort, respectively; the recommended temperature setpoint prediction task outputs a recommended temperature. The calculation formula is as follows: in This indicates the recommended temperature setting. This represents the weight matrix of the temperature prediction layer. This represents the output features of the hidden layer of the network. This represents the bias term of the temperature prediction layer; A meta-learning mechanism is introduced, which learns general initialization parameters by meta-training on multiple user datasets. When facing new users, only a small amount of labeled data is needed to quickly fine-tune the model to adapt to the thermal comfort characteristics of new users and output personalized thermal comfort levels and recommended temperatures.

4. The intelligent air conditioning personalized temperature control method based on UWB radar characteristic perception as described in claim 3, characterized in that, Step 3.3 includes: constructing a deep reinforcement learning temperature control decision network, wherein the network uses human thermal comfort as the core reward function, and the reward value... The calculation formula is as follows: in These correspond to bonus values ​​for different thermal comfort levels; An improved strategy gradient algorithm is used to optimize the temperature control strategy. The strategy update formula is as follows: in This indicates the updated policy parameters. Indicates the current policy parameters. Indicates the learning rate. Indicates the state Take action below The probability, Represents the dominance function; Establish a user feedback mechanism, and calculate the feedback correction coefficient by analyzing implicit user behavior signals and explicit setting adjustments. The formula is as follows: in This represents the feedback correction coefficient. This indicates that the user can manually adjust the number of times. This indicates the number of temperature control cycles during the total operating time, using... The reward function is dynamically adjusted to continuously optimize the temperature control strategy.

5. A personalized temperature control system for intelligent air conditioning based on UWB radar signature perception, used to implement the personalized temperature control method for intelligent air conditioning based on UWB radar signature perception as described in any one of claims 1-4, characterized in that, include: Human-Environment Status Sensing Module: This module includes an ultra-wideband (UWB) radar sensor array deployed in key monitoring areas of the air conditioning system and environmental monitoring sensors deployed in different areas of the room. The UWB radar sensor array is used for non-contact synchronous acquisition of raw echo signals containing vital signs and spatial location. Vital signs include respiratory rate, heart rate, and body motion frequency, while spatial location includes the real-time coordinates (x, y, z) of the human body in a three-dimensional coordinate system. The environmental monitoring sensors include a temperature sensor, a humidity sensor, and a wind speed sensor, which respectively collect environmental parameters such as ambient temperature, relative humidity, and air velocity. All sensors ensure the consistency of data acquisition time through a unified clock synchronization mechanism. Data acquisition and preprocessing module: used to receive and parse UWB radar signals and environmental sensor data; then execute preprocessing algorithms such as adaptive filtering and baseline drift correction to suppress noise; and initially extract basic time-domain and frequency-domain vital signs and basic human activity characteristics. The intelligent thermal comfort recognition module is used to identify the user's thermal comfort status in real time and generate a recommended temperature based on preprocessed multimodal data. Temperature control decision execution module: used to convert the decisions of the thermal comfort intelligent recognition module into specific air conditioning control commands; Feedback optimization module: used to improve the model and strategy online using actual operating data; the feedback optimization module continuously monitors two signals: one is the predicted state output by the thermal comfort intelligent recognition module, and the other is the user's manual adjustment operation of the air conditioner; the user's manual operation is regarded as a reward signal, and the strategy parameters of the temperature control decision network are dynamically updated by improving the strategy gradient algorithm, thereby realizing the iterative optimization of the temperature control strategy.

Images