Intelligent temperature control system for cockpit of flight simulator

Abstract

The present application belongs to the technical field of flight simulation training equipment, and particularly relates to a flight simulator cockpit intelligent temperature control system, which solves the problem that the prior art cannot meet the requirements of multi-dimensional physiological perception fusion, multi-target collaborative optimization, and rapid personalized adaptation and active safety protection of the flight simulator temperature control. The multi-modal perception layer processes multi-channel skin physiological data, physical environment data and pilot electrocardio signals, constructs a blood flow corrected human body heat balance equation, and generates a physiological feature vector containing predicted core body temperature, predicted average vote and predicted dissatisfaction percentage. The decision layer generates temperature control instructions and confidence evaluation values through a meta-learning model, and the reward function is a four-dimensional dynamic weighted form, and the weight is adjusted according to the state through an IF-THEN rule. The dynamic execution layer adopts PID closed-loop regulation to divide the temperature and trigger the response in stages. The present application realizes precise and personalized temperature control, and guarantees the training quality and safety.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of flight simulation training equipment, and particularly relates to an intelligent temperature control system for a seat cabin of a flight simulator. BACKGROUND

[0002] As a key equipment for training pilots' core skills, the seat cabin of a flight simulator plays a decisive role in long-time and high-intensity simulation training, which directly affects the physiological stability, cognitive load level and operation performance of pilots. Therefore, constructing an efficient and comfortable temperature control system that is suitable for flight training scenarios has become a core requirement for ensuring training quality and the physiological safety of pilots.

[0003] In the prior art, the mainstream scheme (such as Chinese patent CN119674344A) discloses a technology that generally adopts a rough control logic based on a fixed threshold. Such a system relies on the monitoring data of a single or small number of environmental sensors (such as temperature and humidity sensors) in the seat cabin, and triggers the start and stop of the air conditioning system when the value exceeds the preset threshold. The core defect of this system lies in the limitation of the decision-making dimension: it completely ignores the physiological differences of individual pilots and the dynamic physiological changes under different training tasks, and cannot achieve the precise goal of "on-demand regulation". In actual application, problems such as response lag and energy waste caused by over-regulation often occur.

[0004] In addition, although some vehicle-mounted intelligent temperature control schemes (such as CN119396221A) introduce intelligent algorithms such as reinforcement learning, due to the essential differences between the application scenarios (daily driving vs. high-intensity flight training), the model architecture and control strategy of these schemes are difficult to be directly migrated. At the same time, these schemes have three inherent defects:

[0005] Reward function design is disconnected from actual needs: the reward function of reinforcement learning is mostly based on energy consumption E and temperature fluctuation as a single optimization target (such as ), without incorporating the PMV / PPD (Predicted Mean Vote / Predicted Percentage of Dissatisfied) parameter, which is the "gold standard" of thermal comfort, into the optimization system, resulting in a significant deviation between the system regulation target and the actual thermal sensation of people;

[0006] Insufficient individualized adaptive ability: when facing new users (new pilots), traditional reinforcement learning models need to be trained from scratch or undergo large-scale data retraining, and the cold start period is more than 2 hours, which cannot meet the scene requirements of rapid rotation of personnel in flight training;

[0007] Lack of safety protection mechanism: existing systems are mostly passive response control, lacking predictive assessment and hierarchical intervention mechanisms for the physiological limit state of pilots (such as excessively high core body temperature), and there is a potential health risk in high-intensity training scenarios;

[0008] In summary, the current technology cannot meet the needs of flight simulators for multi-dimensional physiological perception fusion, multi-target collaborative optimization and rapid individual adaptation, and an intelligent temperature control system for flight training scenarios is urgently needed. SUMMARY

[0009] To solve the above problems in the prior art, i.e., the prior art cannot meet the needs of flight simulator temperature control for multi-dimensional physiological perception fusion, multi-target collaborative optimization, and rapid individual adaptation and active safety protection of new pilots, the first aspect of the present application proposes an intelligent temperature control system for a flight simulator cockpit, which comprises:

[0010] A multi-modal perception layer configured to: obtain multi-channel skin physiological data and perform mean processing to obtain average skin physiological data; obtain physical environment data and calculate heat dissipation in combination with the average skin physiological data; obtain a pilot's electrocardiogram signal, pre-process and extract features to obtain a feature parameter; normalize the feature parameter to obtain a blood flow correction coefficient;

[0011] A human body heat balance equation based on the heat dissipation and the blood flow correction coefficient is constructed to generate a physiological feature vector; the physiological feature vector includes a predicted core body temperature, a predicted average vote, and a predicted dissatisfaction percentage;

[0012] A decision layer configured to: based on the physiological feature vector and a training task load instruction, generate a temperature control instruction and a confidence evaluation value using a meta-learning model; wherein the reward function of the meta-learning model is a multi-target dynamic weighted reward function, including four-dimensional targets of comfort, energy consumption, stability, and response speed; the weights of the reward function are dynamically adjusted according to real-time states through IF-THEN rules; the real-time states include: predicted core body temperature, predicted dissatisfaction percentage, training task load level, and safety margin;

[0013] A dynamic execution layer configured to: based on the temperature control instruction and the average skin physiological data, adjust the operating parameters of the airflow regulation equipment of each temperature control partition of the flight simulator cockpit using a PID closed loop, and then regulate the temperature of each partition;

[0014] Based on the real-time numerical interval of the predicted core body temperature and the predicted dissatisfaction percentage, a progressive temperature control response is triggered in stages, including local adjustment, compound adjustment, and global intervention alarm.

[0015] In some preferred embodiments, the pilot's electrocardiogram signal is obtained, pre-processed, and features are extracted to obtain a feature parameter; the feature parameter is normalized to obtain a blood flow correction coefficient, and the method is:

[0016] Obtain the pilot's electrocardiogram signal;

[0017] preprocessing the electrocardiogram signal, the preprocessing comprising detecting R peaks of the electrocardiogram signal using a Pan-Tompkins algorithm and generating a sequence of RR intervals, removing outliers in the sequence of RR intervals and interpolating the sequence of RR intervals to obtain a preprocessed electrocardiogram signal;

[0018] extracting features from the preprocessed electrocardiogram signal by short-time Fourier transform to obtain characteristic parameters, including low-frequency power, high-frequency power and LF / HF ratio;

[0019] obtaining a blood flow correction coefficient based on the extracted LF / HF ratio by a normalized sigmoid function ;

[0020] ;

[0021] wherein, is the blood flow correction coefficient at time t, is the minimum correction coefficient, is an adjustment parameter, is the real-time LF / HF ratio at time t, is the LF / HF threshold value.

[0022] In some preferred embodiments, the human heat balance equation corrected by the blood flow correction coefficient is:

[0023] ;

[0024] wherein, is the metabolic heat production, is the convective heat loss, is the radiative heat loss, is the evaporative heat loss, is the blood flow correction coefficient, is the heat capacity of the human body, is the core body temperature of the human body, is the rate of change of the core body temperature of the human body with respect to time.

[0025] In some preferred embodiments, a blood flow corrected human heat balance equation is constructed based on the heat loss and the blood flow correction coefficient, and a physiological feature vector is generated, and the method is:

[0026] integrating the human heat balance equation to obtain a predicted core body temperature;

[0027] obtaining a predicted average vote and a predicted percentage of dissatisfaction based on the predicted core body temperature, the physical environment data and the average skin physiological data by a thermal comfort model.

[0028] In some preferred embodiments, the physical environment data comprises cabin temperature, humidity, air flow speed; and the heat dissipation amount comprises convective heat dissipation amount, radiative heat dissipation amount, evaporative heat dissipation amount.

[0029] The multi-channel skin physiological data comprises 12-channel skin temperature and 12-channel skin humidity; and the average skin physiological data comprises average skin temperature and average skin humidity.

[0030] In some preferred embodiments, the meta-learning model acquisition method comprises:

[0031] defining a meta-training data set D_meta-train, wherein the D_meta-train comprises N different pilots' tasks, each task being defined as a time series data sequence {(S, A, R, S')} from a specific pilot in a training; S is a state vector, A is an action vector, R is a reward value, and S' is a new state vector after executing the action vector;

[0032] constructing a meta-learning model with a teacher network and a student network structure, migrating the knowledge of the teacher network to the student network and training the student network based on the meta-training data set through knowledge distillation to obtain the meta-trained student network parameter θ, wherein the teacher network is trained using historical data, and the loss function comprises a physical constraint loss; the loss function of the student network comprises a KL divergence and a hard label loss;

[0033] decoupling the student network parameter θ into general layer parameters, task layer parameters and individual layer parameters through hierarchical parameter decoupling;

[0034] when a new pilot joins, collecting the time series data of the new pilot for a preset time length as a fine-tuning data set D_finetune;

[0035] loading the meta-trained student network parameter θ, and performing limited gradient descent update using the D_finetune as input to obtain an individualized student network parameter θ';

[0036] using the student network updated by the parameter θ' to obtain a temperature control instruction and a confidence evaluation value.

[0037] In some preferred embodiments, the limited gradient descent update adopts dynamic parameter freezing and selective fine-tuning.

[0038] The dynamic parameter freezing and selective fine-tuning comprises: calculating a parameter freezing ratio according to the similarity of the student network parameters of the new pilot and the historical pilots, combining the parameter freezing ratio with the training time length to select parameters of different dimensions to be unfrozen in stages, performing limited gradient descent update, and then obtaining individualized model parameters.

[0039] In some preferred embodiments, the loaded student network parameters θ after training are updated by limited gradient descent with the D_finetune as input to obtain personalized student network parameters θ', and further comprising feature space alignment and migration optimization update: the feature distribution difference between the new and old pilots is minimized by maximum mean difference, and the personalized layer feature weight is dynamically adjusted by a progressive fusion function, and then the parameter gradient of the limited gradient descent update is corrected.

[0040] In some preferred embodiments, the parameter freezing ratio calculation formula is:

[0041] ;

[0042] wherein, is the parameter freezing ratio, is a natural exponential function for mapping similarity to a freezing ratio conforming to exponential decay law, is a similarity calculation function, representing the similarity measure of the initial model parameters of the new pilot and the cluster centers of the historical pilot model parameters, is a freezing adjustment parameter for controlling the influence degree of similarity on the freezing ratio;

[0043] The parameter freezing ratio is combined with the training duration to select the unfreezing parameters of different dimensions in stages, including: in the initial stage, only the personalized layer parameters are unfreezed; in the adaptive stage, the personalized layer parameters are unfreezed and part of the task layer parameters are unfreezed; in the stable stage, all parameters are unfreezed and elastic weight regularization is used to constrain the key parameter update.

[0044] In some preferred embodiments, the IF-THEN rule includes: safety critical rule, energy efficiency priority rule, and steady state optimization rule.

[0045] The beneficial effects of the present application are:

[0046] 1) In view of the limitations of traditional single environmental sensing, the sensing dimension is expanded to "physiology + environment + hemodynamics" three-dimension by 12-channel skin sensing and blood flow correction fusion. Combined with HRV features and heat balance equation, the core body temperature prediction error is reduced to ±0.2℃, and the influence of peripheral vasoconstriction on heat dissipation is accurately captured, providing comprehensive physiological-environmental data support for regulation;

[0047] ​2) The PMV / PPD thermal comfort parameters are incorporated into the reward function core, and the weight is dynamically adjusted through the "IF-THEN" rule: the comfort weight is increased by 50% in the safety critical period, the energy consumption weight is increased by 30% in the high energy consumption period, the PMV control accuracy is increased by 30%, the comprehensive energy consumption is reduced by 28%, the scene balance of comfort and energy saving is realized, and the target and feeling are avoided.

[0048] 3) The hierarchical parameter decoupling and meta-learning adaptation are adopted, 70% of the redundant training is reduced through dynamic parameter freezing, the new pilot model convergence time is reduced from 2 hours to 10 minutes based on 10-minute fine-tuning data and feature distribution alignment, and the rapid rotation demand of flight training "replace and adapt" is met;

[0049] 4) A three-level response mechanism is constructed: the core temperature is greater than 37.3 DEG C to start local air supply, PMV is greater than 1.0 to add cooperative cooling, PMV is greater than 2.0 to alarm globally, the thermal discomfort response time is reduced from 30 seconds to 8 seconds, the thermal stress risk is reduced by more than 60%, the "prediction-intervention-alarm" active safety closed loop is formed, and the physiological safety is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0050] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the accompanying drawings:

[0051] Figure 1 is the flow chart of the human body heat balance equation of the blood flow correction of the application.

[0052] Figure 2 is the architecture diagram of the intelligent temperature control system of the flight simulator cockpit of the application. DETAILED DESCRIPTION

[0053] The application will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related application, and not to limit the application. In addition, it should be noted that only the parts related to the application are shown in the drawings for ease of description.

[0054] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and embodiments.

[0055] In order to more clearly describe the intelligent temperature control system of the flight simulator cockpit of the application, the following will be combined with Figures 1-2 The steps in the embodiments of the application will be described in detail.

[0056] The application proposes an intelligent temperature control system for a flight simulator cockpit, as shown in Figure 1The system comprises a multi-modal perception layer, see Figure 2 , configured to acquire multi-channel skin physiological data and average processing to obtain average skin physiological data;

[0057] In this embodiment, the multi-channel skin physiological data includes 12 channels of skin temperature and 12 channels of skin humidity (12-channel flexible sensor array is arranged in the key areas of the pilot's seat and clothing trunk, limbs, etc. to collect skin temperature and humidity data in real time at a sampling rate of 100 Hz); and the average skin physiological data obtained by average processing includes skin average temperature , skin average humidity , providing physiological basis for differentiated temperature control;

[0058] Acquire physical environment data, and calculate heat dissipation amount in combination with the average skin physiological data;

[0059] In this embodiment, the physical environment data includes cabin temperature , humidity , air flow speed ; the heat dissipation amount realizes precise quantification of heat dissipation state, including convective heat dissipation amount , radiative heat dissipation amount , evaporative heat dissipation amount ; wherein the calculation formula is:

[0060] , wherein, is the convective heat transfer coefficient, A is the body surface area; the calculation formula is: , wherein, is the emissivity, is the Boltzmann constant;

[0061] The calculation formula is:

[0062] ;

[0063] , wherein, is the evaporative heat transfer coefficient, , are the skin surface and cabin water vapor partial pressure, respectively;

[0064] Acquire the pilot's electrocardio signal, pre-process and extract features to obtain feature parameters; normalize the feature parameters to obtain a blood flow correction coefficient;

[0065] In the embodiment, the pilot's electrocardiogram signal is obtained, preprocessed, and feature extraction is performed to obtain feature parameters; the feature parameters are normalized to obtain a blood flow correction coefficient, and the method is as follows:

[0066] The pilot's electrocardiogram signal is obtained, which can be collected by an ECG / PPG sensor;

[0067] The electrocardiogram signal is preprocessed, and the preprocessing includes detecting the R peak of the electrocardiogram signal by using the Pan-Tompkins algorithm and generating an RR interval sequence, and removing and interpolating outliers in the RR interval sequence to obtain a preprocessed electrocardiogram signal;

[0068] Feature extraction is performed on the preprocessed electrocardiogram signal by short-time Fourier transform (STFT) to obtain feature parameters , including low-frequency power , high-frequency power , and LF / HF ratio ; the expression of the feature parameters is:

[0069] ;

[0070] The dynamic increase of the LF / HF ratio is regarded as a core physiological parameter of peripheral vasoconstriction and decreased heat dissipation capacity;

[0071] Based on the extracted LF / HF ratio, a blood flow correction coefficient is obtained by a normalized S-shaped function, which captures the peripheral vasoconstriction effect caused by sympathetic nerve excitation and solves the defect of ignoring the blood flow influence of the traditional heat balance model;

[0072] ;

[0073] wherein, is the blood flow correction coefficient at time t, preferably [0.6, 1.0], is the minimum correction coefficient, preferably 0.6, is the adjustment parameter, preferably 5.0, is the real-time LF / HF ratio at time t, is the LF / HF threshold value, preferably 2.5;

[0074] A blood flow corrected human heat balance equation is constructed based on the heat dissipation amount and the blood flow correction coefficient, and a physiological feature vector is generated; the physiological feature vector includes predicted core temperature , predicted average vote , and predicted percentage of dissatisfaction ; the expression of the physiological feature vector is: , to realize accurate mapping from physiological signals to thermal comfort state;

[0075] The blood flow correction coefficient corrects the human body heat balance equation as follows:

[0076] ;

[0077] wherein, is metabolic heat production (unit: W), is convective heat loss (unit: W), is radiative heat loss (unit: W), is evaporative heat loss (unit: W), is a blood flow correction coefficient, is the heat capacity of the human body (unit: J / K), is the core body temperature of the human body (unit °C), is the rate of change of the core body temperature of the human body with time;

[0078] The original heat transfer equation of the blood flow corrected human body heat balance model is

[0079] wherein, is the thermal diffusivity, is the metabolic heat density, is the density of human tissue, is the specific heat capacity, is the core-skin temperature difference coefficient, is the blood flow influence factor;

[0080] Based on the heat loss and the blood flow correction coefficient, a blood flow corrected human body heat balance equation is constructed, and a physiological feature vector is generated, and the method is as follows:

[0081] The human body heat balance equation is integrated to obtain a predicted core body temperature;

[0082] The integral operation formula is as follows:

[0083] ;

[0084] wherein, is the predicted core body temperature at time t, is the initial core body temperature, t is the integral time variable, the lower limit is 0 and the upper limit is t, is the blood flow correction coefficient at time , the value range is [0.6, 1.0], between the lower limit 0 and the upper limit t;

[0085] Based on the predicted core body temperature , the physical environment data (cabin temperature , humidity , air flow velocity , the average skin physiological data (T ), the predicted average vote is obtained by a thermal comfort model Fanger , and then the predicted percentage of dissatisfaction is obtained ;

[0086] Specifically:

[0087] ;

[0088] ;

[0089] wherein, W is external work, the value of W is estimated by a joystick pressure sensor, conventional flight: W ≈ 0 (joystick force is very small); stunt / emergency: W = 15-25 W / m² (high G force against work); ejection escape: W = 30-50 W / m² (instantaneous impact work);

[0090] The decision layer is configured to: based on the physiological feature vector, train the task load instruction, generate a temperature control instruction and a confidence evaluation value by using a meta-learning model, wherein a reward function of the meta-learning model is a multi-objective dynamic weighted reward function, which includes four-dimensional targets of comfort, energy consumption, stability and response speed;

[0091] The weights of the reward function are dynamically adjusted according to real-time states by IF-THEN rules; the real-time states include: predicted core body temperature, predicted percentage of dissatisfaction, training task load level, and safety margin;

[0092] In this embodiment, the multi-objective dynamic weighted reward function is:

[0093] ;

[0094] wherein, : total reward function value, : system state vector at time t, : control action vector at time t, , , , : dynamic weight coefficient (sum = 1);

[0095] Constraint condition:

[0096] ;

[0097] wherein:

[0098] ;

[0099] ;

[0100] ;

[0101]

[0102] reward for comfort, predicted average vote for target, current actual predicted average vote (i.e. predicted average vote obtained by thermal comfort model ), reward for energy consumption, total power consumption of system, rated power of system, reward for temperature stability, standard deviation, recent temperature of cabin, reward for response speed, temperature after system response, i.e. temperature reached by cabin after executing control action, optimal temperature, representing ideal cabin temperature;

[0103] The IF-THEN rules include: safety critical rule, energy efficiency priority rule, steady state optimization rule; intelligent adaptation of optimization target is achieved by precise matching of scene characteristics, and dynamic weight vector W ( st ) as a function of system state st is not a fixed constant, but is adaptively adjusted according to training task type, real-time physiological state of pilot and external environment parameters, and the specific rule logic is as follows: safety critical rule: when the system identifies a safety critical state (such as detecting a high stress scene through HRV feature parameters, sudden change of HRV signal, etc. physiological warning signals), the "comfort weight priority promotion" mechanism is automatically triggered, and the comfort weight W 1 is dynamically increased by 50% to prioritize pilot physiological safety and operation stability; energy efficiency priority rule: when in low load cruise phase or high energy consumption period (such as power grid peak electricity consumption), the "energy saving target intensification" strategy is started, and the energy consumption weight W 3 is increased by 30%, guiding the system to take energy efficiency optimization control strategy under the premise of meeting basic comfort; steady state optimization rule: in the steady state phase of regular training, based on the cabin environment stability index and response speed demand, the proportion of stability weight W 2 and response speed weight W 4 is dynamically balanced to ensure that the temperature control system realizes optimal balance between precision and dynamic response;

[0104] The set of structured rules guarantees the scientificity of weight adjustment and realizes flexible switching of optimization goals in different scenarios, effectively improving the rationality and robustness of system control.

[0105] The meta-learning model acquisition method comprises the following steps:

[0106] The meta-training phase specifically comprises the following steps: task construction, a "task" is defined as time series data from a specific pilot in a training, and time series data of a specific object single task is defined as a meta-learning data unit Wherein, the state vector S (State) such as multi-modal perception layer data, the action vector A (Action) such as temperature control instruction, the reward value R (Reward) and the new state vector S' after executing the action vector, the meta-training dataset Comprises N different pilot tasks.

[0107] The meta-learning model of the teacher network and the student network is constructed, the knowledge of the teacher network is migrated to the student network based on the meta-training dataset and is trained, and the student network parameter θ after meta-training is obtained, wherein the teacher network is trained using historical data (i.e. state vector S), the loss function contains physical constraint loss, and the output is temperature control instruction prediction distribution and confidence; the loss function of the student network includes KL divergence and hard label loss, realizes lightweight deconstruction, and solves the problem of insufficient model generalization ability;

[0108] The loss function of the teacher network is :

[0109] ;

[0110] Wherein, is the physical constraint loss, is the comfort loss function, is the comfort label / target value of the i-th sample, is the temperature control output prediction value of the teacher network for the i-th input sample, is the weight coefficient of the physical constraint loss. The loss function of the student network is :

[0111]

[0112] ;

[0113] Wherein, is the weight coefficient, ​​​This is the raw output of the student network to input x. This is the raw output of the teacher network to input x. Temperature parameter controls the smoothness of the soft label. KL divergence measures the difference between the distributions of student and teacher outputs. Hard label loss;

[0114] Hierarchical parameter decoupling is used to decouple student network parameters Divided into general layer parameters Task layer parameters and personality layer parameters These are used to model the universal laws of human thermophysiology (such as basal metabolic rate and heat dissipation mechanisms), the characteristics of different training tasks (such as takeoff and landing, aerobatic maneuvers, and emergency situations), and individual differences among pilots (such as physical condition, preferences, and adaptability); the specific expressions are as follows: ;

[0115] New pilot adaptation phase: When a new pilot joins, collect the time series data (S, A, R, S') of the new pilot for a preset duration (first 10 minutes) as the fine-tuning dataset D_finetune;

[0116] Load the meta-trained student network parameters θ, and perform finite gradient descent update with D_finetune as input to obtain personalized student network parameters θ';

[0117] The finite gradient descent update employs dynamic parameter freezing and selective fine-tuning. The dynamic parameter freezing and selective fine-tuning are as follows: the parameter freezing ratio is calculated based on the similarity between the model parameters of the new pilot and the historical pilot. The parameter freezing ratio is combined with the training time to select and unfreeze parameters of different dimensions in stages to obtain personalized model parameters.

[0118] The formula for calculating the freezing ratio of the parameter is:

[0119] ;

[0120] in, The parameter is the freeze ratio. This is a natural exponential function used to map similarity to a freeze ratio that follows an exponential decay law. Here, represents the similarity calculation function, indicating the initial model parameters for the new pilot. Cluster centers of parameters from historical pilot models Similarity measure This is a freeze adjustment parameter used to control the degree of influence of similarity on the freeze ratio.

[0121] The parameter freezing ratio is combined with training duration to select parameters of different dimensions to be unfrozen in stages, including: an initial stage (0-5 minutes), only the individual layer parameters are unfrozen ; an adaptation stage (5-10 minutes), the individual layer parameters are unfrozen and part of the task layer parameters are unfrozen ; a stable stage (>10 minutes), all parameters are unfrozen and elastic weight regularization is used to constrain the update of key parameters to prevent catastrophic forgetting; the elastic weight regularization formula is:

[0122] ;

[0123] wherein, is the total loss function of the elastic weight regularization, is a diagonal element of the Fisher information matrix, which measures the importance of the parameter, is the current task loss, is a regularization weight coefficient, is the value of the i-th parameter of the current model update, is the optimal value of the i-th parameter after the training of the old task (historical pilots), is the square term of the difference between the new and old parameters; The student network parameter θ after the meta-training is loaded, and the D_finetune is input to perform limited gradient descent update to obtain the individualized student network parameter θ', and the feature space alignment and transfer optimization update are further included: the feature distribution difference between the new and old pilots is minimized by maximum mean difference, the individual layer feature weight is dynamically adjusted by a progressive fusion function, and then the parameter gradient of the limited gradient descent update is corrected to realize individualized adaptation and optimization;

[0124] The feature distribution difference between the new and old pilots is minimized by maximum mean difference, and the method is as follows:

[0125] The state vector S in the fine-tuning data set D_finetune of the new pilot (i.e. the target pilot) is input to the shared feature extractor;

[0126] The distribution difference between the extracted new pilot feature representation and the general feature representation space obtained by the old pilot (learned in the meta-training stage) is calculated;

[0127] The minimum feature alignment loss function is applied to minimize the distribution difference, so that the feature representation of the new pilot is effectively projected into the general feature representation space;

[0128] The aligned target pilot feature representation is used to drive the parameter update of the subsequent decision layer;

[0129] The aligned target pilot feature representation is used to drive the parameter update of the subsequent decision layer; ​

[0130] where the minimization feature alignment loss formula is:

[0131] ;

[0132] where, is the maximum mean difference, is the feature distribution of the source pilot (“old data” distribution), is the feature distribution of the target pilot (“new data” distribution), is the number of source pilot feature samples (sample size), is the number of target pilot feature samples (sample size), is a single feature sample of the source pilot, is a single feature sample of the target pilot, is a feature mapping function, is the mean vector of the source pilot feature in high-dimensional space, is the mean vector of the target pilot feature in high-dimensional space, is the squared norm in the reproducing kernel Hilbert space (RKHS, denoted as H);

[0133] The progressive fusion function is:

[0134] ;

[0135] ;

[0136] where, is the final feature representation after fusion, dynamic weight coefficient of personalized features (changes with time t, value range 【0,1】 universal feature representation obtained by meta-training (basic feature) personalized feature representation of the target pilot (exclusive feature);

[0137] As the adaptation time increases, the personalized layer feature weight gradually increases;

[0138] In this way, through multi-objective dynamic optimization and rapid personalized adaptation, the problems of reward function simplification and long cold start period are solved;

[0139] The student network updated using the parameter θ' generates real-time control instructions to update the learning model, obtaining the temperature control instructions :

[0140] ;

[0141] where, is the set point temperature of the temperature control system, A fan rotation speed control parameter, A ventilation mode selection;

[0142] A confidence evaluation value: ;

[0143] The value is calculated based on the decision uncertainty of the personalized model θ' for the new input state or the similarity to the distribution of the meta-training data; when the Confidence_score is lower than a preset threshold, a data sampling update request is triggered or an adjustment to a conservative control strategy is made to provide personalized temperature control;

[0144] A dynamic execution layer configured to: based on the temperature control instruction and the average skin physiological data, adopt PID closed-loop adjustment to adjust the operating parameters of the air flow regulation equipment of each temperature control partition of the flight simulator cockpit, and then regulate the temperature of each partition to achieve precise and differentiated air supply to each part of the pilot's body.

[0145] Each temperature control partition of the flight simulator cockpit includes multiple independent temperature control zones, such as the head, torso, and legs; each region is equipped with independent air flow regulation equipment, such as fans and valves, and air conditioners.

[0146] Based on the real-time numerical interval of the predicted core body temperature and the predicted dissatisfaction percentage, progressive temperature control responses are triggered in stages, including local adjustment, composite adjustment, and global intervention warning, to achieve a three-level active safety response.

[0147] The 1st level response (early warning intervention) of local adjustment is triggered when the predicted core body temperature exceeds 37.3℃, and immediate local strong air supply in the high-temperature area is started.

[0148] The 2nd level response (cooperative cooling) of composite adjustment is triggered when the PMV is greater than 1.0 in addition to the 1st level response, and the seat ventilation is started and the fresh air exchange rate is increased.

[0149] The 3rd level response (forced refrigeration) of global intervention warning is triggered when the predicted dissatisfaction percentage further deteriorates to greater than 2.0, and global high-power refrigeration is forced to start, and an alarm is sent to the monitoring console to ensure the absolute physiological safety of the pilot.

[0150] In this way, through precise partition control and active safety protection, a full-chain safety protection from early warning to forced intervention is formed to eliminate potential health risks and solve the problem of traditional control roughness and passive response.

[0151] It should be noted that the flight simulation cockpit intelligent temperature control system provided by the above embodiment is only exemplified by the division of the above functional modules, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the modules or steps in the embodiment of the application are further decomposed or combined, for example, the modules of the above embodiment can be combined into one module, or can be further split into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the application are only for the purpose of distinguishing various modules and steps, and should not be considered as an improper limitation of the application.

[0152] Those skilled in the art will appreciate that the modules, method steps of the examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The software modules, method steps corresponding to the program can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been described in the above description in general terms. Whether the functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0153] The terms "first", "second", and the like are used to distinguish similar objects, not to describe or indicate a particular order or sequence.

[0154] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, so that the process, method, article or device / apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed, or inherent to the process, method, article or device / apparatus.

[0155] So far, the technical solution of the application has been described in conjunction with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without departing from the principles of the application, and the technical solutions after the changes or replacements will fall within the protection scope of the application.

Claims

1. An intelligent temperature control system for a flight simulator cockpit, characterized in that, The system comprises: A multi-modal perception layer configured to: acquire multi-channel skin physiological data and average processing to obtain average skin physiological data; acquire physical environment data and calculate heat dissipation in combination with the average skin physiological data; acquire a pilot's electrocardio signal, pre-process and extract features to obtain a feature parameter; normalize the feature parameter to obtain a blood flow correction coefficient; Based on the heat dissipation and the blood flow correction coefficient, a blood flow corrected human heat balance equation is constructed to generate a physiological feature vector; the physiological feature vector includes a predicted core body temperature, a predicted average vote, and a predicted dissatisfaction percentage; A decision layer configured to: based on the physiological feature vector and a training task load instruction, generate a temperature control instruction and a confidence evaluation value using a meta-learning model; wherein the reward function of the meta-learning model is a multi-objective dynamic weighted reward function, including comfort, energy consumption, stability, and response speed in four dimensions; the weight of the reward function is dynamically adjusted according to the real-time state through IF-THEN rules; the real-time state includes the predicted core body temperature, the predicted dissatisfaction percentage, the training task load level, and the safety margin; A dynamic execution layer configured to: based on the temperature control instruction and the average skin physiological data, adjust the operating parameters of the airflow regulation equipment of each temperature control partition of the flight simulator cockpit using a PID closed-loop to control the temperature of each partition; Based on the real-time numerical interval of the predicted core body temperature and the predicted dissatisfaction percentage, a progressive temperature control response is triggered in stages, including local adjustment, compound adjustment, and global intervention alarm; The method for obtaining a pilot's electrocardio signal and pre-processing and feature extraction to obtain a feature parameter is: Obtaining a pilot's electrocardio signal; Pre-processing the electrocardio signal, which includes detecting the R peak of the electrocardio signal using the Pan-Tompkins algorithm and generating an RR interval sequence, and removing and interpolating outliers in the RR interval sequence to obtain a pre-processed electrocardio signal; Feature extraction of the pre-processed electrocardio signal through short-time Fourier transform to obtain a feature parameter, including low-frequency power, high-frequency power, and LF / HF ratio; Based on the extracted LF / HF ratio, a blood flow correction coefficient is obtained by a normalized sigmoid function ; ； wherein, is a blood flow correction coefficient at time t, is a minimum correction coefficient, is an adjustment parameter, is a real-time LF / HF ratio at time t, is an LF / HF threshold value.

2. The intelligent temperature control system for a cockpit of a flight simulator according to claim 1, wherein, The human heat balance equation corrected by the blood flow correction coefficient is: ； wherein, is the metabolic heat production, is the convective heat loss, is the radiative heat loss, is the evaporative heat loss, is the blood flow correction factor, is the human body heat capacity, is the human body core temperature, is the rate of change of the human body core temperature with respect to time.

3. The intelligent temperature control system for a cockpit of a flight simulator according to claim 1, wherein, Based on the heat dissipation and the blood flow correction coefficient, a blood flow corrected human heat balance equation is constructed to generate a physiological feature vector, and the method is: Integrating the human heat balance equation to obtain a predicted core body temperature; Based on the predicted core body temperature, the physical environment data, and the average skin physiological data, a predicted average vote is obtained through a thermal comfort model, and then a predicted dissatisfaction percentage is obtained.

4. The intelligent temperature control system for a cockpit of a flight simulator according to claim 1, wherein, The physical environment data includes cabin temperature, humidity, and airflow speed; the heat dissipation includes convective heat dissipation, radiative heat dissipation, and evaporative heat dissipation; The multi-channel skin physiological data includes 12-channel skin temperature and 12-channel skin humidity; the average skin physiological data includes average skin temperature and average skin humidity.

5. The intelligent temperature control system for a cockpit of a flight simulator according to claim 1, wherein, The meta-learning model acquisition method is: Defining a meta-training dataset D_meta-train, wherein D_meta-train contains N tasks of different pilots, each task is defined as a time series data sequence {(S, A, R, S')} from a specific pilot in a training; S is a state vector, A is an action vector, R is a reward value, and S' is a new state vector after executing the action vector; Building a meta-learning model of a teacher network and a student network structure, based on the meta-training dataset, transferring the knowledge of the teacher network to the student network through knowledge distillation and training to obtain the meta-trained student network parameters θ, wherein the teacher network is trained using historical data, and the loss function contains a physical constraint loss; the loss function of the student network includes KL divergence and hard label loss; Decoupling the student network parameters θ into general layer parameters, task layer parameters and individual layer parameters by hierarchical parameter decoupling; When a new pilot joins, collect the time series data of the new pilot for a preset time length as a fine-tuning dataset D_finetune; Loading the meta-trained student network parameters θ, and performing limited gradient descent update with the D_finetune as input to obtain personalized student network parameters θ'; Using the student network updated by the parameters θ' to obtain the temperature control instruction and the confidence evaluation value.

6. The intelligent temperature control system for a cockpit of a flight simulator according to claim 5, wherein, The limited gradient descent update adopts dynamic parameter freezing and selective fine-tuning; The dynamic parameter freezing and selective fine-tuning is: calculating a parameter freezing ratio according to the similarity of the student network parameters of the new pilot and the historical pilots, combining the training time length to select the unfreezing of parameters of different dimensions in stages, performing limited gradient descent update, and then obtaining personalized model parameters.

7. The intelligent temperature control system for a cockpit of a flight simulator according to claim 5, wherein, The loading of the meta-trained student network parameters θ, and the performing of the limited gradient descent update with the D_finetune as input to obtain the personalized student network parameters θ', further includes feature space alignment and migration optimization update: minimizing the feature distribution difference between the new and old pilots by maximum mean difference, dynamically adjusting the individual layer feature weight by using a progressive fusion function, and then correcting the parameter gradient of the limited gradient descent update.

8. The intelligent temperature control system for a cockpit of a flight simulator according to claim 6, wherein, The parameter freezing ratio calculation formula is: ； wherein, is a parameter freezing ratio, is a natural exponential function, used to map the similarity into a freezing ratio that complies with the exponential decay law, is a similarity calculation function, representing the similarity measure between the initial model parameters of the new pilot and the cluster centers of the historical pilot model parameters, is a freezing adjustment parameter, used to control the degree of influence of the similarity on the freezing ratio;​ The parameter freezing ratio combines the training time length to select the unfreezing of parameters of different dimensions in stages, including: in the initial stage, only the individual layer parameters are unfreezed; in the adaptation stage, the individual layer parameters are unfreezed and part of the task layer parameters are unfreezed; in the stable stage, all parameters are unfreezed and elastic weight regularization is used to constrain the key parameter update.

9. The intelligent temperature control system for a cockpit of a flight simulator according to claim 1, wherein, The IF-THEN rule includes: safety critical rule, energy efficiency priority rule, and steady state optimization rule.

Images