Indoor environment monitoring method based on thermal comfort analysis

Abstract

The invention discloses an indoor environment monitoring method based on thermal comfort analysis, and particularly relates to the field of indoor environment industrial data processing and monitoring, which comprises the following steps: deploying a temperature and humidity sensor and a carbon dioxide sensor based on gridding, dynamically clustering a personnel dense area and migrating nodes, the Manhattan distance and random sampling consistency technology is combined to execute fault self-inspection and calibration; infrared and millimeter wave radars and a color depth camera are integrated to acquire radiation temperature and physiological data. The sensor layout is optimized through the dynamic clustering algorithm, and the fault self-inspection technology is integrated, so that the monitoring reliability of the dense area is improved; multi-modal data are aligned by using space-time diagram convolution, and privacy security is ensured in combination with a dual-encryption mechanism; the multi-region air distribution is coordinated based on the block chain to avoid conflicts; adopting reinforcement learning to construct a heat transfer model fused with physiological parameters to enhance simulation adaptability; and on the basis of regulation and control of federal learning linkage equipment, global optimization of industrial data processing of personalized thermal comfort and energy consumption balance is realized.

Description

Technical Field

[0001] The present invention relates to the field of indoor environment industrial data processing and monitoring, and more particularly to an indoor environment monitoring method based on thermal comfort analysis. Background Art

[0002] Indoor environmental monitoring is a technology system that uses an IoT sensor network to collect parameters such as temperature, humidity, CO2 concentration, PM2.5, and light intensity in real time, building a dynamic perception system through wireless communication. It adopts a distributed deployment strategy, integrating temperature, humidity, gas, and particulate matter sensors to achieve multi-dimensional data collection, and aggregates data to the cloud using low-power transmission protocols such as LoRa / NB-IoT. It combines time series prediction algorithms with anomaly detection models to accurately identify the risk of sudden changes in environmental parameters.

[0003] Existing indoor environment monitoring technologies have limitations in the processing and regulation of industrial data on thermal comfort. Fixed grid sensors are difficult to adjust with occupant density, resulting in monitoring blind spots and waste of resources. Multimodal data is spatiotemporally asynchronous, with low fusion accuracy and insufficient privacy protection. Fault detection relies on a single threshold and fails to consider sensor migration errors, resulting in a high false alarm rate. Variable air volume systems lack global coordination and industrial data processing, leading to serious air supply conflicts. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an indoor environment monitoring method based on thermal comfort analysis, which improves monitoring accuracy through dynamic clustering and sensor migration, and combines multimodal data alignment and encryption technology to ensure synchronization and privacy; optimizes the air supply strategy based on the contribution value algorithm and blockchain to achieve fair regulation and traceability; integrates environmental and physiological data to construct a thermal comfort model, strengthens learning adaptive correction, and improves simulation accuracy to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring indoor environment based on thermal comfort analysis, comprising:

[0006] Based on the grid deployment of temperature, humidity and carbon dioxide sensors, the system dynamically clusters densely populated areas and migrates nodes, and combines Manhattan distance and random sampling consistency technology to perform fault self-detection and calibration.

[0007] Deploy temperature, humidity, and carbon dioxide sensors in a grid pattern across the area, and record the initial coordinates and measurement data of each node;

[0008] Identify densely populated areas through dynamic clustering algorithms and calculate the spatial distribution relationship between cluster centers and sensor nodes;

[0009] Manhattan distance is used to calculate the deviation of measurement values between sensors and screen out faulty nodes. After screening out the faulty nodes, the data of the screened nodes is compared using a random sampling consistency algorithm to determine whether they are abnormal data sources and mark the faulty nodes.

[0010] Based on the fault mark, the self-test program is executed to determine whether the sensor needs to be automatically calibrated or a maintenance instruction is triggered. For nodes that need calibration, the system starts the automatic calibration process; for nodes that cannot be self-calibrated, the maintenance instruction is triggered for manual intervention, and the node migration strategy is updated based on historical monitoring data.

[0011] In a preferred embodiment, an infrared sensor, a millimeter-wave radar, and a color depth camera are integrated to collect radiation temperature and physiological data, and the collected radiation temperature and physiological data are aligned in time and space to ensure synchronization of the sensor data;

[0012] The aligned data is processed through the spatiotemporal graph convolutional network to extract features and establish data associations.

[0013] In a preferred embodiment, data from anemometers and gas sensors are collected, and the wind speed change rate and gas concentration gradient of each area are calculated to determine whether the current air volume distribution meets the air supply requirements of each area. If not, the air volume adjustment process is initiated;

[0014] Calculate the air supply demand weight of each area and determine whether the calculated air volume distribution matches the current variable air volume supply parameters. If not, adjust the variable air volume supply set value;

[0015] The adjustment records are stored in the blockchain, and the air volume distribution status of each area is updated synchronously. The wind speed changes after adjustment are monitored to determine whether there is an air supply conflict. If a conflict is detected, the conflict resolution strategy is executed. If the air volume adjustment still cannot meet the demand, the air supply mode reconstruction is triggered and the global air volume distribution plan is recalculated.

[0016] In a preferred embodiment, environmental data and physiological data are collected, temperature gradient, humidity change rate and human body heat exchange parameters are calculated, and stored in a data set to determine data integrity;

[0017] A thermal comfort model is constructed based on the collected data, and the heat transfer coefficient is calculated to initialize the boundary conditions. A reinforcement learning algorithm is used to optimize the thermal comfort model, and the heat transfer coefficient is dynamically adjusted to adapt to different working conditions. The convergence of the modified boundary conditions is determined.

[0018] The operating condition simulation is performed in combination with the corrected boundary conditions, and the deviation between the simulation results and the historical data is calculated. If the deviation of the simulation results exceeds the threshold, the parameter weights of the thermal comfort model are adjusted and the simulation verification is re-executed. The simulation results are continuously monitored, the optimization process is recorded, and the robustness of the thermal comfort model under different operating conditions is evaluated.

[0019] In a preferred embodiment, local human body feature data is collected, skin temperature, metabolic rate and posture information are extracted, and encryption is performed to determine data integrity. If key features are missing or encryption fails, data re-collection or exception handling mechanisms are triggered;

[0020] Federated learning is used to train local models, calculate individual thermal comfort parameters, generate optimization strategies, and determine the convergence of local models.

[0021] Calculate the global radiation temperature adjustment plan, and link the building management system to dynamically adjust the cooling and heating equipment to determine the thermal comfort feedback of the human body after the adjustment.

[0022] In a preferred embodiment, a dynamic clustering algorithm is constructed based on Equation 1-1, which includes the node coordinates and real-time population density in the initial sensor network. Based on the population density distribution, the position of the cluster center is dynamically adjusted using an exponential decay weight, so that the cluster center is shifted toward the densely populated area. By numerically solving the equation system of Equation 1-1, the decay rate is controlled to increase the contribution of close-range nodes to the migration of the cluster center. The updated cluster center triggers the physical migration of the sensor nodes.

[0023] Formula 1-1 is expressed as:

[0024]

[0025] in, is the new center position of the k-th cluster after t+1 iterations; is the summation operation of all data points in the cluster, N k represents the set of sensor nodes in the kth cluster; ω i is the density weight of data point i, ρ i is the density of people around data point i; x i is the eigenvector of data point i; is an exponential decay function; γ represents the migration attenuation coefficient, R max is the maximum value of the sensor’s sensing radius, To calculate the sample x i With the current cluster center The L1 norm distance and Manhattan distance of ; ∥1 is the Manhattan distance operator;

[0026] Construct the fault detection robustness criterion based on formula 1-2;

[0027] Formula 1-2 is expressed as:

[0028]

[0029] Among them, D(s i ) is the fault judgment index; is the neighborhood summation operation, N i represents the set of neighboring nodes of node i; T i is the temperature and humidity of node i; C i is the measured value of carbon dioxide concentration; T j is the time attribute of neighbor node j; C j is the cost attribute of neighbor node j; Δt ij is the time synchronization error difference between nodes i and j; α T is the time-based weight; α C is the cost basis weight; β T is the time dynamic weight coefficient; β C is the dynamic weight coefficient of cost; τ d is the fault judgment threshold; in formula 1-2, dynamic denominator design and multi-parameter joint detection are used.

[0030] In a preferred embodiment, a multimodal alignment and anonymization formula is constructed based on Equation 2-1, which includes a multimodal spatiotemporal graph convolutional network for alignment processing of infrared, millimeter-wave radar, and color depth camera data, and protects biometric privacy through adversarial robustness and homomorphic encryption technology;

[0031] Formula 2-1 is expressed as:

[0032]

[0033] Among them, H (l+1) is the feature matrix of the l+1th layer node; σ is the activation function; is the summation operation of multi-scale convolution kernels; T k (Λ) is the Chebyshev polynomial basis function; ⊙ is the Kronecker product operator; G k is the space-time mixing matrix; H (l) is the feature matrix of the l-th layer node; is the trainable parameter matrix;

[0034] Construct adversarial perturbation generation based on Equation 2-2;

[0035] Formula 2-2 is expressed as:

[0036]

[0037] in, To minimize the perturbation vector δ, δ represents the perturbation vector, and its dimension is the same as the input data x; To combat the perturbation loss function; is to constrain the perturbation direction along the direction with the lowest model sensitivity; λ·KL(f(x+δ)∥f(x)) is the characteristic distribution after perturbation and the original distribution; J f (x) is the Jacobian matrix of the feature extraction model; KL(.) is the relative entropy divergence; λ is the anti-dynamic strength control parameter.

[0038] In a preferred embodiment, the air supply priority of high gradient areas is strengthened by exponential weight distribution, and the volume term avoids oversaturation of small areas to achieve Pareto optimal air volume distribution. Based on this, the air volume distribution is optimized through formula 3-1;

[0039] The air volume distribution optimization of formula 3-1 is expressed as:

[0040]

[0041] Among them, w r is the contribution value weight of region r; exp is the exponential function; -η is the contribution sensitivity coefficient; is the α volume scaling factor; V r is the volume of region r; is the wind pressure gradient in region r; is the pollutant concentration gradient in region r; is the mapping of the quadratic gradient term to a linear scale; r′∈R exp(·) is global normalization; Formula 3-1 realizes Pareto optimal air volume distribution through exponential weight distribution.

[0042] In a preferred embodiment, the thermal comfort prediction error is minimized by dynamically adjusting the convective heat transfer coefficient to balance the human body heat load and the environmental heat transfer efficiency in real time;

[0043] The dynamic correction expression of heat transfer coefficient in formula 4-1 is:

[0044]

[0045] in, is the convective heat transfer coefficient of the next time step; is the convective heat transfer coefficient at the current moment; ζ is the learning rate attenuation factor; ReLU is the hyperbolic tangent function; is the comfort model for h c The partial derivative of is the loss function of the model for predicting average voting comfort; is the convective heat transfer coefficient, tanh(·) is the hyperbolic tangent function; ΔT skinis the skin temperature gradient; is the time derivative of metabolic rate; Equation 4-1 realizes thermal comfort modeling by dynamically correcting the heat transfer coefficient;

[0046] In federated learning strategy optimization, secure aggregation of distributed encrypted gradients is achieved through homomorphic encryption and dynamic weight distribution of divergence;

[0047] The encrypted gradient aggregation expression of formula 5-1 is:

[0048]

[0049] Among them, g global Encrypt gradients globally; is the client gradient product operation, and M is the number of clients participating in federated learning; Weighted exponentiation of the client-side encrypted gradient, π k is the homomorphically encrypted gradient using end k, θ is the encryption parameter, π k is the dynamic weight coefficient; mod N 2 It is a modular operation.

[0050] The technical effects and advantages of the present invention are as follows:

[0051] 1. Through dynamic clustering algorithms and sensor node migration, it can track densely populated areas in real time. It also combines Manhattan distance and random sampling consistency technology to achieve fault self-detection and calibration, thereby relatively improving monitoring accuracy and system reliability.

[0052] 2. Integrate infrared, millimeter-wave radar and other sensors, use spatiotemporal graph convolutional networks to align multimodal data, and anonymize biometric features through adversarial perturbation and homomorphic encryption technology to ensure data synchronization while protecting user privacy;

[0053] 3. Optimize variable air volume supply parameters based on the contribution value algorithm and combine blockchain to record the adjustment process, achieving efficient multi-region air supply coordination and conflict resolution, ensuring the fairness and traceability of air conditioning;

[0054] 4. Integrate environmental and physiological data to build a dynamic thermal comfort model. Use reinforcement learning to modify the heat transfer coefficient and boundary conditions in real time, improve the robustness of the model under different working conditions, and reduce simulation errors.

[0055] 5. Use federated learning to locally train individual thermal comfort parameters and upload them in encrypted form, linking the building management system to globally adjust equipment, achieving personalized environmental optimization and energy consumption balance while protecting user data privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 The process of the present invention Figure 1 ;

[0057] Figure 2 The process of the present invention Figure 2 ;

[0058] Figure 3 The process of the present invention Figure 3 . DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0060] Refer to the instruction manual Figure 1 An indoor environment monitoring method based on thermal comfort analysis according to an embodiment of the present invention includes:

[0061] Temperature, humidity, and carbon dioxide sensors are deployed in a grid-based manner. Monitoring density is increased by dynamically clustering densely populated areas and migrating nodes. Fault self-diagnosis and calibration are performed using Manhattan distance and random sampling consistency techniques.

[0062] Deploy temperature, humidity, and carbon dioxide sensors in a grid pattern across the area, and record the initial coordinates and measurement data of each node;

[0063] A dynamic clustering algorithm is used to identify densely populated areas and calculate the spatial distribution relationship between cluster centers and sensor nodes. The positions of sensor nodes are adjusted according to the changes in densely populated areas, so that the number of monitoring nodes in high-density areas increases.

[0064] Manhattan distance is used to calculate the deviation of measurement values between sensors and screen out faulty nodes. After screening out the faulty nodes, the data of the screened nodes is compared using a random sampling consistency algorithm to determine whether they are abnormal data sources and mark the faulty nodes.

[0065] Based on the fault flags, the system executes a self-check procedure to determine whether the sensor needs to be automatically calibrated or a maintenance instruction is triggered. For nodes that need calibration, the system starts the automatic calibration process. For nodes that cannot self-calibrate, a maintenance instruction is triggered for manual intervention. The node migration strategy is updated based on historical monitoring data.

[0066] It should also be noted that the system integrates infrared sensors, millimeter-wave radars, and color depth cameras to collect radiation temperature and physiological data, and aligns the collected radiation temperature and physiological data in time and space to ensure synchronization of sensor data.

[0067] The aligned data is processed through a spatiotemporal graph convolutional network to extract features and establish data associations. During the data processing process, an anti-perturbation algorithm is used to detect and suppress potential abnormal noise in the data. The data is encrypted using homomorphic encryption and anonymized. After anonymization is completed, the data integrity is verified and the processed data is stored in a designated database for subsequent analysis.

[0068] Collect data from anemometers and gas sensors, calculate the wind speed change rate and gas concentration gradient in each area, and determine whether the current air volume distribution meets the air supply needs of each area. If not, enter the air volume adjustment process;

[0069] Calculate the air supply demand weight of each area and use the air volume allocation algorithm based on contribution value calculation to calculate the optimal air volume allocation plan based on the impact of each area on the overall air supply system; determine whether the calculated air volume allocation matches the current variable air volume supply parameters; if not, adjust the variable air volume supply set value;

[0070] In addition, the blockchain stores adjustment records and synchronously updates the air volume distribution status of each area. The wind speed changes after adjustment are monitored to determine whether there is an air supply conflict. If a conflict is detected, the conflict resolution strategy is implemented. If the air volume adjustment still cannot meet the demand, the air supply mode is reconfigured and the global air volume distribution plan is recalculated. The air volume adjustment results are continuously recorded, and the effectiveness of the air supply strategy is regularly evaluated to ensure the coordination and stability of multiple areas.

[0071] Collect environmental and physiological data, calculate temperature gradients, humidity change rates, and human heat exchange parameters, and store them in a data set. Check data integrity and, if key parameters are missing, trigger data completion or remove abnormal data.

[0072] A thermal comfort model is constructed based on the collected data, and the heat transfer coefficient is calculated to initialize the boundary conditions. A reinforcement learning algorithm is used to optimize the thermal comfort model and dynamically adjust the heat transfer coefficient to adapt to different working conditions. Convergence of the modified boundary conditions is determined. If not, the learning rate is adjusted and the model is retrained.

[0073] Execute the operating condition simulation based on the corrected boundary conditions and calculate the deviation between the simulation results and the historical data. If the deviation exceeds the threshold, adjust the parameter weights of the thermal comfort model and re-execute the simulation verification. Continuously monitor the simulation results, record the optimization process, and evaluate the robustness of the thermal comfort model under different operating conditions.

[0074] Collect local human feature data, extract skin temperature, metabolic rate and posture information, encrypt it, and judge data integrity. If key features are missing or encryption fails, data re-collection or exception handling mechanisms will be triggered;

[0075] It should also be noted that federated learning trains local models, calculates individual thermal comfort parameters, generates optimization strategies, and determines local model convergence. If the target accuracy is not achieved, the training parameters are adjusted and optimization continues. The encrypted optimization strategies are uploaded to the federated server to participate in global thermal comfort and energy consumption optimization calculations.

[0076] Calculate the global radiant temperature adjustment plan and link it with the building management system to dynamically adjust the cooling and heating equipment. Determine the human thermal comfort feedback after adjustment. If it deviates from the target range, it iteratively update the adjustment strategy, continuously monitor energy consumption data, calculate the adjustment benefits, and adjust the federated learning weights to optimize the long-term strategy.

[0077] A dynamic clustering algorithm is constructed based on Equation 1-1, which includes the node coordinates and real-time population density in the initial sensor network. Based on the population density distribution, the position of the cluster center is dynamically adjusted through exponential decay weights, so that the cluster center shifts toward densely populated areas, and the offset amplitude is proportional to the density, thereby deploying more sensor nodes in dense areas. By numerically solving the equation system of Equation 1-1, the decay rate is controlled, so that the contribution of close-range nodes to the migration of the cluster center is increased, avoiding interference from distant low-density areas. The updated cluster center triggers the physical migration of sensor nodes.

[0078] Formula 1-1 is expressed as:

[0079]

[0080] in, is the new center position of the k-th cluster after t+1 iterations; is the summation operation of all data points in the cluster, N k represents the set of sensor nodes in the kth cluster; ω i is the density weight of data point i, ρ i is the density of people around data point i; x i is the eigenvector of data point i; is an exponential decay function; γ represents the migration attenuation coefficient, R max is the maximum value of the sensor’s sensing radius, To calculate the sample x i With the current cluster center The L1 norm distance and Manhattan distance of the cluster center are calculated by ∥1, and the Manhattan distance is used to adapt to the indoor grid layout characteristics.

[0081] Construct the fault detection robustness criterion based on formula 1-2;

[0082] Formula 1-2 is expressed as:

[0083]

[0084] Among them, D(s i ) is the fault judgment index, which represents the abnormality score of node i. The larger the value, the more serious the deviation of the data from the normal state. i ) is greater than the fault judgment threshold τ d It is judged as a fault when is the neighborhood summation operation, N i represents the set of neighboring nodes of node i; T i is the temperature and humidity of node i; C i is the measured value of carbon dioxide concentration; T j is the time attribute of neighbor node j; C j is the cost attribute of neighbor node j; Δt ij is the time synchronization error difference between nodes i and j; α T is the time-based weight; α C is the cost basis weight; β T is the time dynamic weight coefficient; β C is the dynamic weight coefficient of cost, t env is the environmental parameter relaxation time; τ d is the fault judgment threshold. In Equation 1-2, through dynamic denominator design and multi-parameter joint detection, α+βΔt dynamically adjusts the difference threshold to avoid misjudgments caused by node migration or clock asynchrony. Temperature, humidity, and carbon dioxide concentration are also monitored simultaneously. The time decay term is introduced to dynamically adjust the difference tolerance to avoid misjudgment of faults due to transient environmental fluctuations. The nonlinear design of the denominator enhances sensitivity to gradual anomalies.

[0085] A multimodal alignment and anonymization formula is constructed based on Equation 2-1. Equation 2-1 includes a multimodal spatiotemporal graph convolutional network for aligning infrared, millimeter-wave radar, and color depth camera data, and protects biometric privacy through adversarial robustness and homomorphic encryption technology.

[0086] Formula 2-1 is expressed as:

[0087]

[0088] Among them, H (l+1) is the feature matrix of the l+1th layer node; σ is the activation function; is the summation operation of multi-scale convolution kernels; T k (Λ) is the Chebyshev polynomial basis function; ⊙ is the Kronecker product operator; G k is the space-time mixing matrix, Aspa is the spatial adjacency matrix, A tem is the time diffusion matrix; H (l) is the feature matrix of the l-th layer node; It is a trainable parameter matrix; it fuses spatiotemporal correlations through Kronecker products, and implements local filtering of image signals using Chebyshev polynomials to solve the spatiotemporal misalignment problem of infrared, millimeter wave, and depth data;

[0089] Construct adversarial perturbation generation based on Equation 2-2;

[0090] Formula 2-2 is expressed as:

[0091]

[0092] in, To minimize the perturbation vector δ, δ represents the perturbation vector, and its dimension is the same as the input data x; To combat the perturbation loss function; is to constrain the perturbation direction along the direction with the lowest model sensitivity; λ·KL(f(x+δ)∥f(x)) is the characteristic distribution after perturbation and the original distribution; J f (x) is the Jacobian matrix of the feature extraction model; KL(.) is the relative entropy divergence; λ is the anti-dynamic strength control parameter; by reasonably controlling the disturbance amplitude δ, the model failure caused by data distortion is prevented; by preferentially destroying the Jacobian matrix element with larger value corresponding to the sensitive feature J f , these features have a significant impact on the classifier output; by balancing privacy and utility λ, when λ is larger, the perturbation intensity is higher, which can achieve better privacy protection, but it may also reduce the availability of data;

[0093] In addition, the air supply priority of high gradient areas is strengthened through exponential weight distribution, and the volume term avoids oversaturation of small areas to achieve Pareto optimal air volume distribution. Based on this, the air volume distribution is optimized through formula 3-1;

[0094] The air volume distribution optimization of formula 3-1 is expressed as:

[0095]

[0096] Among them, w r is the contribution value weight of region r; exp is the exponential function; -η is the contribution sensitivity coefficient, is the volume scaling factor α, controlling V r The intensity of the influence on the weight; V r is the volume of region r; is the wind pressure gradient in region r; is the pollutant concentration gradient in region r; Mapping the quadratic gradient term to a linear scale alleviates the problem of rapid weight decay in high gradient areas; r′∈R exp(·) is global normalization; Formula 3-1 realizes Pareto optimal air volume distribution through exponential weight distribution, exp The wind pressure and pollutant gradient are mapped to weights through a negative exponential function. The high-gradient area has a higher weight, ensuring priority air supply, alleviating the nonlinear growth of the secondary gradient term, and avoiding imbalance in weight distribution caused by extreme gradients. r As a volume scaling factor, by Amplify the influence of large-area gradients and suppress weight oversaturation in small areas due to small volumes. Formula 3-1 is conducive to a gradient-sensitive, volume-suppressed, and globally optimal air volume distribution mechanism.

[0097] By dynamically adjusting the convection heat transfer coefficient, the heat load of the human body and the heat transfer efficiency of the environment are balanced in real time, minimizing the error in thermal comfort prediction;

[0098] The dynamic correction expression of heat transfer coefficient in formula 4-1 is:

[0099]

[0100] in, is the convective heat transfer coefficient of the next time step, reflecting the heat transfer performance of the wall or air conditioning system at time t+1; is the convective heat transfer coefficient at the current moment; ζ is the learning rate attenuation factor; ReLU is the hyperbolic tangent function; is the comfort model for h c The partial derivative of The loss function of the model for predicting the average voting comfort level measures the error between the actual temperature and the predicted temperature; is the convective heat transfer coefficient; tanh(·) is the hyperbolic tangent function, Used to smooth the rate of change of physiological parameters and avoid sudden changes; ΔT skin is the skin temperature gradient, Used to reflect the instantaneous changes in human body surface temperature; is the time derivative of metabolic rate, Represents the dynamic changes of the human body's metabolic heat production rate; by analogy with the heat conduction equation, Where, ΔT skin and ΔQ meta is the source term Q source , the thermal conductivity is updated in real time through the correction term ζ·ReLU(·), simulating the intelligent response of the material thermal properties to changes in the environment; Equation 4-1 realizes thermal comfort modeling by dynamically correcting the heat transfer coefficient, Quantization h c The impact on the prediction error of the comfort model, the larger the error threshold, the higher the correction amplitude, by smoothing ΔTskin and The coupling of h and h can avoid mutation interference; the ReLU function only allows positive gradient correction to prevent h c Abnormal decrease due to noise or transient fluctuations, ζ decreases over time, ensuring that the correction term gradually converges to avoid oscillation; pass Minimize the mean square error between the predicted temperature and the measured temperature Dynamic adjustment h c To improve the accuracy of thermal comfort prediction;

[0101] In federated learning strategy optimization, secure aggregation of distributed encrypted gradients is achieved through homomorphic encryption and dynamic weight distribution of divergence;

[0102] The encrypted gradient aggregation expression of formula 5-1 is:

[0103]

[0104] Among them, g global It is the global encrypted gradient, a shared ciphertext generated by aggregating multi-party encrypted gradients, representing the update direction of the global model; the encrypted state result is the value encrypted by homomorphic encryption; is the client gradient product operation, M is the number of clients participating in federated learning, and the encrypted gradients of each client are superimposed according to the weight to simulate centralized training; Weighted exponentiation of the client-side encrypted gradient, π k is the homomorphically encrypted gradient using end k, θ is the encryption parameter, π k is the dynamic weight coefficient, calculated by divergence; mod N 2 is a modular operation to prevent numerical overflow and maintain addition homomorphism, N is the modulus, N 2 Ensure that the result is within the encryption space; perform composite modulus through the homomorphic encryption cryptographic system to achieve encryption and decryption operations, and the encrypted gradient Multiplication operations can be performed directly without decryption to ensure privacy; the difference between the client's data distribution and the global distribution is measured through divergence dynamic weights, and the weight distribution is adjusted dynamically so that the greater the difference, the smaller the weight, thereby suppressing noise data that deviates from the global distribution; through homomorphic encryption and divergence dynamic weights, data is always transmitted and aggregated in encrypted form to avoid leakage of original gradients. Divergence weights suppress the impact of non-independent and identically distributed data, resolving the contradiction between data privacy and model performance in federated learning.

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for indoor environment monitoring based on thermal comfort analysis, characterized in that: include: Temperature, humidity, and carbon dioxide sensors are deployed in a grid-based manner. By dynamically clustering densely populated areas and migrating nodes, fault self-detection and calibration are performed using Manhattan distance and random sampling consistency techniques. Integrating infrared, millimeter-wave radar, and a color depth camera to collect radiation temperature and physiological data, aligning multimodal information through a spatiotemporal graph convolutional network, and performing biometric anonymization using adversarial perturbation and homomorphic encryption; Utilize anemometers and gas sensors to optimize variable air volume supply parameters in real time, and coordinate multi-zone air volume distribution through blockchain and contribution value calculation air volume distribution algorithms; A thermal comfort model was constructed by integrating environmental and physiological data, and the robustness of the working condition simulation was verified by combining reinforcement learning with dynamic modification of boundary conditions for the heat transfer coefficient. Based on federated learning, human body features are extracted locally and uploaded in encrypted form, linking with the building management system to dynamically adjust the radiation temperature and equipment operation.

2. The indoor environment monitoring method based on thermal comfort analysis according to claim 1, characterized in that: Deploy temperature, humidity, and carbon dioxide sensors in a grid pattern across the area, and record the initial coordinates and measurement data of each node; Identify densely populated areas through dynamic clustering algorithms and calculate the spatial distribution relationship between cluster centers and sensor nodes; Manhattan distance is used to calculate the deviation of measurement values between sensors and screen out faulty nodes. After screening out the faulty nodes, the data of the screened nodes is compared using a random sampling consistency algorithm to determine whether they are abnormal data sources and mark the faulty nodes. Based on the fault mark, the self-test program is executed to determine whether the sensor needs to be automatically calibrated or a maintenance instruction is triggered. For nodes that need calibration, the system starts the automatic calibration process; for nodes that cannot be self-calibrated, the maintenance instruction is triggered for manual intervention, and the node migration strategy is updated based on historical monitoring data.

3. The indoor environment monitoring method based on thermal comfort analysis according to claim 2, characterized in that: Integrate infrared sensors, millimeter-wave radars, and color depth cameras to collect radiation temperature and physiological data, and align the collected radiation temperature and physiological data in time and space to ensure synchronization of sensor data. The aligned data is processed through the spatiotemporal graph convolutional network to extract features and establish data associations.

4. The indoor environment monitoring method based on thermal comfort analysis according to claim 3, characterized in that: Collect data from anemometers and gas sensors, calculate the wind speed change rate and gas concentration gradient in each area, and determine whether the current air volume distribution meets the air supply needs of each area. If not, enter the air volume adjustment process; Calculate the air supply demand weight of each area; determine whether the calculated air volume distribution matches the current variable air volume supply parameters; if not, adjust the variable air volume supply set value; The adjustment records are stored in the blockchain, and the air volume distribution status of each area is updated synchronously; Monitor the wind speed changes after adjustment to determine whether there is an air supply conflict. If a conflict is detected, execute the conflict resolution strategy. If the air volume still cannot meet the demand after adjustment, the air distribution mode will be reconstructed and the global air volume distribution plan will be recalculated.

5. The indoor environment monitoring method based on thermal comfort analysis according to claim 4, characterized in that: Collect environmental and physiological data, calculate temperature gradients, humidity change rates, and human body heat exchange parameters, store them in data sets, and determine data integrity; A thermal comfort model is constructed based on the collected data, and the heat transfer coefficient is calculated to initialize the boundary conditions. A reinforcement learning algorithm is used to optimize the thermal comfort model, and the heat transfer coefficient is dynamically adjusted to adapt to different working conditions. The convergence of the modified boundary conditions is determined. The operating condition simulation is performed in combination with the corrected boundary conditions, and the deviation between the simulation results and the historical data is calculated. If the deviation of the simulation results exceeds the threshold, the parameter weights of the thermal comfort model are adjusted and the simulation verification is re-executed. The simulation results are continuously monitored, the optimization process is recorded, and the robustness of the thermal comfort model under different operating conditions is evaluated.

6. The indoor environment monitoring method based on thermal comfort analysis according to claim 5, characterized in that: Collect local human feature data, extract skin temperature, metabolic rate and posture information, encrypt it, and judge data integrity. If key features are missing or encryption fails, data re-collection or exception handling mechanisms will be triggered; Federated learning is used to train local models, calculate individual thermal comfort parameters, generate optimization strategies, and determine the convergence of local models. Calculate the global radiation temperature adjustment plan, and link the building management system to dynamically adjust the cooling and heating equipment to determine the thermal comfort feedback of the human body after the adjustment.

7. The indoor environment monitoring method based on thermal comfort analysis according to claim 6, characterized in that: A dynamic clustering algorithm is constructed based on Equation 1-1, which includes the node coordinates and real-time population density in the initial sensor network. Based on the population density distribution, the position of the cluster center is dynamically adjusted using exponential decay weights, causing the cluster center to shift toward densely populated areas. By numerically solving the equation system of Equation 1-1, the decay rate is controlled, increasing the contribution of close-range nodes to the migration of the cluster center. The updated cluster center triggers the physical migration of the sensor nodes. Formula 1-1 is expressed as: in, is the new center position of the k-th cluster after t+1 iterations; is the summation operation of all data points in the cluster, N k represents the set of sensor nodes in the kth cluster; ω i is the density weight of data point i, ρ i is the density of people around data point i; x i is the eigenvector of data point i; is an exponential decay function; γ represents the migration attenuation coefficient, R max is the maximum value of the sensor’s sensing radius, To calculate the sample x i With the current cluster center The L1 norm distance and Manhattan distance of ; ∥1 is the Manhattan distance operator; Construct the fault detection robustness criterion based on formula 1-2; Formula 1-2 is expressed as: Among them, D(s i ) is the fault judgment index; is the neighborhood summation operation, N i represents the set of neighboring nodes of node i; T i is the temperature and humidity of node i; C i is the measured value of carbon dioxide concentration; T j is the time attribute of neighbor node j; C j is the cost attribute of neighbor node j; Δt ij is the time synchronization error difference between nodes i and j; α T is the time-based weight; α C is the cost basis weight; β T is the time dynamic weight coefficient; β C is the dynamic weight coefficient of cost; τ d is the fault judgment threshold; in formula 1-2, dynamic denominator design and multi-parameter joint detection are used.

8. The indoor environment monitoring method based on thermal comfort analysis according to claim 7, characterized in that: A multimodal alignment and anonymization formula is constructed based on Equation 2-1. Equation 2-1 includes a multimodal spatiotemporal graph convolutional network for aligning infrared, millimeter-wave radar, and color depth camera data, and protects biometric privacy through adversarial robustness and homomorphic encryption technology. Formula 2-1 is expressed as: Among them, H (l+1) is the feature matrix of the l+1th layer node; σ is the activation function; is the summation operation of multi-scale convolution kernels; T k (Λ) is the Chebyshev polynomial basis function; ⊙ is the Kronecker product operator; G k is the space-time mixing matrix; H (l) is the feature matrix of the l-th layer node; is the trainable parameter matrix; Construct adversarial perturbation generation based on Equation 2-2; Formula 2-2 is expressed as: in, To minimize the perturbation vector δ, δ represents the perturbation vector, and its dimension is the same as the input data x; To combat the perturbation loss function; is to constrain the perturbation direction along the direction with the lowest model sensitivity; λ·KL(f(x+δ)∥f(x)) is the characteristic distribution after perturbation and the original distribution; J f (x) is the Jacobian matrix of the feature extraction model; KL(.) is the relative entropy divergence; λ is the anti-dynamic strength control parameter.

9. The indoor environment monitoring method based on thermal comfort analysis according to claim 8, characterized in that: The air supply priority of high gradient areas is strengthened through exponential weight distribution, and the volume term avoids oversaturation of small areas to achieve Pareto optimal air volume distribution. Based on this, the air volume distribution is optimized through formula 3-1; The air volume distribution optimization of formula 3-1 is expressed as: Among them, w r is the contribution value weight of region r; exp is the exponential function; -η is the contribution sensitivity coefficient; is the α volume scaling factor; is the convective heat transfer coefficient; V r is the volume of region r; is the wind pressure gradient in region r; is the pollutant concentration gradient in region r; is the mapping of the quadratic gradient term to a linear scale; r′∈R exp(·) is global normalization; Formula 3-1 realizes Pareto optimal air volume distribution through exponential weight distribution.

10. The indoor environment monitoring method based on thermal comfort analysis according to claim 9, characterized in that: By dynamically adjusting the convection heat transfer coefficient, the heat load of the human body and the heat transfer efficiency of the environment are balanced in real time, minimizing the error in thermal comfort prediction; The dynamic correction expression of heat transfer coefficient in formula 4-1 is: in, is the convective heat transfer coefficient of the next time step; is the convective heat transfer coefficient at the current moment; ζ is the learning rate attenuation factor; ReLU is the hyperbolic tangent function; is the comfort model for h c The partial derivative of is the loss function of the model for predicting average voting comfort; tanh(·) is the hyperbolic tangent function; ΔT skin is the skin temperature gradient; is the time derivative of metabolic rate; Equation 4-1 realizes thermal comfort modeling by dynamically correcting the heat transfer coefficient; In federated learning strategy optimization, secure aggregation of distributed encrypted gradients is achieved through homomorphic encryption and dynamic weight distribution of divergence; The encrypted gradient aggregation expression of formula 5-1 is: Among them, g global Encrypt gradients globally; is the client gradient product operation, M is the number of clients participating in federated learning; (g k θ ) πk Weighted exponentiation of the client-side encrypted gradient, π k is the homomorphically encrypted gradient using end k, θ is the encryption parameter, π k is the dynamic weight coefficient; mod N 2 It is a modular operation.

Images