Extreme environment-oriented space adaptive control method

Abstract

The invention discloses an extreme environment-oriented space adaptive control method, which comprises the following steps: 1, multi-source heterogeneous data synchronous acquisition and feature extraction, specifically, deploying and integrating multiple types of sensing units, forming an asynchronous sensing network and acquiring an original data stream in real time: 2, data fusion and unified state vector construction, specifically, configuring an edge computing unit, and constructing a unified state vector; the method comprises the following steps: receiving an original data stream, performing space-time alignment, abnormal value detection and cleaning, and generating unified state description for decision making; 3, based on digital twinning multi-dimensional dynamic prediction, inputting the state vector into a plurality of prediction models running in parallel, and deducing core indexes in a future short period; 4, multi-target model predictive control MPC optimization solution is carried out; 5, cooperative distribution and safety verification of control instructions are carried out; and 6, performing closed-loop execution and multi-granularity feedback learning. According to the invention, upgrading from environment regulation and control to human factor guarantee is realized, dynamic balance under multi-target conflicts is satisfied, and predictability and self-evolution ability of the system are endowed.

Description

A Space Adaptive Control Method for Extreme Environments Technical Field

[0001] This invention belongs to the field of intelligent environmental control, computational perception and optimization decision-making, and particularly relates to a spatial adaptive control method for extreme environments. Background Technology

[0002] With the increasing prevalence of operations in extreme environments, the requirements for environmental control systems in enclosed spaces such as research cabins are becoming increasingly stringent. Current environmental control solutions in this field mainly suffer from the following problems:

[0003] (1) Single perception dimension and not integrated with the state of the personnel: Most existing systems rely on discrete environmental sensors such as temperature, humidity and CO2. Control decisions are based on preset fixed thresholds and lack non-invasive, real-time perception of the actual thermophysiological state of the personnel in the cabin, resulting in a disconnect between environmental regulation and the actual needs of the human body.

[0004] (2) Control objectives are isolated and lack collaborative optimization: Traditional control systems usually treat temperature control, humidity control, fresh air control, structural pressure control, etc. as independent "island-style" single-loop PID control, lacking a unified multi-objective collaborative optimization framework to balance objectives such as "safety", "comfort" and "energy saving".

[0005] (3) The decision-making model is static and lacks prediction and adaptive capabilities: Most existing control logics are based on static and empirical threshold regulation, lacking dynamic modeling and prediction capabilities. They cannot predict short-term (such as a snowstorm in the next hour) and long-term (such as material fatigue) risks, thus failing to achieve preventive regulation and self-optimization.

[0006] Therefore, there is an urgent need for an intelligent adaptive control system and method that can deeply integrate personnel and environmental state perception, establish dynamic prediction models, achieve real-time collaborative optimization of multiple objectives, and possess online learning and proactive fault tolerance capabilities, so as to fundamentally improve space security capabilities and operational efficiency in extreme environments. Summary of the Invention

[0007] The purpose of this invention is to provide a space adaptive control method for extreme environments, so as to solve the technical problems existing in the background art.

[0008] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0009] The present invention provides a space adaptive control method for extreme environments, comprising the following steps:

[0010] Step 1: Synchronous acquisition and feature extraction of multi-source heterogeneous data, specifically involving the deployment and integration of multiple types of sensing units to form an asynchronous sensing network for real-time acquisition of raw data streams:

[0011] Step 2: Data fusion and unified state vector construction, specifically, configuring edge computing units to receive the raw data stream from Step 1, performing spatiotemporal alignment, outlier detection and cleaning, and fusing to generate a unified state description for decision-making;

[0012] Step 3: Multi-dimensional dynamic prediction based on digital twins, specifically, inputting the state vector S_t into multiple prediction models running in parallel to extrapolate the core indicators in the near future;

[0013] Step 4: Multi-objective model predictive control MPC optimization solution, specifically, constructing and solving a finite-time rolling optimization problem;

[0014] Step 5: Coordinated allocation and security verification of control commands, specifically decomposing the abstract optimal control target U_t and converting it into executable commands for specific environmental control devices;

[0015] Step 6: Closed-loop execution and multi-granularity feedback learning, specifically the system executes control commands and continuously learns and optimizes based on the execution results.

[0016] Furthermore, step 1 specifically includes the following steps:

[0017] Step 1.1, Personnel Status Perception Flow: The outline of personnel in the space is captured by a visual sensor at a frequency of not less than 1Hz; a lightweight convolutional neural network is used to perform real-time segmentation and pose recognition of the image, and based on the body surface temperature distribution and pose information, the metabolic rate M, clothing thermal resistance I_cl and real-time coordinate position in the space of each individual are estimated by a pre-trained regression model.

[0018] The inputs for training the regression model include the average temperature of human body regions extracted from infrared images; the activity metabolic rate is determined by a pose estimation algorithm from the images; and environmental features: air temperature T_in and wind speed V_wind·in at the location of the person.

[0019] The regression model adopts the principle of human body thermal balance equation. Based on multiple recordings of the person's metabolic rate M and clothing thermal resistance I_cl, as well as the corresponding infrared images, posture and environmental data, it is obtained through a lightweight multilayer perceptron or feature fusion network.

[0020] Step 1.2, Space Environment Sensing Flow: Through distributed sensor nodes, synchronously collect environmental parameters inside the space, including air temperature T_in, relative humidity RH_in, carbon dioxide concentration CO2, particulate matter concentration, mean radiation temperature Tr_in, wind speed V_wind·in inside the space, and climate parameters outside the space, including wind speed V_wind·out, wind direction θ_wind, ambient temperature T_out, and solar irradiance G_solar.

[0021] Step 1.3, System Status Sensing Stream: Collect the operating status of each environmental control device and the status of the energy system.

[0022] Furthermore, step 2 specifically includes the following steps:

[0023] Step 2.1, Time Synchronization and Spatial Registration: Using a network time protocol-based mechanism, all sensor data are unified to the same time reference, personnel position coordinates are matched with the spatial 3D model, and associated with environmental sensor position information; for multiple measurement values ​​of the same type of parameter, a spatial distribution field is generated;

[0024] Step 2.2, State Vector Generation: Output the fused unified state vector S_t, whose structure is as follows:

[0025] S_t = {t, P_occ: [{pos, M, I_cl}, …], E_in: {T_in, RH_in, CO2, Tr_in, V_wind·in…}, E_out: {V_wind·out, θ_wind, T_out, …}, Sys: {SOC,device_states, …}};

[0026] Where t is the timestamp, P_occ is the personnel status set, pos is the real-time spatial coordinates of the personnel in the three-dimensional coordinate system inside the cabin (unit: m or grid ID), generated by visual positioning and spatial model matching; M is the metabolic rate (unit: met, 1 met = 58.2 W / m). 2 The thermal resistance (I_cl) reflects the intensity of human activity; a sitting temperature of approximately 1.0 met, while brisk walking can reach over 2.0 met. I_cl is the thermal resistance of clothing (unit: clo, 1 clo = 0.155 m). 2·K / W), representing the thermal insulation performance of clothing, polar clothing can reach over 2.0 clo, E_in represents the internal environmental state, E_out represents the external climate state, and Sys represents the system state. E_in represents the cabin environmental state, including: T_in is the air temperature (unit: °C), obtained by multi-point sensor fusion interpolation; RH_in is the relative humidity (dimensionless, %); CO2 is the carbon dioxide concentration (unit: ppm); PM2.5 is the fine particulate matter concentration (unit: μg / m³). 3 ); Tr_in is the average radiant temperature (unit: °C), reflecting the radiative heat transfer characteristics of the cabin walls and equipment; V_wind_in is the air velocity inside the cabin (unit: m / s).

[0027] E_out represents the external climate conditions, including: V_wind_out (external wind speed, m / s); θ_wind (wind direction, °, 0° is due north); T_out (external ambient temperature, °C); and G_solar (solar irradiance, W / m²). 2 P_atm is atmospheric pressure (unit: kPa).

[0028] Sys represents the system status, including: SOC, which is the state of charge of the energy storage unit (in %), representing the remaining battery capacity; E_renewable, which is the real-time power generation of renewable energy (in W); and device_states, which is a dictionary of the operating status of various environmental control devices, including at least the current mode, speed, valve opening, and real-time power consumption of devices such as HVAC systems, pressure regulation systems, and fresh air systems.

[0029] Furthermore, step 3 specifically includes the following steps:

[0030] Step 3.1 Thermal Comfort Prediction: Based on the ISO 7730 standard PMV-PPD (Predicted Average Votes - Predicted Percentage of Dissatisfaction) model, this step introduces predictions based on individual and environmental characteristics. Individual characteristics include metabolic rate M and clothing thermal resistance I_cl. Environmental characteristics include air temperature T_in, air velocity V_wind, relative humidity RH_in, and mean radiant temperature Tr_in. The predicted average votes (PMV) for the overall space and its zones at the current and future times are calculated, where -3 ≤ PMV ≤ +3. "-3" represents "very cold", "-2" represents "cold", "-1" represents "cool", "0" represents "neutral", "1" represents "warm", "2" represents "hot", and "3" represents "very hot". The PMV is then mapped to [0,1] to obtain the thermal comfort index C_comfort. The closer the value is to 1, the higher the comfort level.

[0031] Step 3.2, Envelope Risk Prediction: Establish a parameterized envelope response model based on a deep neural network. Its input vector includes: external wind speed and direction, external temperature, internal pressure setpoint P_set, and a set of parameterized features describing the current geometry of the cabin. The deep neural network is trained using a large amount of finite element simulation or measured data under different shapes and load conditions to learn the complex mapping from input conditions to stress and strain in key structural parts. Its output is the instability risk probability 0≤Risk_struct≤1, which is defined by the logistic regression output layer of the deep neural network and represents the possibility that the structural response exceeds the safety threshold under given input conditions.

[0032] Step 3.3, System Energy Consumption Prediction: Establish an equipment energy consumption model library; based on the current equipment status, setpoints, and environmental parameters, predict the energy consumption curves of the air conditioning, pressure regulation, and scientific research subsystems in the future prediction time domain and their impact on total energy consumption E_cost, where 0≤E_cost≤1, calculated as follows:

[0033] ;

[0034] Where Ptotal_(k) is the predicted total power consumption (W) for the k-th step; E_budget(k) is the energy budget for the k-th step, which is calculated as follows:

[0035] ;

[0036] E_budget(k) is the remaining battery energy at the current step plus the predicted total renewable energy generation up to this step.

[0037] Furthermore, step 4 specifically includes the following steps:

[0038] Step 4.1 Optimization Problem Modeling: Define the cost function in the prediction time domain N:

[0039] ;

[0040] Wherein, ω_c(k), ω_s(k), and ω_e(k) are the time-varying weight coefficients of the three optimization objectives of comfort, safety, and energy efficiency, respectively, satisfying ω_c(k)+ω_s(k)+ω_e(k)=1, and dynamically and smoothly adjusted according to the real-time state of the system to achieve adaptive and coordinated control of "safety first, comfort and energy saving dynamic trade-off" in extreme environments; C_comfort(k) is the thermal comfort index of the k-th step, normalized from PMV to [0,1], and the closer the value is to 1, the more comfortable the people are; struct(k) is the probability of instability risk of the building envelope in the k-th step, output by the deep neural network in the interval [0,1], and the closer it is to 0, the safer the structure is; E_cost(k) is the normalized energy consumption cost of the k-th step, calculated as the ratio of total power consumption to available energy budget, and the closer it is to 0, the lighter the energy burden is.

[0041] k is the step size in the prediction time domain, ω_c, ω_s, ω_e are time-varying weight coefficients, and ω_c+ω_s+ω_e=1;

[0042] Step 4.2, Dynamic Weight Adjustment: The weight coefficients are dynamically adjusted according to the system state, using a smooth transition mechanism;

[0043] ;

[0044] Where ω(k) is the weight value of a certain objective (comfort, safety, or energy efficiency) in the current control cycle (step k).

[0045] ω(k+1) is the updated value of this weight in the next control cycle, ω_target is the target weight dynamically determined by preset rules based on the real-time status of the system; α is the smoothing factor, α≤0.1, ω_target is determined according to the threshold rules of safety risk, SOC, and comfort deviation. When the predicted value of Risk_struct exceeds the warning threshold, ω_s is automatically increased, 0.3≤ω_s≤0.5; when the energy storage SOC is lower than the warning value, ω_e is increased, 0.2≤ω_e≤0.4; if PMV deviates from the comfort range C_comfort≤0.25 or C_comfort≥0.75 for more than 30 minutes, then ω_c is moderately increased, 0.1≤ω_c≤0.3.

[0046] Step 4.3, Real-time Solution: In each control cycle, the interior-point method is used to optimize the solver. Based on the current state S_t and the prediction model, an optimal sequence {U_t, U_{t+1}, ..., U_{t+N-1}} of future control actions is solved. The first control vector U_t in the sequence is output as the optimal control objective at the current moment. The form of U_t is:

[0047] U_t = {T_set, RH_set, P_set, ACH_set, (other device settings)};

[0048] T_set is the temperature; RH_set is the humidity; P_set is the pressure; ACH_set is the fresh air volume setting.

[0049] The constraints include: equipment physical limits, safe range of internal environmental parameters, personnel comfort boundaries, and energy system operation constraints.

[0050] Furthermore, step 5 specifically includes the following steps:

[0051] Step 5.1, Instruction Decomposition: The strategy allocator maps U_t to a device-level instruction set based on the characteristic model and control interface of each actuator; decomposes T_set and ACH_set into HVAC (Heating, Ventilation and Air Conditioning) system mode, fan speed, valve opening degree, and fresh air valve opening degree instructions; and converts P_set into fan speed and valve control signals for the pressure regulation system.

[0052] Step 5.2, Conflict Resolution and Smoothing: Inspect potential conflicts between different device commands, and resolve conflicts by setting priorities and limiting the rate of change; the priority setting is safety > energy efficiency > comfort;

[0053] Step 5.3, Security Boundary Verification: Before all instructions are issued, a final security verification is performed to ensure that they do not exceed the security boundaries of the device and system.

[0054] Furthermore, step 6 specifically includes the following steps:

[0055] Step 6.1, Effect Monitoring and Evaluation: In the next control cycle, collect a new state vector S_{t+1}, compare the actual state with the predicted state in step 3, and evaluate the control effect;

[0056] Step 6.2, Online learning of model parameters: Short-term adaptive modeling uses recursive least squares to fine-tune the parameters of the equipment characteristic model online based on the "set value - actual response" data; long-term learning involves periodically uploading the operating data to the cloud; offline training and updating of personalized parameters in the thermal comfort prediction model are performed using historical data, and then the updated model parameters are distributed to the edge side.

[0057] Step 6.3, Strategy Library Enhancement: Abstract successful control cases into strategy experience and store them in the case library to enrich the initial solutions to optimization problems.

[0058] The space adaptive control method for extreme environments proposed in this invention has the following advantages:

[0059] (1) Achieve an upgrade from "environmental regulation" to "human factor protection": By deeply integrating infrared vision and algorithms with the real physiological state of people, the control system can make decisions with "human comfort and health" as the core, and achieve personalized and precise environmental protection.

[0060] (2) Satisfying dynamic trade-offs under multi-objective conflict: A unified multi-objective optimization framework and model predictive control system have been established, which can solve the dynamic optimal solution of "comfort, safety and energy saving" online in real time and has good collaborative control capability in extreme environments.

[0061] (3) It endows the system with predictive and self-evolutionary capabilities: By using dynamic models and machine learning algorithms, the system has the ability to feedforward control and risk prediction, and can continuously learn to optimize its own parameters and control strategies, thus significantly enhancing its adaptability and robustness. Attached Figure Description

[0062] Figure 1 is a flowchart of a space adaptive control method for extreme environments according to the present invention. Detailed Implementation

[0063] To better understand the purpose, structure, and function of this invention, the following detailed description of a space adaptive control method for extreme environments, in conjunction with the accompanying drawings, is provided.

[0064] As shown in Figure 1, a space adaptive control method for extreme environments according to the present invention includes the following steps:

[0065] Step 1: Synchronous acquisition and feature extraction of multi-source heterogeneous data, specifically involving the deployment and integration of multiple types of sensing units to form an asynchronous sensing network for real-time acquisition of raw data streams:

[0066] Step 2: Data fusion and unified state vector construction, specifically, configuring edge computing units to receive the raw data stream from Step 1, performing spatiotemporal alignment, outlier detection and cleaning, and fusing to generate a unified state description for decision-making;

[0067] Step 3: Multi-dimensional dynamic prediction based on digital twins, specifically, inputting the state vector S_t into multiple prediction models running in parallel to extrapolate the core indicators in the near future;

[0068] Step 4: Multi-objective model predictive control MPC optimization solution, specifically, constructing and solving a finite-time rolling optimization problem;

[0069] Step 5: Coordinated allocation and security verification of control commands, specifically decomposing the abstract optimal control target U_t and converting it into executable commands for specific environmental control devices;

[0070] Step 6: Closed-loop execution and multi-granularity feedback learning, specifically the system executes control commands and continuously learns and optimizes based on the execution results.

[0071] Step 1 specifically includes the following steps:

[0072] Step 1.1, Personnel Status Perception Flow: The outline of personnel in the space is captured by a visual sensor at a frequency of not less than 1Hz; a lightweight convolutional neural network is used to perform real-time segmentation and pose recognition of the image, and based on the body surface temperature distribution and pose information, the metabolic rate M, clothing thermal resistance I_cl and real-time coordinate position in the space of each individual are estimated by a pre-trained regression model.

[0073] The inputs for training the regression model include the average temperature of human body regions extracted from infrared images; the activity metabolic rate is determined by a pose estimation algorithm from the images; and environmental features: air temperature T_in and wind speed V_wind·in at the location of the person.

[0074] The regression model adopts the principle of human body thermal balance equation. Based on multiple recordings of the person's metabolic rate M and clothing thermal resistance I_cl, as well as the corresponding infrared images, posture and environmental data, it is obtained through a lightweight multilayer perceptron or feature fusion network.

[0075] Step 1.2, Space Environment Sensing Flow: Through distributed sensor nodes, synchronously collect environmental parameters inside the space, including air temperature T_in, relative humidity RH_in, carbon dioxide concentration CO2, particulate matter concentration, mean radiation temperature Tr_in, wind speed V_wind·in inside the space, and climate parameters outside the space, including wind speed V_wind·out, wind direction θ_wind, ambient temperature T_out, and solar irradiance G_solar.

[0076] Step 1.3, System Status Sensing Stream: Collect the operating status of each environmental control device and the status of the energy system.

[0077] Step 2 specifically includes the following steps:

[0078] Step 2.1, Time Synchronization and Spatial Registration: Using a network time protocol-based mechanism, all sensor data are unified to the same time reference, personnel position coordinates are matched with the spatial 3D model, and associated with environmental sensor position information; for multiple measurement values ​​of the same type of parameter, a spatial distribution field is generated;

[0079] Step 2.2, State Vector Generation: Output the fused unified state vector S_t, whose structure is as follows:

[0080] S_t = {t, P_occ: [{pos, M, I_cl}, …], E_in: {T_in, RH_in, CO2, Tr_in, V_wind·in…}, E_out: {V_wind·out, θ_wind, T_out, …}, Sys: {SOC,device_states, …}};

[0081] Where t is the timestamp, P_occ is the personnel status set, pos is the real-time spatial coordinates of the personnel in the three-dimensional coordinate system inside the cabin (unit: m or grid ID), generated by visual positioning and spatial model matching; M is the metabolic rate (unit: met, 1 met = 58.2 W / m). 2 The thermal resistance (I_cl) reflects the intensity of human activity; a sitting temperature of approximately 1.0 met, while brisk walking can reach over 2.0 met. I_cl is the thermal resistance of clothing (unit: clo, 1 clo = 0.155 m). 2 ·K / W), representing the thermal insulation performance of clothing, polar clothing can reach over 2.0 clo, E_in represents the internal environmental state, E_out represents the external climate state, and Sys represents the system state. E_in represents the cabin environmental state, including: T_in is the air temperature (unit: °C), obtained by multi-point sensor fusion interpolation; RH_in is the relative humidity (dimensionless, %); CO2 is the carbon dioxide concentration (unit: ppm); PM2.5 is the fine particulate matter concentration (unit: μg / m³). 3 ); Tr_in is the average radiant temperature (unit: °C), reflecting the radiative heat transfer characteristics of the cabin walls and equipment; V_wind_in is the air velocity inside the cabin (unit: m / s).

[0082] E_out represents the external climate conditions, including: V_wind_out (external wind speed, m / s); θ_wind (wind direction, °, 0° is due north); T_out (external ambient temperature, °C); and G_solar (solar irradiance, W / m²). 2 P_atm is atmospheric pressure (unit: kPa).

[0083] Sys represents the system status, including: SOC, which is the state of charge of the energy storage unit (in %), representing the remaining battery capacity; E_renewable, which is the real-time power generation of renewable energy (in W); and device_states, which is a dictionary of the operating status of various environmental control devices, including at least the current mode, speed, valve opening, and real-time power consumption of devices such as HVAC systems, pressure regulation systems, and fresh air systems.

[0084] Step 3 specifically includes the following steps:

[0085] Step 3.1 Thermal Comfort Prediction: Based on the ISO 7730 standard PMV-PPD (Predicted Average Votes - Predicted Percentage of Dissatisfaction) model, this step introduces predictions based on individual and environmental characteristics. Individual characteristics include metabolic rate M and clothing thermal resistance I_cl. Environmental characteristics include air temperature T_in, air velocity V_wind, relative humidity RH_in, and mean radiant temperature Tr_in. The predicted average votes (PMV) for the overall space and its zones at the current and future times are calculated, where -3 ≤ PMV ≤ +3. "-3" represents "very cold", "-2" represents "cold", "-1" represents "cool", "0" represents "neutral", "1" represents "warm", "2" represents "hot", and "3" represents "very hot". The PMV is then mapped to [0,1] to obtain the thermal comfort index C_comfort. The closer the value is to 1, the higher the comfort level.

[0086] Step 3.2, Envelope Risk Prediction: Establish a parameterized envelope response model based on a deep neural network. Its input vector includes: external wind speed and direction, external temperature, internal pressure setpoint P_set, and a set of parameterized features describing the current geometry of the cabin. The deep neural network is trained using a large amount of finite element simulation or measured data under different shapes and load conditions to learn the complex mapping from input conditions to stress and strain in key structural parts. Its output is the instability risk probability 0≤Risk_struct≤1, which is defined by the logistic regression output layer of the deep neural network and represents the possibility that the structural response exceeds the safety threshold under given input conditions.

[0087] Step 3.3, System Energy Consumption Prediction: Establish an equipment energy consumption model library; based on the current equipment status, setpoints, and environmental parameters, predict the energy consumption curves of the air conditioning, pressure regulation, and scientific research subsystems in the future prediction time domain and their impact on total energy consumption E_cost, where 0≤E_cost≤1, calculated as follows:

[0088] ;

[0089] Where Ptotal_(k) is the predicted total power consumption (W) for the k-th step; E_budget(k) is the energy budget for the k-th step, which is calculated as follows:

[0090] ;

[0091] E_budget(k) is the remaining battery energy at the current step plus the predicted total renewable energy generation up to this step.

[0092] Step 4 specifically includes the following steps:

[0093] Step 4.1 Optimization Problem Modeling: Define the cost function in the prediction time domain N:

[0094] ;

[0095] Wherein, ω_c(k), ω_s(k), and ω_e(k) are the time-varying weight coefficients of the three optimization objectives of comfort, safety, and energy efficiency, respectively, satisfying ω_c(k)+ω_s(k)+ω_e(k)=1, and dynamically and smoothly adjusted according to the real-time state of the system to achieve adaptive and coordinated control of "safety first, comfort and energy saving dynamic trade-off" in extreme environments; C_comfort(k) is the thermal comfort index of the k-th step, normalized from PMV to [0,1], and the closer the value is to 1, the more comfortable the people are; struct(k) is the probability of instability risk of the building envelope in the k-th step, output by the deep neural network in the interval [0,1], and the closer it is to 0, the safer the structure is; E_cost(k) is the normalized energy consumption cost of the k-th step, calculated as the ratio of total power consumption to available energy budget, and the closer it is to 0, the lighter the energy burden is.

[0096] k is the step size in the prediction time domain, ω_c, ω_s, ω_e are time-varying weight coefficients, and ω_c+ω_s+ω_e=1;

[0097] Step 4.2, Dynamic Weight Adjustment: The weight coefficients are dynamically adjusted according to the system state, using a smooth transition mechanism;

[0098] ;

[0099] Where ω(k) is the weight value of a certain objective (comfort, safety, or energy efficiency) in the current control cycle (step k).

[0100] ω(k+1) is the updated value of this weight in the next control cycle, ω_target is the target weight dynamically determined by preset rules based on the real-time status of the system; α is the smoothing factor, α≤0.1, ω_target is determined according to the threshold rules of safety risk, SOC, and comfort deviation. When the predicted value of Risk_struct exceeds the warning threshold, ω_s is automatically increased, 0.3≤ω_s≤0.5; when the energy storage SOC is lower than the warning value, ω_e is increased, 0.2≤ω_e≤0.4; if PMV deviates from the comfort range C_comfort≤0.25 or C_comfort≥0.75 for more than 30 minutes, then ω_c is moderately increased, 0.1≤ω_c≤0.3.

[0101] 4.3 Real-time Solution: In each control cycle, an interior-point method is used to optimize the solver. Based on the current state S_t and the prediction model, an optimal sequence {U_t, U_{t+1}, ..., U_{t+N-1}} of future control actions is solved. The first control vector U_t in the sequence is output as the optimal control objective at the current moment. The form of U_t is:

[0102] U_t = {T_set, RH_set, P_set, ACH_set, (other device settings)};

[0103] T_set is the temperature; RH_set is the humidity; P_set is the pressure; ACH_set is the fresh air volume setting.

[0104] The constraints include: equipment physical limits, safe range of internal environmental parameters, personnel comfort boundaries, and energy system operation constraints.

[0105] Step 5 specifically includes the following steps:

[0106] Step 5.1, Instruction Decomposition: The strategy allocator maps U_t to a device-level instruction set based on the characteristic model and control interface of each actuator; decomposes T_set and ACH_set into HVAC (Heating, Ventilation and Air Conditioning) system mode, fan speed, valve opening degree, and fresh air valve opening degree instructions; and converts P_set into fan speed and valve control signals for the pressure regulation system.

[0107] Step 5.2, Conflict Resolution and Smoothing: Inspect potential conflicts between different device commands, and resolve conflicts by setting priorities and limiting the rate of change; the priority setting is safety > energy efficiency > comfort;

[0108] Step 5.3, Security Boundary Verification: Before all instructions are issued, a final security verification is performed to ensure that they do not exceed the security boundaries of the device and system.

[0109] Step 6 specifically includes the following steps:

[0110] Step 6.1, Effect Monitoring and Evaluation: In the next control cycle, collect a new state vector S_{t+1}, compare the actual state with the predicted state in step 3, and evaluate the control effect;

[0111] Step 6.2, Online learning of model parameters: Short-term adaptive modeling uses recursive least squares to fine-tune the parameters of the equipment characteristic model online based on the "set value - actual response" data; long-term learning involves periodically uploading the operating data to the cloud; offline training and updating of personalized parameters in the thermal comfort prediction model are performed using historical data, and then the updated model parameters are distributed to the edge side.

[0112] Step 6.3, Strategy Library Enhancement: Successful control cases are abstracted into strategic experiences and stored in the case library to enrich the initial solutions to the optimization problem. Example:

[0113] (1) Scene initialization:

[0114] Time: Polar night. Abrupt change in external climate: Wind speed V_wind·out increases to 18 m / s, wind direction is directly towards the entrance, temperature T_out drops sharply to -35°C. There are 2 people inside the cabin, initial internal temperature T_in is -5°C, and energy storage SOC is 70%.

[0115] (2) Step 1-2: Perception and Fusion

[0116] Infrared visual identification showed that person A (active) had a metabolic rate of 1.3 met and clothing thermal resistance of 1.6 clo; person B (sedentary) had a metabolic rate of 1.0 met and clothing thermal resistance of 2.0 clo; environmental sensors confirmed that T_in = -5°C, CO2 = 800 ppm, and external wind speed and temperature were as set in the scenario.

[0117] The edge computing unit generates a state vector S_t.

[0118] (3) Step 3: Dynamic prediction

[0119] Thermal comfort model prediction: The current overall PMV ≈ -1.8 (tending towards "cold"), PPD > 60%. Without intervention, the local PMV in the area where person B is located may be lower than -2.0.

[0120] Structural risk model prediction: Under the current internal and external pressure difference, the probability of instability in the windward negative pressure zone will rapidly rise to 45% within the next 10 minutes.

[0121] Energy consumption model prediction: If T_in is increased to a comfortable range (e.g., 15°C), energy consumption in the next hour will account for more than 50% of SOC.

[0122] (4) Step 4: MPC optimization solution

[0123] Because the Risk_struct prediction value is high, the optimizer automatically increases the safety weight ω_s.

[0124] After weighing the pros and cons of "increasing temperature to improve comfort" versus "increasing internal pressure to ensure safety" and "controlling energy consumption," the solver arrives at the current optimal solution U_t: {T_set: 0°C, P_set: +20% (relative to the reference pressure difference), ACH_set: 0.4 times / hour}

[0125] Decision logic: Prioritize significantly increasing internal pressure (P_set) to suppress the probability of structural risk to below 10%; on this basis, increase the temperature to 0°C to a limited extent to alleviate extreme cold, and strictly control the fresh air volume to save energy.

[0126] (5) Step 5: Control Allocation

[0127] Send a command to the pressure regulation system: "Within 90 seconds, steadily increase the overall internal pressure by 20%."

[0128] Send the instruction to the HVAC system: "Heating mode, target supply air temperature 0°C, low fan speed operation, avoid direct airflow to personnel area B."

[0129] Send a command to the fresh air system: "Maintain an air exchange rate of 0.4 times / hour".

[0130] A prompt was sent to the human-machine interface: "Strong wind and extreme cold warning. Structural pressure has increased. It is recommended to add warm clothing. Those sitting still can move to the vicinity of the activity area."

[0131] (6) Step 6: Closed Loop and Learning

[0132] Ten minutes later, monitoring showed that the pressure stabilized and the structural risk probability decreased to 8%; T_in rose to -1°C; and the estimated body surface temperature of person B slightly increased. The system recorded the entire chain of data for this coordinated control of "strong winds and extreme cold".

[0133] The comparison revealed that the actual temperature rise was slightly slower than the model prediction. The system made a fine-tuning of the performance coefficient model parameters of HVAC equipment at the current extremely low temperature to make future predictions more accurate. These were then marked as typical scenarios and stored in the strategy library.

[0134] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A space adaptive control method for extreme environments, characterized in that, The process includes the following steps: Step 1: Synchronous acquisition and feature extraction of multi-source heterogeneous data, specifically deploying and integrating multiple types of sensing units to form an asynchronous sensing network and acquiring raw data streams in real time; Step 2: Data fusion and unified state vector construction, specifically configuring edge computing units to receive the raw data stream from Step 1, performing spatiotemporal alignment, outlier detection and cleaning, and fusing to generate a unified state description for decision-making; Step 3: Multi-dimensional dynamic prediction based on digital twins, specifically inputting the state vector S_t into multiple prediction models running in parallel to extrapolate key indicators in the near future; Step 4: Multi-objective model predictive control (MPC) optimization solution, specifically constructing and solving a finite-time-domain rolling optimization problem; Step 5: Coordinated allocation and security verification of control commands, specifically decomposing the abstract optimal control objective U_t and converting it into executable commands for specific environmental control devices; Step 6: Closed-loop execution and multi-granularity feedback learning, specifically the system executes control commands and continuously learns and optimizes based on the execution results.

2. The space adaptive control method for extreme environments according to claim 1, characterized in that, Step 1 specifically includes the following steps: Step 1.1, Personnel Status Perception Flow: Capture the outline of personnel in space using a visual sensor at a frequency of no less than 1Hz; use a lightweight convolutional neural network to perform real-time image segmentation and pose recognition, and estimate the metabolic rate M, clothing thermal resistance I_cl, and real-time coordinate position in space for each individual based on body surface temperature distribution and pose information through a pre-trained regression model; the input to training the regression model includes the average temperature of human body regions extracted from infrared images; use a pose estimation algorithm from the images to determine the activity metabolic rate; environmental features: air temperature T_in and wind speed V_wind·in at the location of the personnel; the regression model adopts the principle of human body thermal balance equation, and after multiple recordings of the personnel's metabolic rate M and clothing... Based on the thermal resistance I_cl and the corresponding infrared image, attitude and environmental data, it is obtained through a lightweight multilayer sensor or feature fusion network; Step 1.2, Space Environment Sensing Flow: Through distributed sensor nodes, the environmental parameters inside the space are collected synchronously, including air temperature T_in, relative humidity RH_in, carbon dioxide concentration CO2, particulate matter concentration, mean radiation temperature Tr_in, wind speed V_wind·in inside the space, and climate parameters outside the space, including wind speed V_wind·out, wind direction θ_wind, ambient temperature T_out, and solar irradiance G_solar; Step 1.3, System Status Sensing Flow: Collect the operating status of each environmental regulation device and the status of the energy system.

3. The space adaptive control method for extreme environments according to claim 2, characterized in that, Step 2 specifically includes the following steps: Step 2.1, Time Synchronization and Spatial Registration: Using a network time protocol-based mechanism, all sensor data are unified to the same time reference, the personnel position coordinates are matched with the spatial 3D model, and associated with the environmental sensor position information; for multi-point measurements of the same type of parameter, a spatial distribution field is generated; Step 2.2, State Vector Generation: The fused unified state vector S_t is output, with the following structure: S_t = {t, P_occ: [{pos, M, I_cl}, …], E_in: {T_in, RH_in, CO2, Tr_in, V_wind·in…}, E_out: {V_wind·out, θ_wind, T_out, …}, Sys: {SOC, device_states, …}}; where t is the timestamp, P_occ is the personnel state set, pos is the real-time spatial coordinates of the personnel in the cabin 3D coordinate system; M is the metabolic rate; I_cl The system states are: E_in (clothing thermal resistance), E_out (cabin interior environment state), E_out (cabin exterior climate state), Sys (system state); T_in (air temperature); RH_in (relative humidity); Tr_in (mean radiant temperature); V_wind_in (cabin interior air velocity); V_wind_out (cabin exterior wind speed); θ_wind (wind direction); T_out (cabin exterior ambient temperature); G_solar (solar irradiance); P_atm (atmospheric pressure); SOC (state of charge of energy storage units); E_renewable (real-time renewable energy generation capacity); and device_states (dictionary of operating states of various environmental control devices).

4. The space adaptive control method for extreme environments according to claim 3, characterized in that, Step 3 specifically includes the following steps: Step 3.1, Thermal Comfort Prediction: Based on the ISO 7730 standard PMV-PPD model, introduces data based on individual personnel characteristics and environmental characteristics. Individual personnel characteristics are metabolic rate M and clothing thermal resistance I_cl, and environmental characteristics are air temperature T_in, air velocity V_wind, relative humidity RH_in, and mean radiant temperature Tr_in; calculate the overall and zoned predicted average PMV for the current and future times, -3≤PMV≤+3, where "-3" represents "very cold", and "-2" represents "very cold". "" represents "cold", "-1" represents "cool", "0" represents "neutral", "1" represents "warm", "2" represents "hot", and "3" represents "very hot". The PMV is mapped to [0,1] to obtain the thermal comfort index C_comfort. The closer the value is to 1, the higher the comfort level. Step 3.2: Envelope risk prediction: Establish a parameterized envelope response model based on a deep neural network. Its input vector includes: external wind speed and direction, external temperature, internal pressure setpoint P_set, and a set of parameterized features describing the current geometry of the cabin. The deep neural network uses a large number of finite elements under different shapes and load conditions. The system is trained using meta-simulation or measured data to learn the complex mapping from input conditions to stress and strain in key structural components. Its output is the probability of instability risk (0 ≤ Risk_struct ≤ 1), defined by the logistic regression output layer of a deep neural network, representing the likelihood that the structural response will exceed a safety threshold under given input conditions. Step 3.3: System energy consumption prediction: Establish an equipment energy consumption model library; based on the current equipment status, setpoints, and environmental parameters, predict the energy consumption curves of the air conditioning, pressure regulation, and research subsystems in the future prediction time domain and their impact on total energy consumption (E_cost), where 0 ≤ E_cost ≤ 1, calculated as follows: Where Ptotal_(k) is the predicted total power consumption (W) for the k-th step; E_budget(k) is the energy budget for the k-th step, which is calculated as follows: E_budget(k) is the remaining battery energy in the current step plus the predicted total renewable energy generation up to this step.

5. The space adaptive control method for extreme environments according to claim 4, characterized in that, Step 4 specifically includes the following steps: Step 4.1 Optimization problem modeling: Define the cost function in the prediction time domain N: Where ω_c(k), ω_s(k), and ω_e(k) are the time-varying weight coefficients of the three optimization objectives of comfort, safety, and energy efficiency, respectively, satisfying ω_c(k)+ω_s(k)+ω_e(k)=1; C_comfort(k) is the thermal comfort index of step k; struct(k) is the probability of instability risk of the building envelope in step k; E_cost(k) is the normalized energy consumption cost of step k; ω_c, ω_s, and ω_e are time-varying weight coefficients, and ω_c+ω_s+ω_e=1; Step 4.2, Dynamic Weight Adjustment: The weight coefficients are dynamically adjusted according to the system state, using a smooth transition mechanism; Where ω(k) is the weight value of a certain target in the current control cycle, ω(k+1) is the updated value of the weight in the next control cycle, ω_target is the target weight dynamically determined by preset rules according to the real-time state of the system; α is a smoothing factor. When the predicted value of Risk_struct exceeds the warning threshold, ω_s is automatically increased, 0.3≤ω_s≤0.5; when the energy storage SOC is lower than the warning value, ω_e is increased, 0.2≤ω_e≤0.4; if PMV deviates from the comfort range C_comfort≤0.25 or C_comfort≥0.75 for more than 30 minutes, then ω_c is moderately increased, 0.1≤ω_c≤0.3; Step 4.3, Real-time solution: In each control cycle, the interior point method is used to optimize the solver. Based on the current state S_t and the prediction model, the optimal sequence of a series of future control actions {U_t, U_{t+1}, ..., U_{t+N-1}} is solved. The first control vector U_t in the sequence is... The output is the optimal control target at the current moment; U_t is in the form of: U_t = {T_set, RH_set, P_set, ACH_set, (other device settings)}; T_set is temperature; RH_set is humidity; P_set is pressure; ACH_set is fresh air volume setting; constraints include: equipment physical limits, safe range of internal environmental parameters, personnel comfort boundaries, and energy system operation constraints.

6. The space adaptive control method for extreme environments according to claim 5, characterized in that, Step 5 specifically includes the following steps: Step 5.1, Instruction Decomposition: The strategy allocator maps U_t to a device-level instruction set based on the characteristic model and control interface of each actuator; decomposes T_set and ACH_set into HVAC system mode, fan speed, valve opening degree, and fresh air valve opening degree instructions; and converts P_set into fan speed and valve control signals for the pressure regulation system; Step 5.2, Conflict Resolution and Smoothing: Checks for potential conflicts between different device instructions, and resolves conflicts by setting priorities and limiting the rate of change; the priority is set as safety > energy saving > comfort; Step 5.3, Safety Boundary Verification: Before all instructions are issued, a final safety verification is performed to ensure that they do not exceed the safety boundaries of the equipment and system.

7. The space adaptive control method for extreme environments according to claim 6, characterized in that, Step 6 specifically includes the following steps: Step 6.1, Effect Monitoring and Evaluation: In the next control cycle, a new state vector S_{t+1} is collected, and the actual state is compared with the predicted state in Step 3 to evaluate the control effect; Step 6.2, Online Learning of Model Parameters: Short-term adaptive learning uses recursive least squares to fine-tune the parameters of the equipment characteristic model online based on the "setpoint-actual response" data; long-term learning involves periodically uploading operating data to the cloud; offline training and updating of personalized parameters in the thermal comfort prediction model are performed using historical data, and then the updated model parameters are distributed to the edge side; Step 6.3, Strategy Library Enhancement: Successful control cases are abstracted into strategy experience and stored in the case library to enrich the initial solution of the optimization problem.

Images