Building indoor environment adaptive regulation system and method based on face feature depth analysis and multi-source weighting

Abstract

The application discloses a building indoor environment adaptive regulation and control system and method based on face feature depth analysis and multi-source weighting, relates to the field of adaptive regulation and control, and comprises the following modules: an indoor environment perception and monitoring module, which is used for collecting indoor environment data; a face feature collection and analysis module, which is used for collecting face analysis data; a basic model decision matching module, which is used for collecting weighted multi-source fusion features, matching models in a model library, and obtaining a basic decision model; a weighted data regulation and control instruction module, which is used for constructing a data weighted fusion model, generating a building indoor environment adaptive regulation and control instruction, and adjusting the opening degree, power and mode of equipment; and a running feedback iteration module, which is used for collecting feedback data and optimizing a scenario decision model library. The application identifies the physiological state and thermal comfort demand of a person by constructing a building indoor environment adaptive regulation and control system, and improves the comfort of indoor environment regulation and control.

Description

Technical Field

[0001] This invention relates to the field of adaptive control, and more specifically, to an adaptive control system and method for building indoor environments based on deep facial feature analysis and multi-source weighting. Background Technology

[0002] The building indoor environment adaptive control system based on deep facial feature analysis and multi-source weighting serves as the core support carrier for building intelligent upgrading and personalized optimization of the indoor environment. Its environmental perception accuracy, personnel demand recognition capability, model matching and adaptability, control response efficiency, and closed-loop iteration performance directly determine the comfort level, energy utilization efficiency, and personnel experience of the building indoor environment.

[0003] The diversification of building interior scenarios, the differences in people's physiological states and thermal comfort needs, the heterogeneous coupling of multi-source indoor environmental data and personnel characteristic data, the difficulty in coordinating control precision and energy-saving goals, and the lack of standardization of control processes are the core constraints affecting the efficient implementation and stable operation of building indoor environment adaptive control systems. These systems are dynamically affected by multiple factors such as dynamic fluctuations in indoor environmental parameters, changes in personnel identity and physiological states, differences in control needs in different scenarios, inconsistent dimensions of multi-source data, and model adaptation deviations. At the same time, the deep analysis of facial features and the weighted fusion control technology for multi-source data in building interior scenarios are key means to solve the problems of low accuracy, insufficient personalized adaptation, and serious energy waste in traditional indoor environment control, and directly ensure the accurate matching of indoor environment control with people's physiological states and thermal comfort needs.

[0004] However, existing building indoor environment control systems and methods often rely on fixed threshold control or single environmental parameter feedback modes in practical applications. They lack the ability to deeply analyze facial and behavioral features, ignore the value of weighted fusion of multi-source data, and lack accurate matching and real-time verification mechanisms for scenario-based decision-making models. They cannot cover the personalized thermal comfort needs of people in different indoor scenarios and physiological states. This results in a one-sided approach to control systems, emphasizing unified control over personalized adaptation and environmental parameter collection over human needs identification. At the same time, existing methods often adopt static control strategies or single model configuration modes, failing to solve the problem of multi-stage technical coordination. This leads to low adaptability between environmental control and human needs, loose correlation of data features, insufficient control accuracy, and serious ineffective energy consumption. The process requires manual intervention to set parameters and repeated debugging, increasing the deployment difficulty and operating cost of the control system, and failing to achieve dynamic adaptive control of the indoor environment.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes an adaptive control system and method for building indoor environments based on deep facial feature analysis and multi-source weighting, in order to overcome the aforementioned technical problems existing in the current related technologies.

[0007] To achieve the above objectives, the specific technical solution adopted by the present invention is as follows: According to one aspect of the present invention, an adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting is provided. The system includes: an indoor environment perception and monitoring module, a facial feature acquisition and analysis module, a basic model decision matching module, a weighted data control instruction module, an equipment adaptive control module, and an operation feedback iteration module. The indoor environment sensing and monitoring module is used to deploy multi-dimensional environmental sensors to collect data on indoor temperature, humidity, light, air quality and equipment operation status in real time, and summarize them to form indoor environment data. The facial feature acquisition and analysis module is used to collect facial image data and behavioral feature data of people inside the building, and extract and integrate physiological state parameters, thermal comfort requirement parameters and identity feature information of people to form facial analysis data; As a preferred embodiment, the face feature acquisition and analysis module includes: a face image acquisition module, a behavior feature capture module, a physiological state analysis module, and a comfort requirement output module; The face image acquisition module is used to acquire raw face image data of people inside the building in real time through indoor camera equipment, and to perform image noise reduction and normalization preprocessing. The behavioral feature capture module is used to capture data on people's body movements, spatial location distribution, and facial expression changes through indoor camera equipment as behavioral feature data, and bind the corresponding person's identity feature information. The physiological state analysis module is used to perform deep analysis of facial image data and behavioral feature data based on a deep learning model to obtain physiological state parameters and thermal comfort requirement parameters of personnel. As a preferred embodiment, the physiological state analysis module includes: a facial key feature extraction module, a behavioral feature quantification module, a multi-feature fusion analysis module, and a physiological comfort output module; The facial key feature extraction module is used to extract features from facial image data based on a deep convolutional neural network model, and obtain facial key features including facial key point coordinates, skin color heat distribution features, and eye opening and closing status. The behavioral feature quantification module is used to quantify the range of human body movements, spatial distribution density, and range of facial expression changes in behavioral feature data to obtain behavioral quantification features including limb extension, human position distribution, and facial pleasure. The multi-feature fusion and parsing module is used to fuse key facial features with behavioral quantitative features and perform correlation analysis using a convolutional long short-term memory neural network model to obtain physiological state parameters of the person, including body temperature range and fatigue level classification. As a preferred embodiment, the multi-feature fusion parsing module includes: a feature normalization processing unit, a weighted feature fusion unit, a temporal correlation analysis unit, and a physiological state output unit; The feature normalization processing unit is used to perform dimensionless and standardized processing on key facial features and quantitative behavioral features, respectively, and to eliminate dimensional differences and numerical deviations between features of different dimensions. The weighted feature fusion unit is used to assign corresponding fusion weights to the standardized facial key features and behavior quantification features, and to perform weighted superposition operation to obtain a high-dimensional fusion feature vector. The temporal correlation analysis unit is used to input high-dimensional fused feature vectors into a convolutional long short-term memory neural network model to perform in-depth mining and correlation reasoning on the feature change patterns in the temporal dimension. The physiological state output unit is used to determine the physiological state parameters of personnel, including body temperature range and fatigue level, based on the time-series correlation analysis results and a preset classification threshold.

[0008] The physiological comfort output module is used to match personnel physiological state parameters with identity feature information to generate personnel thermal comfort requirement parameters including preferred temperature range, suitable humidity range and light tolerance level.

[0009] The comfort requirement output module is used to standardize and encode personnel physiological state parameters, thermal comfort requirement parameters and identity feature information, and establish an association mapping according to the identity feature information to form face parsing data containing personnel identity identifiers, physiological state parameters and thermal comfort requirement parameters.

[0010] The basic model decision matching module is used to preset the scenario-based decision model library and data weight ratio, allocate weights to indoor environmental data and face analysis data according to the data weight ratio, form weighted multi-source fusion features, and match the weighted multi-source fusion features with the models in the scenario-based decision model library to obtain the basic decision model. As a preferred embodiment, the basic model decision matching module includes: a weight parameter preset module, a model library management module, a data weight allocation module, and a decision model matching module; The weight parameter preset module is used to preset the weight allocation benchmark parameters of indoor environmental data and face analysis data to form a data weight ratio system. The model library management module is used to store control decision models corresponding to different indoor scenarios and build a scenario-based decision model library. The data weight allocation module is used to perform weight allocation processing on indoor environmental data and face analysis data according to a standardized data weight ratio system, so as to form a weighted multi-source fusion feature. The decision model matching module is used to perform feature matching and fit calculation between the weighted multi-source fusion features and the models in the scenario-based decision model library, and select the decision model with the highest fit as the basic decision model.

[0011] As a preferred embodiment, the decision model matching module includes: a feature preprocessing module, a fitness calculation module, a model screening module, and a model validation module; The feature preprocessing module is used to standardize and normalize the weighted multi-source fusion features output by the data weight allocation module, and remove redundant features and abnormal data to obtain standardized fusion features. The fitness calculation module is used to calculate the feature similarity between the standardized fusion features and each regulation decision model in the scenario-based decision model library using the cosine similarity algorithm, so as to obtain the fitness value corresponding to each model. The model screening module is used to preset the fit screening threshold and screen the control decision model based on the fit screening threshold. The model verification module is used to verify the selected basic decision-making models in real time and confirm the adaptability of the model output results to the current indoor scene and personnel needs. If the validation passes, the model will be output as the base decision model. If the verification fails, return to the model filtering module for re-filtering.

[0012] The weighted data control instruction module is used to input indoor environmental data and facial recognition data into the basic decision model, construct a data weighted fusion model, preset comfort thresholds and scene control strategies, and then combine the data weighted fusion model to generate adaptive control instructions for the building's indoor environment. As a preferred embodiment, the weighted data control instruction module includes: a fusion model construction module, a comfort threshold configuration module, a scenario strategy matching module, and a control instruction generation module; The fusion model construction module is used to input indoor environmental data and facial recognition data into the basic decision model, perform multi-source data weighted calculation and model parameter initialization, and obtain a data weighted fusion model. The comfort threshold configuration module is used to preset the comfort thresholds for temperature, humidity, light, and air quality for different personnel identities, physiological states, and indoor scenes, forming a standardized comfort threshold library. The scene strategy matching module is used to preset a scene control strategy library and match indoor environmental data and face parsing data with the scene control strategy library; The control instruction generation module is used to calculate target control parameters based on a data weighted fusion model, combined with a standardized comfort threshold library and matching scenario control strategies, and generate adaptive control instructions for the building indoor environment that include adjustment parameters for air conditioning, fresh air, and lighting equipment.

[0013] The device adaptive control module is used to adjust the opening degree, power and mode of indoor environment control equipment such as air conditioner, fresh air and lighting according to the building indoor environment adaptive control command; The operation feedback iteration module is used to collect indoor environmental operation data and personnel comfort feedback data after equipment adjustment, and to compare and analyze the control effect to optimize the scenario-based decision model library.

[0014] As a preferred embodiment, the operation feedback iteration module includes: an operation data acquisition module, a comfort feedback receiving module, a regulation effect analysis module, and a fusion model optimization module; The operation data acquisition module is used to collect indoor environmental operation data such as indoor temperature, humidity, light, air quality and equipment operation status in real time after the equipment is adjusted. The comfort feedback receiving module is used to acquire subjective comfort feedback data of people in the building interior and form a personnel comfort evaluation dataset. The regulation effect analysis module is used to compare indoor environmental operation data, personnel comfort evaluation dataset, and standardized comfort threshold library to perform quantitative analysis on the accuracy of environmental regulation and comfort satisfaction, and obtain regulation effect evaluation results. As a preferred embodiment, the regulation effect analysis module includes: a data comparison and statistics module and a comprehensive accuracy evaluation module; The data comparison and statistics module is used to compare the indoor environmental operation data with the standardized comfort threshold library for deviation, and to perform positive and negative feedback classification statistics on the personnel comfort evaluation dataset to obtain the environmental parameter deviation value and the comfort feedback ratio. The comprehensive accuracy assessment module is used to perform quantitative calculations and comprehensive judgments based on environmental parameter deviation values ​​and the proportion of comfort feedback, and to generate a control effect assessment result that includes control accuracy, comfort satisfaction, and effect level.

[0015] The fusion model optimization module is used to iteratively optimize the weight parameters and calculation rules of the data weighted fusion model based on the evaluation results of the regulation effect, and synchronously update the optimized model to the scenario-based decision model library.

[0016] According to another aspect of the present invention, an adaptive control method for building indoor environment based on deep facial feature analysis and multi-source weighting is provided, the method comprising the following steps: S1. Deploy multi-dimensional environmental sensors to collect real-time data on indoor temperature, humidity, light, air quality, and equipment operating status, and summarize them to form indoor environmental data; S2. Collect facial image data and behavioral feature data of people inside the building, and extract and integrate physiological state parameters, thermal comfort requirement parameters and identity feature information of people to form facial analysis data; S3. Preset scenario-based decision model library and data weight ratio. Based on the data weight ratio, assign weights to indoor environmental data and face analysis data to form weighted multi-source fusion features. Then match the weighted multi-source fusion features with the models in the scenario-based decision model library to obtain the basic decision model. S4. Input indoor environmental data and facial recognition data into the basic decision model, construct a data weighted fusion model, preset comfort thresholds and scene control strategies, and then combine the data weighted fusion model to generate adaptive control instructions for the building's indoor environment. S5. Based on the building's indoor environment adaptive control instructions, adjust the opening degree, power and mode of indoor environment control equipment such as air conditioning, fresh air, and lighting; S6. Collect indoor environmental operation data and personnel comfort feedback data after equipment adjustment, compare and analyze the control effect, and optimize the scenario-based decision-making model library.

[0017] The beneficial effects of this invention are as follows: 1. This invention constructs an adaptive control system for building indoor environments that integrates multi-dimensional perception of the indoor environment, deep analysis of facial and behavioral features, scenario-based model matching, weighted fusion of multi-source data, adaptive control of terminal devices, and closed-loop iteration of operational feedback. Combining deep learning feature extraction, temporal feature correlation analysis, cosine similarity model matching, and quantitative evaluation iteration mechanism, it identifies the physiological state and thermal comfort needs of personnel, achieving efficient matching and dynamic adaptation of indoor environmental parameters and personalized needs of personnel, thereby improving the accuracy and comfort of indoor environmental control. At the same time, through multi-source data weighted calculation, standardized comfort threshold configuration, one-click matching of scenario-based strategies, and automatic generation of control instructions, it simplifies the complex logic and manual intervention process of traditional indoor environmental control, reducing the deployment and operation costs of the control system.

[0018] 2. This invention deploys multi-dimensional environmental sensors through an indoor environment sensing and monitoring module to achieve real-time collection and aggregation of data on temperature, humidity, light intensity, air quality, and equipment operating status. Combined with image preprocessing, feature extraction, and quantization fusion processes by the facial feature acquisition and analysis module, it ensures the integrity, accuracy, and standardization of multi-source data. Through the weight parameter preset, model library management, feature preprocessing, and adaptation calculation of the basic model decision matching module, it achieves accurate matching and real-time verification of the scenario-based decision model, avoiding control failures caused by model adaptation deviations. At the same time, through the multi-source data weighted calculation, standardized comfort threshold configuration, scenario strategy matching, and automatic generation of control instructions by the weighted data control instruction module, it achieves precise control of the opening, power, and mode of terminal devices such as air conditioners, fresh air systems, and lighting. This simplifies the complex process of manually setting parameters and repeatedly debugging in traditional indoor environment control, reducing the deployment difficulty and operating cost of the control system. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a system block diagram of an adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting, according to an embodiment of the present invention. Figure 2 This is a flowchart of a method for adaptive control of building indoor environment based on deep facial feature analysis and multi-source weighting, according to an embodiment of the present invention.

[0021] In the picture: 1. Indoor environment perception and monitoring module; 2. Facial feature acquisition and analysis module; 3. Basic model decision matching module; 4. Weighted data control instruction module; 5. Equipment adaptive control module; 6. Operational feedback and iteration module. Detailed Implementation

[0022] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to one aspect of the present invention, an adaptive control system for building indoor environment based on deep analysis of facial features and multi-source weighting is provided. The system includes: an indoor environment perception and monitoring module 1, a facial feature acquisition and analysis module 2, a basic model decision matching module 3, a weighted data control instruction module 4, an equipment adaptive control module 5, and an operation feedback iteration module 6. The indoor environment sensing and monitoring module 1 is used to deploy multi-dimensional environmental sensors to collect data on indoor temperature, humidity, light, air quality and equipment operation status in real time, and summarize them to form indoor environment data. Specifically, multi-dimensional environmental sensors need to be scientifically deployed in conjunction with the building's indoor functional zoning, spatial structure, and human activity patterns, following the principles of full coverage, key area control, and interference-resistant deployment to achieve real-time collection of all elements of the indoor environment and equipment status. Temperature and humidity sensors should be evenly deployed in various functional areas such as offices, meeting rooms, corridors, and rest areas, installed in the middle of the walls, away from interference sources such as air conditioning vents, heating equipment, and strong direct sunlight, to ensure accurate temperature and humidity data. Light sensors should be deployed on the indoor ceiling and near windows to simultaneously collect data on natural and artificial lighting. Air quality sensors should be deployed in densely populated areas and poorly ventilated areas to monitor key indicators such as PM2.5, CO2, and VOCs in real time.

[0025] Simultaneously, operational status acquisition units are installed at terminal control devices such as air conditioners, fresh air systems, and lighting, directly connecting to the device ports to acquire real-time operational parameters such as device start / stop, power, and opening degree. Each sensor and acquisition unit synchronously uploads raw data at a fixed frequency via wired or wireless communication modules. After preprocessing the data, such as noise reduction and outlier removal, the system classifies, integrates, and uniformly encodes the data according to monitoring points, parameter types, and spatial areas, ultimately summarizing it into standardized indoor environmental data that includes real-time environmental parameters, device operating status, and an overall environmental overview.

[0026] The facial feature acquisition and analysis module 2 is used to collect facial image data and behavioral feature data of people inside the building, and extract and integrate physiological state parameters, thermal comfort requirement parameters and identity feature information of people to form facial analysis data; In this embodiment of the invention, the face feature acquisition and analysis module 2 includes: a face image acquisition module, a behavior feature capture module, a physiological state analysis module, and a comfort requirement output module; The face image acquisition module is used to acquire raw face image data of people inside the building in real time through indoor camera equipment, and to perform image noise reduction and normalization preprocessing. Specifically, firstly, high-definition indoor camera equipment is scientifically deployed in various functional areas of the building to build a full-area unobstructed facial image acquisition system. Combining the indoor space structure, personnel activity trajectories and shooting coverage requirements, the equipment is installed in non-interference locations such as walls and ceilings, avoiding direct sunlight, equipment interference, and blind spots caused by personnel obstruction. The indoor images are continuously captured at a fixed resolution and frame rate, and the original facial image data of personnel is captured and cached in real time to ensure the integrity and real-time nature of the collected data.

[0027] Subsequently, image noise reduction processing was carried out. To address interference information such as Gaussian noise and salt-and-pepper noise generated during the acquisition process due to light fluctuations and equipment operation, a composite algorithm combining Gaussian filtering and median filtering was used to process the original image frame by frame, eliminating invalid noise data, retaining core effective features such as facial key points and skin texture, improving image clarity and usability, and eliminating the interference of noise on subsequent analysis.

[0028] Finally, image normalization preprocessing is performed to uniformly adjust the denoised images to the model's fit size. A brightness equalization algorithm is used to eliminate brightness deviations caused by uneven indoor lighting. This completes precise cropping of the face region, pose correction, and spatial alignment, resolving the issue of inconsistent image specifications caused by differences in shooting angle and distance. The processed output is standardized face image data with uniform format and regular features, eliminating the blurriness and differences of the original image.

[0029] The behavioral feature capture module is used to capture data on people's body movements, spatial location distribution, and facial expression changes through indoor camera equipment as behavioral feature data, and bind the corresponding person's identity feature information. Specifically, firstly, high-definition indoor camera equipment is scientifically deployed in various functional areas of the building according to spatial layout and human activity patterns to construct a real-time visual acquisition network with no blind spots. This eliminates acquisition limitations such as wall obstruction and blind spots, and captures in real time data on the range of motion of people's limbs, limb postures, and movement trajectories, as well as spatial distribution data such as the area where people are located, their gathering density, and spatial points, and facial expression data such as changes in facial muscles, emotional expression, and changes in the intensity of facial expressions. The above multi-dimensional information is integrated to form raw behavioral feature data.

[0030] Subsequently, the raw behavioral feature data is precisely preprocessed. Based on target detection, multi-target tracking and posture estimation algorithms, different people indoors are individually distinguished and continuously tracked. Redundant noise information such as environmental interference, non-person targets and invalid frames are removed, and effective, regular and representative behavioral features are extracted to ensure the accuracy and usability of the data.

[0031] Finally, by combining the personnel identification information output by the facial feature acquisition and analysis module, the corresponding body movements, spatial location distribution, and facial expression change data of each person are accurately associated and bound with the unique identity feature information, establishing a one-to-one mapping relationship between behavioral features and identity information, completing the standardized encoding and structured storage of the data, and forming standardized behavioral feature data with identity tags.

[0032] The physiological state analysis module is used to perform deep analysis of facial image data and behavioral feature data based on a deep learning model to obtain physiological state parameters and thermal comfort requirement parameters of personnel. In this embodiment of the invention, the physiological state analysis module includes: a facial key feature extraction module, a behavioral feature quantification module, a multi-feature fusion analysis module, and a physiological comfort output module; The facial key feature extraction module is used to extract features from facial image data based on a deep convolutional neural network model, and obtain facial key features including facial key point coordinates, skin color heat distribution features, and eye opening and closing status. Specifically, the standardized face image data, after noise reduction and normalization preprocessing, is first input into a pre-trained and optimized deep convolutional neural network model to complete the tensor transformation of the image data and the initialization of the model's forward inference, providing a standardized and unified input basis for face feature extraction.

[0033] Subsequently, relying on the multi-layer convolution, pooling, and feature activation operation mechanism of deep convolutional neural networks, progressive feature mining from shallow to deep layers is carried out on facial images. The coordinate information of key organs such as eyes, eyebrows, nose bridge, and lips is accurately located and extracted through the facial key point detection branch. The facial region pixels are analyzed through the skin color thermal distribution analysis branch to obtain skin color thermal distribution and skin color uniformity features. The eye state recognition branch determines the eyelid and eye fissure areas to obtain standardized eye opening and closing state parameters.

[0034] Finally, the extracted multi-dimensional features are normalized, outlier removed, and structured and integrated to eliminate dimensional differences and numerical deviations between different features, forming complete facial key feature data including facial key point coordinates, skin color thermal distribution features, and eye opening and closing status.

[0035] The behavioral feature quantification module is used to quantify the range of human body movements, spatial distribution density, and range of facial expression changes in behavioral feature data to obtain behavioral quantification features including limb extension, human position distribution, and facial pleasure. Specifically, the collected and preprocessed human behavior feature data is used as the basis for quantitative processing. The data on the amplitude of limb movements, spatial distribution density, and amplitude of facial expression changes are frame aligned, interference is removed, and target tracking is calibrated to eliminate invalid information such as environmental occlusion, image jitter, and non-target individuals, thus ensuring the accuracy and stability of the quantitative data source.

[0036] Subsequently, multi-dimensional behavioral feature quantification calculations were carried out. For the range of limb movements, the displacement and angle changes of joint points were extracted through human posture key point detection algorithms, and the range of movements was normalized and mapped to form a standardized limb extension index. For spatial distribution density, based on indoor spatial grid division and personnel coordinate positioning, the degree of personnel gathering and spatial point distribution in each area were calculated, and the personnel location distribution data were quantified. For the range of facial expression changes, the degree of deformation and range of motion of facial muscle key points were analyzed, and the intensity of expression was judged according to preset grading rules to generate an expression pleasure index.

[0037] Finally, the three types of quantitative results are subjected to dimensionless and standardized processing to eliminate numerical differences and dimensional biases of different dimensions, complete data regularization and structured integration, and finally output behavioral quantitative features including limb extension, personnel location distribution and facial pleasure.

[0038] The multi-feature fusion and parsing module is used to fuse key facial features with behavioral quantitative features and perform correlation analysis using a convolutional long short-term memory neural network model to obtain physiological state parameters of the person, including body temperature range and fatigue level classification. In this embodiment of the invention, the multi-feature fusion parsing module includes: a feature normalization processing unit, a weighted feature fusion unit, a temporal correlation analysis unit, and a physiological state output unit; The feature normalization processing unit is used to perform dimensionless and standardized processing on key facial features and quantitative behavioral features, respectively, and to eliminate dimensional differences and numerical deviations between features of different dimensions. Specifically, the extracted key facial features and quantitative behavioral features are first processed. The two types of multi-dimensional feature data are pre-cleaned and validated to remove outliers, missing fields and noisy data, ensuring that the input data is complete and reliable, and providing a stable data foundation for subsequent dimensionless and standardized processing.

[0039] Subsequently, the key facial features and behavioral quantitative features were processed to be dimensionless. For facial features such as facial key point coordinates, skin color heat distribution, and eye opening and closing, as well as behavioral features such as limb extension, personnel position distribution, and expression pleasure, extreme value normalization and Z-score standardization were combined to map feature data of different units, different magnitudes, and different numerical ranges to a unified standard range. This broke down the computational barriers caused by various physical dimensions such as coordinates, scale, and density, and weakened the problems of magnitude differences and uneven distribution of the original data.

[0040] Finally, the dimensionless feature data is standardized and regulated. Through mean correction, variance alignment and numerical offset calibration, the data dispersion bias is further corrected, the feature distribution form and numerical rules are unified, and the dimensional differences and numerical biases between features of different dimensions are completely eliminated.

[0041] The weighted feature fusion unit is used to assign corresponding fusion weights to the standardized facial key features and behavior quantification features, and to perform weighted superposition operation to obtain a high-dimensional fusion feature vector. Specifically, the core processing objects are the standardized facial key features and behavioral quantitative features. The feature importance assessment mechanism is used to comprehensively consider the impact weight of the two types of features on the analysis of people's physiological state and the determination of thermal comfort needs, and to assign corresponding fusion weights to the two types of features and their subordinate sub-features.

[0042] Among the key facial features, the coordinates of facial key points and the heat distribution of skin color, as well as the limb extension and facial expression pleasure among the behavioral quantification features, are assigned higher weights because they are more closely related to people's physiological state and comfort needs. Secondary dimension features are assigned relatively lower weights. At the same time, the weight parameters are dynamically fine-tuned based on the model training feedback to ensure that the weight allocation is scientific and reasonable and fits the actual analysis needs.

[0043] Subsequently, weighted operations are performed on the two types of standardized features respectively. The value of each sub-feature is multiplied by the corresponding fusion weight to obtain the weighted feature value of each sub-feature. Then, all weighted feature values ​​are superimposed and integrated dimension by dimension and arranged in the order of preset feature dimensions to complete the aggregation processing of high-dimensional features. Finally, a high-dimensional fusion feature vector containing two core feature information of face and behavior, with unified dimensions and balanced values ​​is generated.

[0044] The temporal correlation analysis unit is used to input high-dimensional fused feature vectors into a convolutional long short-term memory neural network model to perform in-depth mining and correlation reasoning on the feature change patterns in the temporal dimension. Specifically, the high-dimensional fused feature vectors after weighted superposition are first normalized and preprocessed according to the time sequence. The feature vectors are then sorted according to the time sequence, and time sequence labeling information is added. At the same time, abnormal feature points in the time sequence dimension are removed and missing data is filled in to ensure the temporal continuity and integrity of the input data and to adapt to the input requirements of the convolutional long short-term memory neural network model.

[0045] Subsequently, the normalized temporal high-dimensional fused feature vector is input into a pre-trained and optimized convolutional long short-term memory neural network model. Relying on the dual characteristics of convolutional operation and long short-term memory, the high-dimensional features of each temporal node are first extracted locally through the convolutional layer, preserving the core correlation information between key facial features and behavioral quantitative features. Then, through the gating mechanism (input gate, forget gate, output gate), the feature changes in the temporal dimension are dynamically memorized and filtered to deeply explore the evolution of features over time, such as the temporal change trends of skin color heat distribution, facial expression pleasure, and limb extension.

[0046] Finally, through the model's correlation reasoning module, the features of different time-series nodes are analyzed to uncover the inherent temporal correlation between facial features and behavioral features, such as the temporal correlation between changes in eye opening and closing state and fatigue level, and the dynamic correlation between limb extension and thermal comfort needs. This completes the in-depth mining and correlation reasoning of temporal features and outputs feature analysis results containing temporal correlation information.

[0047] The physiological state output unit is used to determine the physiological state parameters of personnel, including body temperature range and fatigue level, based on the time-series correlation analysis results and a preset classification threshold.

[0048] Specifically, the analysis results are first based on the temporal correlation analysis output of the convolutional long short-term memory neural network model. The analysis results are then structured and organized to extract core temporal features related to body temperature and fatigue levels. Irrelevant features and abnormal analysis results are removed, and feature association annotations are added to ensure that the data used for judgment is accurate and meets the needs of physiological state analysis.

[0049] Then, a preset classification threshold system is invoked. This system combines physiological and medical standards with the needs of indoor building environment control scenarios. Pre-set thresholds for judging body temperature range and fatigue level classification are set. The body temperature range is divided into normal, high and low corresponding value thresholds. The fatigue level classification is set into mild, moderate and severe, with time-series characteristic thresholds such as eye opening and closing frequency and facial expression change amplitude. The threshold parameters can be dynamically fine-tuned according to the actual scenario.

[0050] Next, the normalized temporal correlation features are compared with the preset classification thresholds one by one. Through threshold interval matching, the temporal change features of skin color heat distribution are mapped to the corresponding body temperature intervals. The temporal change features such as eye opening and closing state and limb extension are judged as the corresponding fatigue level grades. The judgment results are then structured and integrated to clearly mark the personnel identification, body temperature range, fatigue level and judgment basis, forming standardized and structured personnel physiological state parameters.

[0051] The physiological comfort output module is used to match personnel physiological state parameters with identity feature information to generate personnel thermal comfort requirement parameters including preferred temperature range, suitable humidity range and light tolerance level.

[0052] Specifically, the core processing objects are structured personnel physiological state parameters and personnel identity feature information. The consistency of the two types of data is first checked and standardized, and invalid data with missing identity identifiers or abnormal physiological parameters are removed to ensure that the identity feature information (unique identifier) ​​corresponds one-to-one with the physiological state parameters (body temperature range, fatigue level classification), laying the foundation for subsequent matching work.

[0053] Subsequently, a mechanism for linking physiological states with identity characteristics was established. Each individual's identity identifier was used as an index, binding it to their corresponding physiological state parameters. Combined with pre-defined comfort requirement matching rules, and referencing comfort requirement benchmarks corresponding to different identities (e.g., office workers, visitors) and different physiological states (e.g., fatigue levels, body temperature ranges), personalized matching was performed. For example, individuals with higher fatigue levels were matched with a lower preferred temperature range; individuals with higher body temperatures were matched with a suitable low humidity range. The appropriate light tolerance level was determined based on the daily activity scenarios corresponding to each identity. Finally, the matching results were structured, integrated, and standardized, clearly labeling the individual's identity identifier, preferred temperature range, suitable humidity range, and light tolerance level, generating standardized thermal comfort requirement parameters for each individual.

[0054] The comfort requirement output module is used to standardize and encode personnel physiological state parameters, thermal comfort requirement parameters and identity feature information, and establish an association mapping according to the identity feature information to form face parsing data containing personnel identity identifiers, physiological state parameters and thermal comfort requirement parameters.

[0055] Specifically, the core processing objects are personnel physiological state parameters, thermal comfort requirements parameters, and identity characteristics information. Data pre-verification and standardization are carried out first to eliminate invalid data with missing identity identifiers, abnormal parameters, or inconsistent formats, ensuring that the three types of data are completely corresponding and the information is accurate, thus laying the foundation for standardized coding and association mapping.

[0056] Subsequently, standardized coding processing is carried out, using unified coding rules to assign a unique standardized identity identifier to the identity feature information, and classifying and coding physiological state parameters (body temperature range, fatigue level classification) and thermal comfort demand parameters (preferred temperature range, suitable humidity range, light tolerance level) separately. The text descriptions and numerical ranges are converted into a standardized and unified coding format, eliminating differences in different expressions and format deviations, and ensuring that the coding of various parameters is standardized, identifiable, and calculable.

[0057] Next, using standardized identity identifiers as the core index, an association mapping mechanism is established to accurately bind each identity identifier to its corresponding encoded physiological state parameters and thermal comfort requirement parameters, constructing a one-to-one mapping relationship to ensure the uniqueness and accuracy of data association. Finally, the associated encoded data is structured and integrated, sorted according to a preset format, and clearly includes complete information containing personnel identity identifiers, physiological state parameters, and thermal comfort requirement parameters, forming standardized and structured face parsing data.

[0058] The basic model decision matching module 3 is used to preset the scenario-based decision model library and data weight ratio, allocate weights to indoor environmental data and face parsing data according to the data weight ratio, form weighted multi-source fusion features, and match the weighted multi-source fusion features with the models in the scenario-based decision model library to obtain the basic decision model. In this embodiment of the invention, the basic model decision matching module 3 includes: a weight parameter preset module, a model library management module, a data weight allocation module, and a decision model matching module; The weight parameter preset module is used to preset the weight allocation benchmark parameters of indoor environmental data and face analysis data to form a data weight ratio system. Specifically, firstly, based on the core requirements of the building's indoor environment adaptive control system, and clarifying the core roles of indoor environmental data and facial recognition data, a weight allocation benchmark parameter is preset, and a complete weight ratio system is constructed. The core dimensions of the two types of data are first identified: indoor environmental data covers environmental parameters such as temperature and humidity, as well as equipment operating status; facial recognition data includes information such as personnel identity, physiological state, and thermal comfort needs. Combining control accuracy requirements and personnel comfort priorities, and referencing relevant industry standards and historical system operating data, the basic weight benchmarks for the two types of data are determined.

[0059] Subsequently, for the sub-dimensions of the two types of data, sub-benchmark parameters were preset according to their importance. Among them, the sub-dimension of facial analysis data closely related to human comfort and physiological state (such as thermal comfort needs and fatigue level) was assigned a higher benchmark weight, the core parameters of indoor environment (such as temperature and humidity) were assigned a moderate benchmark weight, and the secondary auxiliary parameters were assigned a lower benchmark weight. At the same time, a dynamic adjustment threshold for the weight was preset. Combined with changes in indoor scene, number of people and differences in needs, the triggering conditions and adjustment range of weight adjustment were clarified to form a dynamically adaptable benchmark parameter system. Finally, all benchmark parameters were integrated, the weight allocation rules were standardized, the parameter value range and adjustment logic were clarified, and the construction of a standardized weight ratio system was completed.

[0060] The model library management module is used to store control decision models corresponding to different indoor scenarios and build a scenario-based decision model library. Specifically, firstly, considering the different control needs of different functional scenarios in the building interior (such as office areas, meeting rooms, rest areas, corridors, etc.), we first sort out the core control objectives, environmental parameter ranges, and personnel needs of various scenarios. Based on this, we collect and screen control decision models that are suitable for each scenario, including basic control models, personalized adaptation models, and optimized iterative models, to ensure that the models accurately correspond to the needs of the scenarios.

[0061] Subsequently, high-performance storage servers were selected as storage carriers, and a classified storage mechanism was adopted to archive various control decision models according to indoor scene types. Each model was labeled with key information such as scene identifier, adaptive parameter range, core control logic, and applicable personnel scale. At the same time, a dual backup mechanism was established to prevent model data loss or damage and to ensure the security and integrity of model storage.

[0062] Next, we construct the basic architecture of the scenario-based decision-making model library, establish a scenario index system, associate scenario types with corresponding control decision-making models, design standardized model input, query, and update interfaces to facilitate subsequent model calls and management, and finally establish a dynamic maintenance mechanism for the model library, preset model update, iteration, and removal rules, and promptly supplement new scenario models, optimize existing models, and remove ineffective models based on changes in indoor scenarios and feedback on control effects, forming a scenario-based decision-making model library that is structurally complete, highly adaptable, and dynamically updatable.

[0063] The data weight allocation module is used to perform weight allocation processing on indoor environmental data and face analysis data according to a standardized data weight ratio system, so as to form a weighted multi-source fusion feature. Specifically, based on the pre-set standardized data weighting system, the indoor environment data and facial recognition data are first verified and standardized to remove missing data, outliers, and invalid information with inconsistent formats, ensuring that the two types of data are of uniform specifications, complete information, and meet the processing requirements for weight allocation.

[0064] Subsequently, strictly following the benchmark parameters and allocation rules in the weighting system, weight allocation processing was performed on the two types of data respectively: First, according to the basic weights specified in the system, overall weights were allocated to indoor environmental data (temperature, humidity, etc.) and facial analysis data (physiological state, thermal comfort needs, etc.). Then, for each sub-dimension of the two types of data, corresponding sub-weights were allocated according to the sub-weight standards in the system and the degree of influence of the sub-dimensions on the control decision, to ensure that the weight allocation conforms to the system requirements and adapts to the control needs.

[0065] Then, each data point is multiplied by its corresponding weight to obtain the weighted feature value of each dimension. The difference in the units of measurement of data in different dimensions is eliminated simultaneously. Finally, all weighted feature values ​​are integrated and arranged in an orderly manner for each dimension to complete the weighted fusion of multi-source data and form a weighted multi-source fusion feature that contains dual information of indoor environment and personnel needs, with unified dimensions and balanced values.

[0066] The decision model matching module is used to perform feature matching and fit calculation between the weighted multi-source fusion features and the models in the scenario-based decision model library, and select the decision model with the highest fit as the basic decision model.

[0067] In this embodiment of the invention, the decision model matching module includes: a feature preprocessing module, a fitness calculation module, a model screening module, and a model validation module; The feature preprocessing module is used to standardize and normalize the weighted multi-source fusion features output by the data weight allocation module, and remove redundant features and abnormal data to obtain standardized fusion features. Specifically, the weighted multi-source fusion features output by the data weight allocation module are used as the processing object. Pre-data verification is carried out to check for missing fields, extreme outliers and non-standard format information in the feature data one by one. Missing data is filled by interpolation and extreme outliers are removed by statistical thresholds to ensure the integrity and validity of the input feature data, laying the foundation for subsequent standardization, normalization and redundancy and anomaly handling.

[0068] Then, standardization and normalization processing is performed. A combination of Z-score standardization and extreme value normalization is used to uniformly map the feature data of different dimensions and magnitudes in the weighted multi-source fusion features to a preset standard range, eliminating the difference in dimensions and numerical offset, so that the features of each dimension have comparable and computable consistency, and avoid the interference of a single feature with an excessively high magnitude to subsequent processing.

[0069] Next, redundant features and abnormal data removal were carried out. Through feature correlation analysis algorithms, redundant features with high correlation and information duplication were screened and eliminated, while core effective features were retained. At the same time, combined with the 3σ principle and scenario adaptability judgment, abnormal feature data that deviated from the normal distribution and were irrelevant to the indoor environment control scenario were further eliminated, reducing data redundancy and improving feature purity. Finally, the processed feature data was structured and standardized, and the feature format and arrangement order were unified to obtain standardized fusion features that are uniform in specifications, complete in information, free of redundancy and abnormality.

[0070] The fitness calculation module is used to calculate the feature similarity between the standardized fusion features and each regulation decision model in the scenario-based decision model library using the cosine similarity algorithm, so as to obtain the fitness value corresponding to each model. Specifically, the core processing object is each regulation decision model in the standardized fusion feature and scenario-based decision model library. The feature dimensions of the two are aligned first, and the feature vector dimensions of each regulation decision model in the model library are standardized to ensure that they are completely consistent with the dimensions of the standardized fusion feature. At the same time, invalid and redundant dimensional information in the model features are removed to ensure that the feature data input to the algorithm is standardized and uniform, laying the foundation for cosine similarity calculation.

[0071] Subsequently, a cosine similarity algorithm was introduced. While clarifying the core logic of the algorithm, standardized fusion features were used as the baseline feature vector. The core feature vectors of each regulation decision model in the scenario-based decision model library were extracted one by one. According to the vector dot product operation rules, the dot product and modulus product of the baseline feature vector and the feature vectors of each model were calculated respectively. Then, the cosine value of the angle between the two was calculated by the cosine function. The range of the cosine value is [-1, 1]. The closer the value is to 1, the higher the feature similarity between the two and the stronger the model fit. The closer the value is to -1, the lower the similarity and the weaker the fit.

[0072] Finally, the cosine value obtained from each calculation is used as the fitness value of the corresponding control decision model. All fitness values ​​are standardized and regularized, and the corresponding model identifiers are marked to ensure that the numerical format is uniform and comparable. Finally, the fitness value corresponding to each control decision model is obtained.

[0073] The model screening module is used to preset the fit screening threshold and screen the control decision model based on the fit screening threshold. Specifically, firstly, based on the precision requirements of building indoor environment control, scenario adaptation needs, and historical system operation data, a preset adaptation screening threshold is established to construct a threshold setting system. The screening threshold needs to take into account both adaptation precision and model diversity, refer to the optimal adaptation range of past model matching, and combine the control priority of different indoor scenarios (office areas, meeting rooms, etc.) to preset basic screening thresholds. At the same time, dynamic adjustment space is reserved, and the triggering conditions for threshold adjustment are clarified (such as overall low adaptation, scenario changes, etc.) to ensure that the thresholds not only meet the control needs but also flexibly adapt to different scenarios.

[0074] Subsequently, the fitness values ​​of each control decision model calculated by the cosine similarity algorithm are compared with the preset fitness screening threshold one by one. The screening follows the core rule of "fit value ≥ screening threshold" to select all control decision models that meet the conditions as candidate basic decision models. If the fitness of all models is lower than the threshold, the threshold dynamic adjustment mechanism is activated to appropriately lower the threshold or return to the data processing stage for re-optimization to ensure that there are candidate models that meet the requirements. Finally, the selected candidate models are initially standardized, and the fitness values ​​of each model are labeled with the corresponding scenario identifier. Models with abnormal fluctuations in fitness are removed to form a standardized candidate model set.

[0075] The model verification module is used to verify the selected basic decision-making models in real time and confirm the adaptability of the model output results to the current indoor scene and personnel needs. If the validation passes, the model will be output as the base decision model. If the verification fails, return to the model filtering module for re-filtering.

[0076] The weighted data control instruction module 4 is used to input indoor environmental data and facial analysis data into the basic decision model, construct a data weighted fusion model, preset comfort thresholds and scene control strategies, and then combine the data weighted fusion model to generate adaptive control instructions for the building's indoor environment. In this embodiment of the invention, the weighted data control instruction module 4 includes: a fusion model construction module, a comfort threshold configuration module, a scene strategy matching module, and a control instruction generation module; The fusion model construction module is used to input indoor environmental data and facial recognition data into the basic decision model, perform multi-source data weighted calculation and model parameter initialization, and obtain a data weighted fusion model. Specifically, the first step is to use the basic decision-making model that has passed screening and verification as the core carrier. The indoor environmental data and facial recognition data are first verified and standardized, and invalid information such as missing data, abnormal format and mismatch with the scene is removed to ensure that the two types of data are of uniform specifications, complete information and fully adapted to the input dimensions and data format of the basic decision-making model. This lays a solid foundation for the subsequent multi-source data weighted calculation and model parameter initialization.

[0077] Subsequently, the standardized indoor environmental data and facial recognition data are simultaneously input into the basic decision-making model. Strictly following the preset standardized data weighting system, weighted calculations are performed on the two types of data and their respective sub-dimensions. The data of each dimension are multiplied with their corresponding weights and then superimposed and integrated dimension by dimension, integrating the dual core information of environmental parameters and personnel needs, and eliminating the differences in the dimensions and numerical deviations of data from different dimensions.

[0078] Simultaneously, combining the core control logic of the basic decision-making model with the characteristics of the current indoor scene, the key parameters such as the model's weight coefficients, fusion rules, and calculation thresholds are initialized and set. The parameter value range is calibrated to ensure that the parameters are adapted to the current data characteristics and control needs. Finally, through preliminary iterative calibration of the model, the weighted calculation accuracy and parameter adaptability are optimized to obtain a data weighted fusion model that takes into account both environmental adaptability and personalized needs of personnel, with accurate calculation and stable operation.

[0079] The comfort threshold configuration module is used to preset the comfort thresholds for temperature, humidity, light, and air quality for different personnel identities, physiological states, and indoor scenes, forming a standardized comfort threshold library. Specifically, the core basis for setting comfort thresholds is to combine human physiological and medical standards, building indoor environment control industry standards, and the differentiated comfort needs of different people with different identities, physiological states, and indoor functional scenarios, and to integrate historical system operation data, measured environmental parameters, and personnel feedback information.

[0080] For the four core environmental parameters of temperature, humidity, light, and air quality, multi-dimensional differentiated presets are made according to the type of personnel, the level of physiological state, and the indoor functional scenario. For people with different identities, different fatigue levels and body temperature ranges, as well as different scenarios such as office, meeting, and rest, corresponding suitable ranges and critical thresholds are matched respectively. Personalized thresholds are distinguished between normal state and special physiological state to ensure that the thresholds meet the human comfort needs and the usage characteristics of the scenario.

[0081] Subsequently, all preset thresholds were standardized and regulated, and numerical formats, unit specifications, and labeling rules were unified. An association mapping index between personnel identity, physiological state, indoor scene, and corresponding comfort threshold was established to form structured and searchable threshold entries. At the same time, a dynamic threshold calibration mechanism was established to periodically verify, optimize, and update the thresholds based on actual control effects, personnel comfort feedback, and scene changes, eliminating unreasonable thresholds and supplementing adaptive thresholds for new scenes. Finally, all standardized and dynamically adjustable comfort thresholds were integrated to build a standardized comfort threshold library covering multiple identities, multiple states, and multiple scenes.

[0082] The scene strategy matching module is used to preset a scene control strategy library and match indoor environmental data and face parsing data with the scene control strategy library; Specifically, firstly, considering different functional scenarios such as office, meeting, and rest areas within a building, as well as differences in personnel identity, physiological state, and thermal comfort needs, the control objectives, execution logic, and priority rules for temperature, humidity, light, and air quality under each scenario are sorted out. Scenario control strategies, including routine control, personalized control, and emergency control, are preset. The strategies are then classified and archived according to scenario type, personnel status, and environmental parameter range. The applicable conditions, control range, and constraints of each strategy are clarified, forming a standardized scenario control strategy library covering multiple scenarios and needs.

[0083] Subsequently, real-time collected and standardized indoor environmental data and facial recognition data are used as matching inputs. The two types of data are aligned in dimensions and formatted to ensure consistency with the matching fields and feature dimensions in the strategy library. Then, a combination of feature matching and comfort threshold comparison is used to accurately match real-time environmental parameters, personnel identification, physiological state, thermal comfort needs, and the applicable conditions of each control strategy in the strategy library. Through fit calculation and threshold judgment, control strategies that are highly compatible with the current indoor scene, environmental state, and personnel needs are selected. Finally, the effectiveness of the matching results is verified, strategies with low fit or that do not meet the control constraints are eliminated, and the best-fit strategy is retained and labeled.

[0084] The control instruction generation module is used to calculate target control parameters based on a data weighted fusion model, combined with a standardized comfort threshold library and matching scenario control strategies, and generate adaptive control instructions for the building indoor environment that include adjustment parameters for air conditioning, fresh air, and lighting equipment.

[0085] Specifically, firstly, real-time standardized indoor environmental data and facial recognition data are retrieved synchronously, and then the standardized comfort threshold library is connected with the scene control strategy that has been accurately matched. This completes the three-dimensional alignment and validity verification of the calculation data source, comfort threshold, and execution strategy, eliminating abnormal information such as parameter conflicts, threshold out-of-bounds errors, and strategy incompatibility, ensuring the stability and reliability of the calculation premise.

[0086] Subsequently, relying on the weighted calculation logic of the data weighted fusion model, the difference between the measured parameters of the current environment, the physiological state and thermal comfort needs of the personnel and the target comfort range in the comfort threshold library is quantitatively calculated. Combined with the constraints such as the control amplitude, priority and execution rhythm set by the scenario control strategy, the target control parameters such as the air conditioning target temperature and operation mode, the fresh air system air volume and air exchange frequency, and the brightness and color temperature level of the lighting equipment are accurately calculated to ensure that the parameters are both in line with the personnel's identity preferences and physiological state and comply with the scenario control specifications, so as to avoid parameter sudden changes or operation exceeding the threshold.

[0087] Finally, the calculated target control parameters are standardized and coded and organized at the equipment level. The specific adjustment parameters are integrated according to three types of equipment: air conditioning, fresh air and lighting. The resulting standardized, clear, and directly executable adaptive control commands for the building's indoor environment are generated. After the legality and security of the commands are verified, they are output to the equipment control terminal to achieve adaptive, refined and personalized control of the indoor environment based on the real-time status of the personnel.

[0088] The device adaptive control module 5 is used to adjust the opening degree, power and mode of indoor environment control equipment such as air conditioner, fresh air and lighting according to the building indoor environment adaptive control command; Specifically, the first step is to verify the legality of the instructions and parse the parameters. This involves checking the unique identifiers of the air conditioning, fresh air, lighting, and other equipment in the instructions, as well as the target control parameters and safety constraints. Abnormal information such as parameters exceeding limits and instructions conflicting are then eliminated to ensure that the instructions are valid and compatible with the corresponding equipment operating specifications.

[0089] Subsequently, in accordance with the instructions, precise adjustments were made to each type of equipment: for air conditioning equipment, the compressor output power, damper opening, air outlet direction, and operating mode were adjusted according to the instructions, and the target temperature and wind speed level were matched simultaneously; for fresh air equipment, the fan speed, air supply duct valve opening, and ventilation mode were adjusted, and the fresh air volume and exhaust ratio were adjusted as needed; for lighting equipment, the drive power, brightness level, color temperature level, and on / off status were adjusted, and the light intensity and scene mode set in the instructions were strictly followed.

[0090] The adjustment process follows the principle of gradual change to avoid the impact of sudden parameter changes on the equipment and indoor environment. At the same time, the equipment operation feedback data is collected in real time, and the actual opening degree, power, mode and command target value are compared and calibrated. If a deviation occurs, it is corrected in a small way immediately. In addition, the linkage equipment safety protection mechanism automatically limits the amplitude protection when the adjustment parameter reaches the safety threshold to ensure stable operation of the equipment. After all equipment has completed the adjustment and met the standard, the execution status and operating parameters are fed back, forming a complete command execution closed loop.

[0091] The operation feedback iteration module 6 is used to collect indoor environmental operation data and personnel comfort feedback data after equipment adjustment, and to compare and analyze the control effect to optimize the scenario-based decision model library.

[0092] In this embodiment of the invention, the operation feedback iteration module 6 includes: an operation data acquisition module, a comfort feedback receiving module, a regulation effect analysis module, and a fusion model optimization module; The operation data acquisition module is used to collect indoor environmental operation data such as indoor temperature, humidity, light, air quality and equipment operation status in real time after the equipment is adjusted. Specifically, after the adjustment of the building's indoor environmental control equipment is completed, multi-type sensor monitoring units are deployed throughout the entire area to collect real-time data on the indoor environment and equipment operation. Temperature, humidity, light intensity, and air quality sensors are reasonably deployed in various functional areas of the building to achieve full coverage and high-frequency real-time collection of core environmental parameters such as temperature, relative humidity, light intensity, PM2.5, and CO2 concentration, accurately reflecting the environmental changes after equipment control.

[0093] Simultaneously, it is equipped with an equipment operation status acquisition module to synchronously acquire real-time operating parameters such as the opening degree, output power, operating mode, and fan speed of control equipment such as air conditioning, fresh air, and lighting, ensuring that environmental data and equipment status data are collected synchronously and correspond one-to-one.

[0094] The collected raw data is uploaded to the data processing unit in real time via a stable communication transmission module. After preprocessing operations such as noise reduction filtering, outlier removal, and missing data completion, signal interference and data deviations are eliminated, ensuring the accuracy and validity of the collected data. A real-time data verification mechanism is established to compare the target values ​​of control commands with the actual collected values, verifying the rationality and completeness of the data. Finally, the processed standardized indoor environmental operation data is archived and dynamically updated in real time, forming a set of operational data that can be accessed in real time and is traceable.

[0095] The comfort feedback receiving module is used to acquire subjective comfort feedback data of people in the building interior and form a personnel comfort evaluation dataset. Specifically, the first step is to build a multi-channel, lightweight subjective feedback collection system, deploy touch feedback terminals in various functional areas of the room, and provide convenient methods such as mobile mini-programs and voice interaction to cover people in all scenarios such as office, meeting, and rest, thereby reducing the threshold for feedback operations.

[0096] Subsequently, a unified subjective comfort feedback index system was designed, focusing on four core dimensions: temperature, humidity, light, and air quality. Quantitative indicators such as comfort level evaluation, overall satisfaction, and personalized adjustment preferences were set up to transform vague subjective feelings into calculable graded values.

[0097] During data collection, personnel identification, collection time, scene, and real-time physiological status are simultaneously bound to ensure that feedback data is accurately correlated with the objective environment and personnel status. The raw feedback data is cleaned in real time to remove duplicate, invalid, and abnormal information, and the rationality and completeness of the data are verified. Then, the valid data is standardized and coded, and structured according to evaluation dimensions and graded values. A correlation mapping is established with environmental operation data and facial recognition data. Finally, the data is classified and archived according to dimensions such as personnel identity, physiological status, and indoor scene to build a queryable, traceable, and iterative personnel comfort evaluation dataset. A dynamic update mechanism is established to continuously collect new feedback data and improve the dataset.

[0098] The regulation effect analysis module is used to compare indoor environmental operation data, personnel comfort evaluation dataset, and standardized comfort threshold library to perform quantitative analysis on the accuracy of environmental regulation and comfort satisfaction, and obtain regulation effect evaluation results. In this embodiment of the invention, the regulation effect analysis module includes: a data comparison and statistics module and a comprehensive accuracy evaluation module; The data comparison and statistics module is used to compare the indoor environmental operation data with the standardized comfort threshold library for deviation, and to perform positive and negative feedback classification statistics on the personnel comfort evaluation dataset to obtain the environmental parameter deviation value and the comfort feedback ratio. Specifically, the indoor environmental operation data collected and preprocessed in real time is first standardized and regulated to remove noise interference, abnormal extreme values ​​and missing information, and to ensure that the formats of parameters such as temperature, humidity, light, and air quality are fully matched with the standardized comfort threshold library, so as to provide a reliable data foundation for deviation comparison.

[0099] Subsequently, the measured parameters of each environment were compared with the target comfort ranges of the corresponding personnel identities, physiological states, and indoor scenes in the threshold library one by one to calculate the difference. The actual deviation magnitude and direction of each parameter were accurately calculated to form a quantitative environmental parameter deviation value, which clearly reflects the objective gap between the current environment and the comfort standard. At the same time, for the personnel comfort evaluation dataset, positive and negative feedback were classified according to the subjective feeling type. Evaluations such as comfort and satisfaction were classified as positive feedback, while evaluations such as cold, hot, and stuffy were classified as negative feedback, thus completing the classification labeling and summarization.

[0100] Based on this, the number of positive and negative feedbacks in each dimension and their proportion of the total feedback were counted to obtain the proportion of comfort feedback corresponding to different environmental parameters, which intuitively reflects the subjective satisfaction of the personnel. Finally, the environmental parameter deviation value and the proportion of comfort feedback were structured and integrated to form a quantitative analysis result that combines objective data deviation and subjective feedback proportion.

[0101] The comprehensive accuracy assessment module is used to perform quantitative calculations and comprehensive judgments based on environmental parameter deviation values ​​and the proportion of comfort feedback, and to generate a control effect assessment result that includes control accuracy, comfort satisfaction, and effect level.

[0102] Specifically, the first step is to use the deviation values ​​of environmental parameters and the proportion of comfort feedback as the core quantitative basis. The standardization and weighting of the two types of data are completed first. According to the control priority of core dimensions such as temperature and humidity, differentiated calculation weights are assigned to the deviation values ​​of each parameter and the corresponding proportion of comfort feedback, eliminating the one-sidedness of single-dimensional data and building a unified comprehensive judgment framework.

[0103] Quantitative calculations were then performed: for control accuracy, the deviation values ​​of each environmental parameter were normalized and then weighted and summed, with the smaller the deviation value, the higher the accuracy score; for comfort and satisfaction, the proportion of positive feedback was used as the core, and the degree of discomfort of negative feedback was combined for weighted correction to quantitatively reflect the subjective satisfaction level of the personnel.

[0104] Next, based on the preset evaluation threshold system, the quantitative scores of accuracy and satisfaction are compared with the judgment thresholds for four levels of effectiveness: excellent, good, average, and poor. The level is determined by considering the priority of scenario-based control. Finally, the calculation and judgment results are structured and integrated, clearly marking the control accuracy value, comfort and satisfaction percentage, and effectiveness level, along with the judgment criteria for the core dimensions, to form a standardized control effect evaluation result.

[0105] The fusion model optimization module is used to iteratively optimize the weight parameters and calculation rules of the data weighted fusion model based on the evaluation results of the regulation effect, and synchronously update the optimized model to the scenario-based decision model library.

[0106] Specifically, we first conduct an in-depth analysis of the evaluation results, combining the deviation values ​​of environmental parameters and the proportion of comfort feedback to identify problems in the model's weight allocation and calculation rules, and clarify optimization directions such as the imbalance of weight ratios between facial recognition data and indoor environmental data, unreasonable sub-dimension weights, and insufficient adaptability of the fusion calculation logic to the actual scenario.

[0107] Subsequently, an adaptive weight optimization algorithm was employed to dynamically adjust the overall weights of the two types of data in the model, as well as the subdivided weights for subdivisions such as temperature, humidity, and physiological state. The calculation rules for weighted superposition and feature fusion were optimized, and the calculation thresholds and constraints were calibrated to improve the model's adaptability to different personnel states and indoor scenarios. After optimization, simulation tests and actual control verification ensured that the control accuracy and comfort level met the standards, with no calculation anomalies or parameter conflicts.

[0108] Finally, the validated optimized model is standardized and packaged, with version and applicable scenarios labeled, and synchronously updated to the scenario-based decision model library, replacing the original model. Simultaneously, an iteration record and traceability mechanism is established to retain historical versions for easy backtracking. Through this closed-loop iterative approach, the model's computational accuracy is continuously improved, providing more efficient and precise decision support for adaptive control of the building's indoor environment.

[0109] According to another aspect of the invention, such as Figure 2 As shown, an adaptive control method for building indoor environment based on deep facial feature analysis and multi-source weighting is provided. This method includes the following steps: S1. Deploy multi-dimensional environmental sensors to collect real-time data on indoor temperature, humidity, light, air quality, and equipment operating status, and summarize them to form indoor environmental data; S2. Collect facial image data and behavioral feature data of people inside the building, and extract and integrate physiological state parameters, thermal comfort requirement parameters and identity feature information of people to form facial analysis data; S3. Preset scenario-based decision model library and data weight ratio. Based on the data weight ratio, assign weights to indoor environmental data and face analysis data to form weighted multi-source fusion features. Then match the weighted multi-source fusion features with the models in the scenario-based decision model library to obtain the basic decision model. S4. Input indoor environmental data and facial recognition data into the basic decision model, construct a data weighted fusion model, preset comfort thresholds and scene control strategies, and then combine the data weighted fusion model to generate adaptive control instructions for the building's indoor environment. S5. Based on the building's indoor environment adaptive control instructions, adjust the opening degree, power and mode of indoor environment control equipment such as air conditioning, fresh air, and lighting; S6. Collect indoor environmental operation data and personnel comfort feedback data after equipment adjustment, compare and analyze the control effect, and optimize the scenario-based decision-making model library.

[0110] 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 within the protection scope of the present invention.

Claims

1. An adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting, characterized in that, The system includes: an indoor environment perception and monitoring module, a facial feature acquisition and analysis module, a basic model decision matching module, a weighted data control instruction module, an equipment adaptive control module, and an operation feedback and iteration module; The indoor environment sensing and monitoring module is used to deploy multi-dimensional environmental sensors to collect data on indoor temperature, humidity, light, air quality and equipment operation status in real time, and summarize them to form indoor environment data. The facial feature acquisition and analysis module is used to collect facial image data and behavioral feature data of people inside the building, and extract and integrate physiological state parameters, thermal comfort requirement parameters and identity feature information of people to form facial analysis data; The basic model decision matching module is used to preset the scenario-based decision model library and data weight ratio, allocate weights to indoor environmental data and face analysis data according to the data weight ratio, form weighted multi-source fusion features, and match the weighted multi-source fusion features with the models in the scenario-based decision model library to obtain the basic decision model. The weighted data control instruction module is used to input indoor environmental data and facial recognition data into the basic decision model, construct a data weighted fusion model, preset comfort thresholds and scene control strategies, and then combine the data weighted fusion model to generate adaptive control instructions for the building's indoor environment. The device adaptive control module is used to adjust the opening degree, power and mode of indoor environment control equipment such as air conditioner, fresh air and lighting according to the building indoor environment adaptive control command; The operation feedback iteration module is used to collect indoor environmental operation data and personnel comfort feedback data after equipment adjustment, and to compare and analyze the control effect to optimize the scenario-based decision model library.

2. The adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting as described in claim 1, characterized in that, The facial feature acquisition and analysis module includes: a facial image acquisition module, a behavioral feature capture module, a physiological state analysis module, and a comfort requirement output module; The face image acquisition module is used to acquire raw face image data of people inside the building in real time through indoor camera equipment, and to perform image noise reduction and normalization preprocessing. The behavioral feature capture module is used to capture data on people's body movements, spatial location distribution, and facial expression changes through indoor camera equipment as behavioral feature data, and bind the corresponding person's identity feature information. The physiological state analysis module is used to perform deep analysis of facial image data and behavioral feature data based on a deep learning model to obtain physiological state parameters and thermal comfort requirement parameters of personnel. The comfort requirement output module is used to standardize and encode personnel physiological state parameters, thermal comfort requirement parameters and identity feature information, and establish an association mapping according to the identity feature information to form face parsing data containing personnel identity identifiers, physiological state parameters and thermal comfort requirement parameters.

3. The adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting as described in claim 1, characterized in that, The basic model decision matching module includes: a weight parameter preset module, a model library management module, a data weight allocation module, and a decision model matching module; The weight parameter preset module is used to preset the weight allocation benchmark parameters of indoor environmental data and face analysis data to form a data weight ratio system. The model library management module is used to store control decision models corresponding to different indoor scenarios and build a scenario-based decision model library. The data weight allocation module is used to perform weight allocation processing on indoor environmental data and face analysis data according to a standardized data weight ratio system, so as to form a weighted multi-source fusion feature. The decision model matching module is used to perform feature matching and fit calculation between the weighted multi-source fusion features and the models in the scenario-based decision model library, and select the decision model with the highest fit as the basic decision model.

4. The adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting as described in claim 1, characterized in that, The weighted data control instruction module includes: a fusion model construction module, a comfort threshold configuration module, a scenario strategy matching module, and a control instruction generation module; The fusion model construction module is used to input indoor environmental data and facial recognition data into the basic decision model, perform multi-source data weighted calculation and model parameter initialization, and obtain a data weighted fusion model. The comfort threshold configuration module is used to preset the comfort thresholds for temperature, humidity, light, and air quality for different personnel identities, physiological states, and indoor scenes, forming a standardized comfort threshold library. The scene strategy matching module is used to preset a scene control strategy library and match indoor environmental data and face parsing data with the scene control strategy library; The control instruction generation module is used to calculate target control parameters based on a data weighted fusion model, combined with a standardized comfort threshold library and matching scenario control strategies, and generate adaptive control instructions for the building indoor environment that include adjustment parameters for air conditioning, fresh air, and lighting equipment.

5. The adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting as described in claim 1, characterized in that, The operation feedback iteration module includes: an operation data acquisition module, a comfort feedback receiving module, a regulation effect analysis module, and a fusion model optimization module; The operation data acquisition module is used to collect indoor environmental operation data such as indoor temperature, humidity, light, air quality and equipment operation status in real time after the equipment is adjusted. The comfort feedback receiving module is used to acquire subjective comfort feedback data of people in the building interior and form a personnel comfort evaluation dataset. The regulation effect analysis module is used to compare indoor environmental operation data, personnel comfort evaluation dataset, and standardized comfort threshold library to perform quantitative analysis on the accuracy of environmental regulation and comfort satisfaction, and obtain regulation effect evaluation results. The fusion model optimization module is used to iteratively optimize the weight parameters and calculation rules of the data weighted fusion model based on the evaluation results of the regulation effect, and synchronously update the optimized model to the scenario-based decision model library.

6. The adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting as described in claim 2, characterized in that, The physiological state analysis module includes: a facial key feature extraction module, a behavior feature quantification module, a multi-feature fusion analysis module, and a physiological comfort output module; The facial key feature extraction module is used to extract features from facial image data based on a deep convolutional neural network model, and obtain facial key features including facial key point coordinates, skin color heat distribution features, and eye opening and closing status. The behavioral feature quantification module is used to quantify the range of human body movements, spatial distribution density, and range of facial expression changes in behavioral feature data to obtain behavioral quantification features including limb extension, human position distribution, and facial pleasure. The multi-feature fusion and parsing module is used to fuse key facial features with behavioral quantitative features and perform correlation analysis using a convolutional long short-term memory neural network model to obtain physiological state parameters of the person, including body temperature range and fatigue level classification. The physiological comfort output module is used to match personnel physiological state parameters with identity feature information to generate personnel thermal comfort requirement parameters including preferred temperature range, suitable humidity range and light tolerance level.

7. The adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting as described in claim 3, characterized in that, The decision model matching module includes: a feature preprocessing module, a fitness calculation module, a model screening module, and a model validation module; The feature preprocessing module is used to standardize and normalize the weighted multi-source fusion features output by the data weight allocation module, and remove redundant features and abnormal data to obtain standardized fusion features. The fitness calculation module is used to calculate the feature similarity between the standardized fusion features and each regulation decision model in the scenario-based decision model library using the cosine similarity algorithm, so as to obtain the fitness value corresponding to each model. The model screening module is used to preset the fit screening threshold and screen the control decision model based on the fit screening threshold. The model verification module is used to verify the selected basic decision-making models in real time and confirm the adaptability of the model output results to the current indoor scene and personnel needs. If the validation passes, the model will be output as the base decision model. If the verification fails, return to the model filtering module for re-filtering.

8. The adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting as described in claim 5, characterized in that, The regulation effect analysis module includes: a data comparison and statistics module and a comprehensive accuracy evaluation module; The data comparison and statistics module is used to compare the indoor environmental operation data with the standardized comfort threshold library for deviation, and to perform positive and negative feedback classification statistics on the personnel comfort evaluation dataset to obtain the environmental parameter deviation value and the comfort feedback ratio. The comprehensive accuracy assessment module is used to perform quantitative calculations and comprehensive judgments based on environmental parameter deviation values ​​and the proportion of comfort feedback, and to generate a control effect assessment result that includes control accuracy, comfort satisfaction, and effect level.

9. The adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting as described in claim 6, characterized in that, The multi-feature fusion and parsing module includes: a feature normalization processing unit, a weighted feature fusion unit, a temporal correlation analysis unit, and a physiological state output unit; The feature normalization processing unit is used to perform dimensionless and standardized processing on key facial features and quantitative behavioral features, respectively, and to eliminate dimensional differences and numerical deviations between features of different dimensions. The weighted feature fusion unit is used to assign corresponding fusion weights to the standardized facial key features and behavior quantification features, and to perform weighted superposition operation to obtain a high-dimensional fusion feature vector. The temporal correlation analysis unit is used to input high-dimensional fused feature vectors into a convolutional long short-term memory neural network model to perform in-depth mining and correlation reasoning on the feature change patterns in the temporal dimension. The physiological state output unit is used to determine the physiological state parameters of personnel, including body temperature range and fatigue level, based on the time-series correlation analysis results and a preset classification threshold.

10. A method for adaptive control of building indoor environment based on deep facial feature analysis and multi-source weighting is used to implement the adaptive control system for building indoor environment based on deep facial feature analysis and multi-source weighting as described in any one of claims 1-9, characterized in that, The method includes the following steps: S1. Deploy multi-dimensional environmental sensors to collect real-time data on indoor temperature, humidity, light, air quality, and equipment operating status, and summarize them to form indoor environmental data; S2. Collect facial image data and behavioral feature data of people inside the building, and extract and integrate physiological state parameters, thermal comfort requirement parameters and identity feature information of people to form facial analysis data; S3. Preset scenario-based decision model library and data weight ratio. Based on the data weight ratio, assign weights to indoor environmental data and face analysis data to form weighted multi-source fusion features. Then match the weighted multi-source fusion features with the models in the scenario-based decision model library to obtain the basic decision model. S4. Input indoor environmental data and facial recognition data into the basic decision model, construct a data weighted fusion model, preset comfort thresholds and scene control strategies, and then combine the data weighted fusion model to generate adaptive control instructions for the building's indoor environment. S5. Based on the building's indoor environment adaptive control instructions, adjust the opening degree, power and mode of indoor environment control equipment such as air conditioning, fresh air, and lighting; S6. Collect indoor environmental operation data and personnel comfort feedback data after equipment adjustment, compare and analyze the control effect, and optimize the scenario-based decision-making model library.

Images