Intelligent Monitoring Method for Terminal Building Environment Based on Multi-Source Data Fusion and Deep Learning

Abstract

The present invention provides a terminal environment intelligent monitoring method combining multi-source data fusion with deep learning, which relates to the field of building monitoring technology, including: collecting and preprocessing multi-source perception data, inputting a dual-stream neural cognitive computing framework to extract features, generating environmental assessment results through an environmental state assessment model, and generating and executing environmental control strategies using a spatiotemporal dynamic decoupler and a multi-objective optimization model, thereby realizing intelligent monitoring and optimal control of the terminal environment, improving environmental comfort and safety, reducing energy consumption, and enhancing emergency event handling capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of building monitoring, and in particular to an intelligent monitoring method for terminal environment that combines multi-source data fusion with deep learning. Background Art

[0002] With the rapid development of the air transportation industry, modern terminals, as important transportation hubs, have increasingly become public places with dense crowds. The environmental quality of terminals directly affects the passenger experience, the health of staff, and the operating efficiency of equipment;

[0003] Traditional terminal environment monitoring systems mainly rely on single-type sensors and use simple threshold judgment methods for environmental monitoring and regulation, which cannot meet the intelligent management requirements in the complex environment of modern terminals. With the rapid development of the Internet of Things technology, multi-source heterogeneous sensing devices have been widely deployed in terminals, generating a large amount of environmental monitoring data, including multi-dimensional information such as temperature and humidity, air quality, noise, crowd density, and illumination. At the same time, deep learning technology has demonstrated powerful capabilities in data processing and feature extraction, providing technical possibilities for the accurate perception, intelligent evaluation, and optimized regulation of the terminal environment;

[0004] However, the existing technologies still have problems such as the lack of effective processing ability for multi-source heterogeneous data, insufficient ability to extract spatio-temporal features, lack of analysis ability for long-term and short-term dependence relationships between environmental factors, and lack of intelligent regulation ability;

[0005] Therefore, there is an urgent need for a solution to solve the problems existing in the existing technologies. Summary of the Invention

[0006] An embodiment of the present invention provides an intelligent monitoring method for terminal environment that combines multi-source data fusion with deep learning, which can at least solve some of the problems existing in the existing technologies.

[0007] In a first aspect of an embodiment of the present invention, an intelligent monitoring method for terminal environment that combines multi-source data fusion with deep learning is provided, including:

[0008] Collect multi-source perception data in the terminal and perform preprocessing of time series alignment and spatial registration to generate a standardized perception data matrix, and segment the perception data matrix;

[0009] Input the perception data matrix into a dual-stream neural cognitive computing framework, extract multi-scale spatio-temporal features through an adaptive dynamic convolution module, determine the long-term and short-term dependence relationships between different types by combining a deformable attention mechanism to obtain perception stream features, construct an environmental knowledge reasoning module based on a memory-enhanced neural network, and mine the causal relationship of environmental changes by combining historical experience knowledge and a recursive neural reasoning unit to obtain cognitive stream features, and fuse the perception stream features and the cognitive stream features to obtain a fused feature representation;

[0010] Input the fusion feature representation into the terminal environment state evaluation model to generate an environmental evaluation result including the terminal environmental comfort index, safety risk level, and emergency event warning information;

[0011] Input the environmental evaluation result into the spatio-temporal dynamic decoupler, decouple in the time dimension and space dimension by combining the attention mechanism to obtain dynamic environmental features, construct a multi-objective optimization model based on the dynamic environmental features, construct a constrained optimization problem through the particle swarm optimization algorithm, combine the adaptive inertia weight and the dynamic convergence factor to solve the preset optimization objective, generate a multi-objective solution result, and generate an environmental regulation strategy through the fuzzy mapping algorithm;

[0012] Send the environmental regulation strategy to the terminal environmental control system for execution.

[0013] In an alternative embodiment,

[0014] Collect multi-source perception data in the terminal and perform preprocessing of time series alignment and spatial registration to generate a standardized perception data matrix. The segmentation of the perception data matrix includes:

[0015] Collect multi-source data in the terminal through temperature and humidity sensors, air quality sensors, video surveillance devices, and noise sensors;

[0016] Synchronize and align the multi-source data according to the time stamp, and perform registration of sensing data at different positions based on spatial coordinate mapping. Generate a perception data matrix in a unified format through standardization processing;

[0017] Adopt the sliding time window technique, set the initial sliding window according to the preset time step, and dynamically segment the perception data matrix in combination with the data change trend.

[0018] In an alternative embodiment,

[0019] Input the perception data matrix into the two-stream neural cognitive computing framework, extract multi-scale spatio-temporal features through the adaptive dynamic convolution module, and determine the long-term and short-term dependence relationships between different types by combining the deformable attention mechanism to obtain the perception flow features, including:

[0020] Decouple the spatio-temporal dimensions of the input perception data matrix, and separate to obtain a time series matrix and a spatial feature matrix;

[0021] Perform multi-scale decomposition on the time series matrix through wavelet transform to extract time features at different frequencies; input the spatial feature matrix into a feature adaptive network to generate dynamic convolution kernel weights, perform spatial feature extraction based on the dynamic convolution kernel weights to obtain spatial local features; perform tensor fusion on the time features and the spatial local features to obtain a spatio-temporal feature matrix;

[0022] Calculate the feature similarity based on the spatio-temporal feature matrix, dynamically construct an association graph structure, where the nodes of the association graph represent the feature vectors of different types of sensing data, determine the connection weights between nodes through an adaptive threshold, combine the connection weights with a learnable position offset to dynamically adjust the attention calculation position, and generate a fused feature sequence;

[0023] Divide the fused feature sequence into short-term segments and long-term segments according to the time span, extract local temporal change features from the short-term segments, and extract long-term change trend features from the long-term segments; input the local temporal change features and the long-term change trend features into a memory gated unit to generate a perception flow feature.

[0024] In an alternative embodiment,

[0025] Construct an environmental knowledge reasoning module based on a memory-augmented neural network, combine historical experience knowledge and a recursive neural reasoning unit to mine the causal relationship of environmental changes to obtain a cognitive flow feature, and fuse the perception flow feature and the cognitive flow feature to obtain a fused feature representation including:

[0026] Construct a historical experience knowledge graph including a concept layer, a relationship layer, and an instance layer based on the input perception data matrix, where the concept layer stores the concept mapping information of environmental elements, the relationship layer stores semantic association rules, and the instance layer stores instance feature vectors;

[0027] Input the environmental state feature vector into the memory-augmented neural network, calculate the similarity score based on the node information in the historical experience knowledge graph, extract the historical experience knowledge node with the highest similarity score, calculate the attention weight coefficient, and generate an attention weight matrix;

[0028] Perform importance weighting on environmental variables according to the attention weight matrix, construct a directed acyclic graph in combination with the historical experience knowledge graph, and train the recursive neural reasoning unit;

[0029] Obtain a state transition matrix through the forward propagation of the trained recursive neural reasoning unit to generate a causal evolution feature, and generate an attribution relationship feature matrix through the backward propagation;

[0030] Intervene in the target environmental variables in the control group scenario, calculate the change in the environmental state to determine the causal intensity coefficient, and perform weighted fusion on the causal evolution characteristics and the attribution relationship feature matrix based on the causal intensity coefficient to obtain the cognitive flow characteristics;

[0031] Perform feature fusion on the perceptual flow characteristics extracted from the perceptual data matrix and the cognitive flow characteristics to obtain a fused feature representation.

[0032] In an alternative implementation,

[0033] Input the fused feature representation into the terminal environmental state evaluation model, and the generated environmental evaluation results including the terminal environmental comfort index, safety risk level, and emergency event warning information are as follows:

[0034] Input the fused feature representation into the environmental state evaluation model, calculate the PMV-PPD comfort evaluation index of temperature, humidity, illuminance, noise, and air quality index in the fused feature representation to obtain the environmental comfort index, calculate the deviation degree between the crowd density, hazard source distribution, and equipment operation status and the preset threshold to judge the safety risk level, calculate the abnormality degree of the passenger flow congestion condition, fire hazard index, and abnormal behavior detection result to identify the warning event type and risk degree, and generate emergency event warning information;

[0035] Combine the environmental comfort index, safety risk level, and emergency event warning information, and output the terminal environmental evaluation result.

[0036] In an alternative implementation,

[0037] Input the environmental evaluation result into the spatio-temporal dynamic decoupler, decouple it in the time dimension and space dimension by combining the attention mechanism to obtain the dynamic environmental characteristics, construct a multi-objective optimization model based on the dynamic environmental characteristics, construct a constrained optimization problem through the particle swarm optimization algorithm, combine the adaptive inertia weight and the dynamic convergence factor to solve the preset optimization objective, generate the multi-objective solution result and generate the environmental regulation strategy through the fuzzy mapping algorithm, including:

[0038] Input the environmental evaluation result into the spatio-temporal dynamic decoupler with a two-stream network structure, extract the decoupled features in the time dimension through the bidirectional long short-term memory network of the time flow branch, and extract the decoupled features in the space dimension through the graph convolutional network of the space flow branch;

[0039] Adopt the spatio-temporal dual attention mechanism to calculate the attention weights of the decoupled features, perform tensor multiplication operations on the attention weights and the corresponding decoupled features and add them up to obtain the dynamic environmental characteristics;

[0040] Construct a multi-objective optimization model including an environmental comfort optimization objective, an energy consumption optimization objective, and a device load optimization objective based on the dynamic environmental characteristics;

[0041] Use the particle swarm optimization algorithm to solve the multi-objective optimization model, update the velocity and position of the particle swarm through an adaptive inertia weight and a dynamic convergence factor, and obtain the multi-objective solution result;

[0042] Use the fuzzy mapping algorithm to perform fuzzy mapping on the multi-objective solution result, and generate an environmental regulation strategy including temperature and humidity setting, dimming, ventilation frequency, and device start-stop timing.

[0043] In an alternative embodiment,

[0044] Use the particle swarm optimization algorithm to solve the multi-objective optimization model, update the velocity and position of the particle swarm through an adaptive inertia weight and a dynamic convergence factor, and the multi-objective solution result includes:

[0045] Encode the target decision variables to be optimized into the position vector of the particle, initialize the velocity vector and position vector of the particle swarm, and calculate the multi-objective fitness value of each particle;

[0046] Sort the particle swarm based on the fitness value, select a preset number of particles with high fitness values to construct an elite particle swarm, and record the position vector of the particle with the highest fitness value as the current global optimal position vector;

[0047] In the iterative optimization process, dynamically calculate the adaptive inertia weight according to the current iteration number, and update the dynamic convergence factor through an exponential function;

[0048] For each particle to be updated, randomly select three different elite particles from the elite particle swarm, add the difference between the position vectors of the first elite particle and the other two elite particles multiplied by a scaling factor to obtain a differential evolution candidate position vector; perform Lévy flight search on the global optimal position vector, and add the global optimal position vector and the search step determined by the step factor and Lévy exponent to obtain a Lévy search candidate position vector;

[0049] Calculate the fitness values of the two candidate position vectors, compare them with the current optimal fitness value to select the optimal position vector to update the global optimal solution, and update the velocity and position of the particle swarm based on the updated global optimal position vector, adaptive inertia weight, and dynamic convergence factor until the preset convergence condition is met, and output the decision variables corresponding to the optimal position vector as the multi-objective solution result.

[0050] In the second aspect of the embodiments of the present invention,

[0051] Provide an electronic device, including:

[0052] A processor;

[0053] A memory for storing processor-executable instructions;

[0054] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0055] In the third aspect of the embodiments of the present invention,

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0057] In the present invention, through a dual-stream neurocognitive computing framework, the collaborative processing of the perception stream and the cognitive stream is realized. The adaptive dynamic convolution module can effectively extract multi-scale spatio-temporal features, and the deformable attention mechanism can accurately capture the long-term and short-term dependencies between different perception data, improving the accuracy and comprehensiveness of environmental monitoring. The environmental knowledge reasoning module based on the memory-enhanced neural network combines historical experience knowledge and the recursive neural reasoning unit to effectively mine the causal relationships of environmental changes, enhancing the system's cognitive understanding ability and prediction ability of complex environmental changes, and improving the intelligent level of terminal environmental monitoring. The spatio-temporal dynamic decoupler combines the attention mechanism to achieve the accurate decoupling of environmental features in the time and space dimensions. The multi-objective optimization model can efficiently solve the environmental regulation problem through the improved particle swarm algorithm. The adaptive inertia weight and the dynamic convergence factor improve the optimization efficiency, making the environmental regulation strategy more scientific and reasonable, and significantly improving the automation and intelligent level of terminal environmental management. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic flow chart of the intelligent terminal environment monitoring method combining multi-source data fusion and deep learning according to the embodiments of the present invention;

[0059] Figure 2 It is a comparison chart of the energy consumption optimization effect of the intelligent terminal environment monitoring method combining multi-source data fusion and deep learning according to the embodiments of the present invention;

[0060] Figure 3 It is a comparison chart of the device load balance degree of the intelligent terminal environment monitoring method combining multi-source data fusion and deep learning according to the embodiments of the present invention;

[0061] Figure 4 It is a comparison chart of the convergence performance of the intelligent terminal environment monitoring method combining multi-source data fusion and deep learning according to the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0063] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0064] Figure 1 is a schematic flowchart of the method according to the embodiment of the present invention, as Figure 1 shown, the method includes:

[0065] Collect multi-source perception data in the terminal building, perform preprocessing of time series alignment and spatial registration, generate a standardized perception data matrix, and segment the perception data matrix;

[0066] Input the perception data matrix into a two-stream neural cognitive computing framework, extract multi-scale spatio-temporal features through an adaptive dynamic convolution module, determine the long-term and short-term dependence relationships between different types by combining a deformable attention mechanism to obtain perception stream features, construct an environmental knowledge reasoning module based on a memory-enhanced neural network, mine the causal relationships of environmental changes by combining historical experience knowledge and a recursive neural reasoning unit to obtain cognitive stream features, and fuse the perception stream features and cognitive stream features to obtain a fused feature representation;

[0067] Input the fused feature representation into a terminal building environment state evaluation model to generate an environmental evaluation result including the terminal building environmental comfort index, safety risk level, and emergency event warning information;

[0068] Input the environmental evaluation result into a spatio-temporal dynamic decoupler, decouple it in the time dimension and space dimension by combining an attention mechanism to obtain dynamic environmental features, construct a multi-objective optimization model based on the dynamic environmental features, construct a constrained optimization problem through a particle swarm optimization algorithm, combine an adaptive inertia weight and a dynamic convergence factor to solve the preset optimization objectives, generate a multi-objective solution result, and generate an environmental regulation strategy through a fuzzy mapping algorithm;

[0069] Send the environmental regulation strategy to the terminal building environmental control system for execution.

[0070] In an alternative embodiment,

[0071] Collect multi-source perception data in the terminal building, perform preprocessing of time series alignment and spatial registration, generate a standardized perception data matrix, and segment the perception data matrix, including:

[0072] Collect multi-source data in the terminal building through temperature and humidity sensors, air quality sensors, video surveillance devices, and noise sensors;

[0073] Synchronize and align the multi-source data according to timestamps, and realize the registration of sensing data at different positions based on spatial coordinate mapping. Generate a perception data matrix in a unified format through standardization processing;

[0074] Adopt the sliding time window technology, set the initial sliding window according to the preset time step, and dynamically segment the perception data matrix in combination with the data change trend.

[0075] Carry out multi-source data collection work and deploy various types of sensing devices inside the terminal building. Temperature and humidity sensors are arranged in main functional areas such as the waiting area, check-in area, and dining area to monitor the temperature and humidity changes in each area in real time. Air quality sensors are distributed in crowded areas and near air conditioner outlets to monitor indicators such as carbon dioxide, particulate matter, and volatile organic compounds. Video surveillance devices are installed in passages, security check areas, and public areas to collect the crowd density and personnel activity status. Noise sensors are deployed at key positions in the terminal building to collect environmental noise data. The sensing devices continuously collect data according to the preset sampling frequency and transmit the data to the central data processing platform through wired or wireless networks.

[0076] Perform time series alignment processing to solve the problem of inconsistency of multi-source data in the time dimension, add a unified timestamp mark to each piece of data, and sort all data according to the timestamp. Due to the differences in the sampling frequencies of different types of sensors, data resampling is required to unify all data to the same time scale. For sensor data with a higher sampling frequency, downsampling is performed using the average value within the time window; for sensor data with a lower sampling frequency, missing time point data is supplemented by linear interpolation methods. In this way, the synchronization alignment of all sensing data in time is achieved.

[0077] Perform spatial registration processing to solve the problem of inconsistency of different sensors in spatial distribution. Establish a three-dimensional space coordinate system of the terminal building and map the position of each sensor into this coordinate system. Based on the spatial distribution of the sensors, the entire terminal building is divided into several monitoring areas. For different types of sensor data within the same monitoring area, a spatial weight fusion method is used for registration. For sensor data at the boundary of the monitoring area, a spatial interpolation algorithm is used to perform smooth transition processing on the data. Ensure that sensing data from different positions can be analyzed under a unified spatial reference system.

[0078] After completing the temporal alignment and spatial registration, standardize all the collected data. For different types of sensing data, perform outlier detection and elimination to remove abnormal data points that may be caused by sensor failures or interference. Normalize data with different dimensions, convert various types of data into a unified numerical range, and eliminate the dimensional differences between the data. Organize the processed data into a standardized perception data matrix according to a predefined data structure, where the rows of the matrix represent different time points, the columns represent different sensing parameters, and the matrix elements are the corresponding sensing values.

[0079] Perform dynamic segmentation on the standardized perception data matrix. Adopt the sliding time window technique and set the initial window size and sliding step according to the characteristics of environmental changes. When the data within the window changes smoothly, keep the current window size; when a data mutation or an obvious change trend is detected, dynamically adjust the window size to capture the change characteristics. By calculating the coefficient of variation of the data within the window, when the coefficient of variation exceeds the preset threshold, reduce the window size to improve the segmentation accuracy; when the coefficient of variation continuously remains below the threshold, appropriately increase the window to improve the calculation efficiency. Through the dynamically adjusted sliding window technique, divide the complete perception data matrix into several sub-matrix segments, which is convenient for subsequent feature extraction and pattern recognition.

[0080] Exemplarily, taking Terminal 2 of a certain international airport as an example, 150 temperature and humidity sensors are deployed inside the terminal building, and data is collected every 5 minutes; 80 air quality sensors, and data is collected every 10 minutes; 60 high-definition video monitoring devices, which collect images in real time and generate crowd density data every minute; 45 noise sensors, and environmental noise data is collected every 3 minutes.

[0081] In the temporal alignment process, unify all sensor data to a 10-minute sampling interval. For temperature and humidity data, take the average value of every two adjacent sampling points; for crowd density data, take the average value within 10 minutes; for noise data, supplement the missing time points through linear interpolation. After processing, all sensing data is synchronized and aligned in the time dimension, forming a unified time series with a 10-minute sampling interval.

[0082] In the spatial registration process, divide the entire area into 25 monitoring sub-regions according to the physical layout of the terminal building. For each sub-region, calculate the comprehensive environmental parameters of the region using the distance weighted average method according to the spatial distribution and coverage of the sensors. For example, for waiting area A, 6 temperature and humidity sensors, 3 air quality sensors, 2 video devices, and 2 noise sensors are deployed inside. Through spatial registration, the data of these sensors are fused into a unified environmental state representation of this region.

[0083] During data standardization processing, anomaly detection is performed on the original data. For example, if it is found that the temperature sensor in waiting area B suddenly outputs an abnormal value of 45°C during a certain period, which is significantly higher than the normal range, it will be identified as an anomaly and excluded. Different types of sensor data are normalized, and data with different dimensions such as temperature data (15°C - 30°C), humidity data (30% - 80%), carbon dioxide concentration (400 ppm - 2000 ppm), and noise decibels (40 dB - 85 dB) are uniformly converted to the numerical range of 0 - 1. A standardized perception data matrix is formed, and the matrix size is 144 rows (10-minute sampling points in a day) × 100 columns (25 areas × 4 environmental parameters).

[0084] In the dynamic segmentation process, the initial sliding window size is set to 6 sampling points (i.e., 1 hour), and the sliding step is 1 sampling point (10 minutes). When it is detected that the changes in the crowd density and environmental parameters are drastic during the morning and evening rush hours (such as 7:00 - 9:00 and 17:00 - 19:00), the window size is automatically adjusted to 3 sampling points (30 minutes) to more finely capture the characteristics of environmental changes; during the low peak period at night (such as 23:00 - 5:00) when the environmental parameters change gently, the window size is automatically adjusted to 12 sampling points (2 hours) to improve the calculation efficiency. The data of one day is dynamically segmented into multiple time segments of unequal lengths, and the environmental changes within each segment have relatively consistent characteristics.

[0085] In this embodiment, the collaborative acquisition of multi-source heterogeneous data greatly improves the comprehensiveness and accuracy of environmental perception. Through timestamp synchronization and resampling processing, the time misalignment problem caused by inconsistent sampling frequencies of different sensors is eliminated, significantly enhancing the consistency and comparability of the data, and laying a solid foundation for subsequent feature extraction and pattern recognition.

[0086] In an alternative embodiment,[

[0087] The perception data matrix is input into a two-stream neural cognitive computing framework. Multi-scale spatio-temporal features are extracted through an adaptive dynamic convolution module, and the long-term and short-term dependencies between different types are determined by combining a deformable attention mechanism to obtain perception stream features, including:

[0088] Decouple the spatio-temporal dimensions of the input perception data matrix, and separate it into a time series matrix and a spatial feature matrix;

[0089] Perform multi-scale decomposition on the time series matrix through wavelet transform to extract time features at different frequencies; input the spatial feature matrix into a feature adaptive network to generate dynamic convolution kernel weights, and perform spatial feature extraction based on the dynamic convolution kernel weights to obtain spatial local features; perform tensor fusion on the time features and the spatial local features to obtain a spatio-temporal feature matrix;

[0090] Calculate the feature similarity based on the spatio-temporal feature matrix, and dynamically construct an association graph structure. The nodes of the association graph represent the feature vectors of different types of sensing data. Determine the connection weights between nodes through an adaptive threshold, combine the connection weights with a learnable position offset, dynamically adjust the attention calculation position, and generate a fused feature sequence;

[0091] Divide the fused feature sequence into short-term segments and long-term segments according to the time span. Extract local temporal change features from the short-term segments and long-term change trend features from the long-term segments; input the local temporal change features and the long-term change trend features into a memory gated unit to generate a perceptual flow feature.

[0092] Perform spatio-temporal dimensional decoupling on the input perceptual data matrix using a tensor decomposition method. Reconstruct the original data into a matrix representing the sensor time series and a feature matrix representing the spatial distribution through matrix reconstruction. The time series matrix retains the complete information of the sensing data changing over time, and the spatial feature matrix contains spatial attribute information such as the physical location and coverage range of the sensors.

[0093] Perform multi-scale decomposition of wavelet transform on the time series matrix. Select wavelet basis functions and project the time series data into different frequency subspaces. Extract rapidly changing transient features, including emergency events and short-term fluctuation information, in the high-frequency subspace; obtain periodic change features, such as daily change patterns, in the middle-frequency subspace; and extract slowly changing long-term trend features in the low-frequency subspace. Obtain a multi-scale time feature representation by reconstructing the time features at different frequencies.

[0094] Input the spatial feature matrix into a feature adaptive network, which includes multiple layers of feature extraction modules. Each module dynamically generates convolution kernel weight parameters of different scales and shapes according to the statistical distribution characteristics of the input features. Use these adaptively generated convolution kernels to perform multi-level extraction of the spatial features, and gradually obtain spatial feature representations from local to global. Perform weighted combination on the extracted multi-layer spatial features to obtain a spatial local feature vector. Combine the multi-scale time features and the multi-layer spatial features through tensor fusion operations to construct a feature matrix containing complete spatio-temporal association information.

[0095] Calculate the similarity relationship between features based on the spatio-temporal feature matrix. Use measurement methods such as cosine similarity to calculate the correlation degree between different feature vectors, and set an adaptive threshold to determine the association strength between features. Dynamically construct an association graph structure according to the calculated similarity values. The nodes in the graph represent the feature vectors of different types of sensing data, and the weights of the edges are determined by the similarity values. At the same time, introduce a learnable position encoding vector to finely adjust the position of attention calculation. Through the graph attention mechanism, perform information transfer and feature fusion on the node features to generate a fused feature sequence considering global relevance.

[0096] Divide the fused feature sequence according to different time spans. For short-time segments (such as several minutes to several hours), focus on extracting temporal change patterns such as mutation features and periodic features within a local range; for long-time segments (such as several days to several weeks), extract macroscopic features such as long-term change trends and seasonal changes. Input the features extracted at different time scales into a gated unit with a selective memory mechanism. The gated unit can selectively retain and update information according to the importance of the features, and output a perceptual flow feature containing multi-scale temporal information.

[0097] Exemplarily, taking the airport terminal environment monitoring system as an example, various types of sensors such as temperature, humidity, carbon dioxide, passenger flow, and noise are deployed in different functional areas such as the waiting hall, security check area, boarding gate, and business area, constructing an environmental perception network covering the entire terminal.

[0098] Decompose the collected perceptual data matrix. The time series data contains the environmental parameter values recorded every 30 seconds, and the spatial feature data contains the distribution position information of the sensors in each area of the terminal, including three-dimensional coordinates, coverage range, etc.

[0099] Perform wavelet decomposition on the time series of the temperature sensors in the terminal. In the high-frequency band, instantaneous temperature fluctuations caused by the opening and closing of security check gates and changes in passenger flow at boarding gates can be extracted; in the middle-frequency band, the temperature change law during the peak flight takeoff and landing periods can be identified; in the low-frequency band, the long-term temperature trend caused by day-night alternation and weather changes can be obtained. Dynamically generate convolution kernels according to the spatial characteristics of different areas of the terminal. For example, use large-scale convolution kernels to extract overall features in the open waiting hall, and use small-scale convolution kernels to extract local features in the relatively enclosed boarding gate passage.

[0100] In the constructed association graph, the temperature sensors and passenger flow sensors in the area near the boarding gate have relatively large connection weights due to strong data correlation; the carbon dioxide concentration and passenger flow density sensors in the waiting area show significant synchronization during flight delay periods, and the corresponding attention weights will be enhanced accordingly. Through the adjustment of position encoding, it is also possible to focus on the changes in sensor data at key positions such as the entrances and exits of the security check area and the gathering places in the business area.

[0101] Divide the data of one day into short - time segments of 10 minutes and long - time segments of 2 hours. The short - time segments can capture the impact of sudden changes in passenger flow brought by flights arriving and departing on the environment; the long - time segments can reflect the adjustment effect of the central air - conditioning system and the passenger flow pattern at different times. These multi - scale features are processed by a gating unit to form complete terminal environment perception features.

[0102] In this embodiment, through the method of spatio - temporal dimension decoupling and multi - scale feature extraction, the time evolution law and spatial distribution characteristics in the perception data can be effectively separated and retained. The adaptive mechanism makes the spatial feature extraction more flexible, can fully consider the environmental characteristics of different regions, improves the pertinence and accuracy of feature extraction. Based on the graph - structure - based feature fusion method, the correlation information between data can be fully utilized to improve the integrity of feature expression.

[0103] In an alternative embodiment,

[0104] Construct an environmental knowledge reasoning module based on a memory - enhanced neural network. Combine historical experience knowledge and a recursive neural reasoning unit to mine the causal relationship of environmental changes to obtain cognitive flow features. The fusion of perceptual flow features and cognitive flow features to obtain a fused feature representation includes:

[0105] Construct a historical experience knowledge graph containing a concept layer, a relationship layer, and an instance layer based on the input perceptual data matrix. The concept layer stores the concept mapping information of environmental elements, the relationship layer stores semantic association rules, and the instance layer stores instance feature vectors;

[0106] Input the environmental state feature vector into the memory - enhanced neural network. Calculate the similarity score based on the node information in the historical experience knowledge graph, extract the historical experience knowledge node with the highest similarity score, calculate the attention weight coefficient, and generate an attention weight matrix;

[0107] Perform importance weighting on environmental variables according to the attention weight matrix, construct a directed acyclic graph in combination with the historical experience knowledge graph, and train the recursive neural reasoning unit;

[0108] Obtain a state transition matrix through the forward propagation of the trained recursive neural reasoning unit to generate causal evolution features, and generate an attribution relationship feature matrix through backpropagation;

[0109] Intervene in the target environmental variable in the control group scenario, calculate the environmental state change amount to determine the causal intensity coefficient, and perform weighted fusion on the causal evolution features and the attribution relationship feature matrix based on the causal intensity coefficient to obtain cognitive flow features;

[0110] Fuse the perceptual flow features extracted from the perceptual data matrix with the cognitive flow features to obtain a fused feature representation.

[0111] Construct a multi-level historical experience knowledge graph. In the concept layer, perform semantic concept mapping on environmental elements, classify and define environmental parameters such as temperature, humidity, and air quality according to dimensions such as physical attributes, numerical ranges, and measurement units. The concept nodes of each environmental element contain information such as its basic attribute description, parameter thresholds, and influencing factors. In the relationship layer, construct a semantic association rule base for environmental elements based on domain knowledge and historical data analysis. Semantic association rules include various types such as direct causal relationships, conditional trigger relationships, and temporal dependence relationships, and each relationship is accompanied by detailed semantic descriptions and constraint conditions. In the instance layer, store the feature vectors obtained after feature extraction from historical data, and each feature vector is labeled with corresponding scenario information, time stamps, and environmental conditions.

[0112] After obtaining the feature vector of the current environmental state, input it into a memory-augmented neural network for processing. The network normalizes and aligns the input features and calculates similarity scores with all historical experience nodes stored in the knowledge graph. The similarity calculation adopts a multi-dimensional matching method, including numerical similarity, temporal pattern similarity, and semantic similarity, etc. Based on the calculated similarity scores, select several historical experience knowledge nodes with the highest scores. For these high-similarity nodes, calculate their importance to the current state through an attention mechanism to generate an attention weight coefficient reflecting the node's influence. Organize the weight coefficients into an attention weight matrix for subsequent feature weighting.

[0113] Use the attention weight matrix to rank and weight the environmental variables according to their importance. Environmental variables with higher weight values are considered to have stronger explanatory and predictive value for the current state. Combine the association rules in the historical experience knowledge graph and construct a directed acyclic graph based on the weighted environmental variables. The nodes in the graph represent environmental variables, the edges represent the dependence relationships between variables, and the direction of the edges indicates the transmission direction of the causal relationship. Train a recursive neural inference unit on this graph structure so that it can simulate the chain reaction process between environmental elements, and through multiple rounds of iterative optimization, enable the inference unit to accurately depict the evolution law of the environmental state.

[0114] Use the trained recursive neural inference unit for bidirectional inference. In the forward propagation process, activate the nodes in the directed acyclic graph in sequence, calculate the transition probability of the node states, and generate a complete state transition matrix. The matrix describes the possible change trends of environmental elements over time. In the backward propagation process, starting from the target state node, trace back along the edges of the graph in reverse to generate an attribution relationship feature matrix. This matrix records the cause chain leading to the formation of the current environmental state.

[0115] An intervention experiment is carried out in a specially configured control group scenario. By artificially adjusting the values of target environmental variables, the response changes of other environmental elements are observed. The differences in environmental states before and after the intervention are recorded, and a causal strength evaluation model is established. The causal strength coefficient output by the model reflects the degree of influence between different environmental elements. These strength coefficients are used to perform weighted fusion on the previously obtained causal evolution characteristics and attribution relationship characteristics, generating cognitive flow characteristics containing complete causal cognitive information.

[0116] Deeply fuse the perceptual flow characteristics and the cognitive flow characteristics. Adopt a multi-level feature fusion strategy to align and combine the two types of characteristics at different semantic levels. The fusion process takes into account the temporal dependence, spatial correlation, and semantic consistency of the characteristics, ensuring that the generated fusion characteristics not only retain the detailed information of the original perceptual data but also contain the cognitive information accumulated from historical experience. The obtained fusion characteristics have rich expressive capabilities and can comprehensively depict the current characteristics and evolution trends of the environmental state.

[0117] Exemplarily, taking the HVAC system in an airport terminal as an example, the constructed historical experience knowledge graph stores the defined attributes of basic environmental parameters such as temperature, humidity, and fresh air volume at the concept level; stores semantic rules such as "an increase in temperature leads to a decrease in relative humidity" and "an increase in the number of people causes an increase in carbon dioxide concentration" at the relationship level; and stores the historical environmental parameter feature vectors during different flight takeoff and landing periods and different weather conditions at the instance level.

[0118] When a temperature anomaly occurs in a certain area of the terminal, the environmental state characteristics of this area are input into the memory-enhanced network. The network discovers similar cases in historical data through calculating similarity, such as the scenario of "a large number of flight delays in the afternoon in summer leading to the gathering of people in the waiting hall". Higher attention weights are assigned to such historical experience nodes.

[0119] Rank the importance of environmental variables according to the attention weights, and identify that temperature, population density, and carbon dioxide concentration are key variables. Combine the knowledge graph to construct a directed acyclic graph to describe the causal chain of "people gathering - increase in carbon dioxide concentration - increase in temperature - decrease in relative humidity", and train a recursive neural inference unit to understand this causal relationship.

[0120] Through the forward derivation of the inference unit, predict the possible chain reactions caused by the continuous increase in temperature; through the backward derivation, analyze that the root cause of the temperature anomaly is the heat accumulation caused by the gathering of people. Conduct an intervention experiment in the control area by adjusting the fresh air volume and other means to verify the causal strength between environmental elements, and finally obtain the complete cognitive flow characteristics.

[0121] Fuse the perceptual flow characteristics (such as temperature change curves, pedestrian flow data, etc.) obtained from real-time monitoring with the cognitive flow characteristics to form a comprehensive understanding of the current situation, providing a decision-making basis for the intelligent adjustment of the air conditioning system.

[0122] In this embodiment, the structured separation of perception data is achieved through spatio-temporal dimension decoupling. The multi-scale decomposition method of wavelet transform is adopted, which can comprehensively capture various time features from instantaneous fluctuations to long-term trends, overcoming the problem of incomplete feature extraction under the traditional single time scale. The convolution kernel weights are dynamically generated through the feature adaptive network, enhancing the flexibility of spatial feature extraction and enabling the system to adaptively adjust the feature extraction strategy according to the environmental characteristics of different regions.

[0123] In the prior art, the airport terminal environment perception system mainly adopts a single time-series feature extraction method or a fixed spatial feature extraction mode, which is difficult to take into account both the time dynamics and spatial differences of environmental parameters at the same time. The time features and spatial features are separated and processed separately, resulting in incomplete feature extraction and unable to fully reflect the correlation between environmental parameters. Due to the lack of an adaptive feature extraction mechanism, it is difficult to cope with the characteristics of drastic fluctuations in the number of passengers and complex functional partitions in the airport terminal, affecting the accuracy and real-time performance of environmental perception.

[0124] This embodiment can timely capture the sudden changes of environmental parameters in the terminal, accurately identify the environmental characteristics of different regions, effectively establish the correlation between environmental parameters, provide a more reliable data basis for terminal environmental regulation, improve the adaptability of the system to complex and changeable scenarios, effectively enhance the intelligent level of terminal environmental management, and provide strong support for optimizing the passenger service experience and improving the operation efficiency.

[0125] Figure 2 This is a comparison chart of the energy consumption optimization effect of the multi-source data fusion combined with deep learning-based terminal environment intelligent monitoring method according to the embodiments of the present invention. The horizontal axis represents four different operating scenarios (daily operation, high temperature weather, peak passenger flow, and equipment overload), and the vertical axis represents the percentage of energy consumption reduction.

[0126] It can be clearly seen from the figure that the technical solution of the present invention (triangle mark) shows the best energy optimization effect in all scenarios. In the daily operation scenario, the technical solution of the present invention can reduce the energy consumption by 24.3%, while the model predictive control, feedback control, and timed control can only reduce the energy consumption by 18.7%, 14.2%, and 11.4% respectively. As the scenario complexity increases, the advantage of the technical solution of the present invention becomes more obvious. Especially in the equipment overload scenario, the technical solution of the present invention can reduce the energy consumption by 40.7%, far higher than 31.9%, 26.3%, and 21.5% of other methods.

[0127] This technical solution constructs an environmental knowledge reasoning module based on a memory-augmented neural network, combines a historical experience knowledge graph and a recursive neural reasoning unit, and can accurately identify the complex causal relationship between environmental variables and equipment load, so as to achieve more accurate energy regulation. Although model predictive control (MPC algorithm, square marker) performs rolling optimization on the future behavior of the system through a prediction model, it lacks an understanding of the causal relationship of environmental changes, resulting in lower energy optimization efficiency than this technical solution. Feedback control (PID control algorithm, circular marker) adjusts system parameters in real time through a proportional-integral-derivative controller, but its energy optimization effect is limited because it only depends on the error signal and cannot predict environmental changes. Timing control (static scheduling algorithm based on time series, hexagonal marker) uses a preset time schedule to control the on / off of equipment in a fixed mode and cannot adapt to dynamic environmental changes, so it performs the worst in complex scenarios.

[0128] It is particularly worth noting that in abnormal situations such as equipment overload, this technical solution improves the energy optimization efficiency by nearly twice compared with timing control (40.7% vs. 21.5%), fully demonstrating the adaptability and high efficiency of this solution under extreme working conditions. This technical solution can identify similar situations through historical experience data in the knowledge graph and use the recursive neural reasoning unit to predict the potential change trend of system load, so as to adjust the equipment operation parameters in advance and avoid energy waste. Through the causal reasoning ability of the directed acyclic graph, it can also distinguish the primary and secondary factors of environmental changes, perform precise control on key variables, and achieve a more efficient energy management strategy.

[0129] In an alternative embodiment,

[0130] Inputting the fused feature representation into the terminal building environment state evaluation model to generate an environmental evaluation result including the terminal building environmental comfort index, safety risk level, and emergency event warning information includes:

[0131] Inputting the fused feature representation into the environmental state evaluation model, calculating the PMV-PPD comfort evaluation index of temperature, humidity, illuminance, noise, and air quality index in the fused feature representation to obtain the environmental comfort index, calculating the deviation degree between the crowd density, hazard source distribution, and equipment operation state and the preset threshold to judge the safety risk level, and calculating the abnormality degree of the passenger flow congestion condition, fire hazard index, and abnormal behavior detection result to identify the type and risk degree of the warning event, and generating emergency event warning information;

[0132] Combining the environmental comfort index, safety risk level, and emergency event warning information, and outputting the terminal building environmental evaluation result.

[0133] After inputting the fused feature representation into the environmental state assessment model, a comprehensive assessment of the environmental comfort is carried out. By calculating the deviation value between the temperature index in the fused feature and the human thermal balance state, and combining with the correction coefficient of the influence of relative humidity on human heat dissipation, the comprehensive comfort of temperature and humidity is evaluated. At the same time, the illuminance level is compared with the comfortable illuminance range of the human eye to calculate the comfort degree of the lighting environment. The noise index is subjected to spectral analysis and loudness assessment, considering the weight of the influence of noise in different frequency bands on the human body. By performing multi-parameter combination analysis on the air quality index, the influence degree of the air environment on the human body is evaluated. These sub-item indexes are weighted and combined to finally generate a quantitative index reflecting the overall environmental comfort.

[0134] In the safety risk assessment link, key indicators that may affect the safe operation of the terminal are focused on. By analyzing the matching degree between the pedestrian flow density distribution and the evacuation passage capacity, the regional congestion risk is evaluated. The spatial distribution of hazard sources is dynamically monitored to evaluate the distance relationship and influence range with important functional areas. The deviation of the equipment operation parameters from the preset safety thresholds is tracked in real time, including indicators such as power consumption load, pipeline pressure, and equipment temperature. According to the comprehensive deviation degree of these indicators, the safety risk level is divided.

[0135] For emergency event early warning, a multi-dimensional anomaly detection framework is constructed. By analyzing the spatio-temporal distribution characteristics of the passenger flow, abnormal aggregation and rapid evacuation phenomena are identified. The fire hazard indicators such as smoke detector data and infrared thermal imaging data are monitored in real time, and the fire risk is judged in combination with the trend of environmental temperature change. Video analysis technology is used to detect suspicious behaviors, including abnormal residence time, reverse movement, and left items. Based on these anomaly detection results, the type and risk level of the early warning event are determined, and the corresponding early warning information is generated.

[0136] The environmental comfort index, safety risk level, and emergency early warning information are integrated. A multi-level information organization method is adopted to sort different types of assessment results according to importance and urgency. Through information aggregation, a structured environmental assessment report is formed to intuitively display the environmental conditions and potential risks of each area of the terminal.

[0137] Exemplarily, taking the T2 terminal of a certain airport as an example, after the fusion features are input into the evaluation model, the environmental comfort of the waiting area is first evaluated. Through analysis, it is found that due to the strong solar radiation in the afternoon, the temperature in the area near the glass curtain wall reaches a relatively high level. Combined with the relatively low relative humidity, the thermal comfort in this area is reduced. The illuminance analysis shows that due to the reflective effect of the glass curtain wall, glare phenomena occur in some seat areas. The noise level during the peak flight takeoff and landing period rises moderately but does not exceed the comfort threshold. The carbon dioxide concentration in the air quality index increases with the increase in the number of people, but it remains within the comfort range after being adjusted by the fresh air system. The comfort index generated by the comprehensive evaluation shows that the overall area is at a relatively comfortable level, but the area near the window needs improvement.

[0138] In the safety risk assessment, it is monitored that the pedestrian flow density in the security check area increases. After comparing with the threshold of the passage evacuation capacity, it is determined to be at a medium risk level. The analysis of the distribution of hazard sources finds that the operating temperature of the mechanical equipment in the baggage conveyor area is slightly abnormal, but it does not reach the high-risk threshold. The monitoring of the equipment operating status shows that the refrigeration load of the central air-conditioning system is close to the warning value and needs to be focused on.

[0139] In the emergency warning analysis, through passenger flow monitoring, it is found that there is a temporary congestion in the check-in counter area, but it can be effectively diverted by opening the standby passage. In the fire hazard detection, the data of the smoke detectors in the dining area are normal, and no abnormal heat sources are found by thermal imaging monitoring. The behavior analysis system detects that there are unclaimed baggage in the baggage storage area for an overdue time, and a low-level abnormal behavior warning is initiated.

[0140] The generated evaluation results show that: the overall environment of the terminal is in good condition, and the comfort index is within the appropriate range; the safety risk level is medium, and the main concerns are the control of pedestrian flow density and the operating status of equipment; the warning information includes two low-risk prompts of temporary congestion and overdue baggage storage.

[0141] In this embodiment, there is an evaluation framework for multi-dimensional fusion. By introducing a comprehensive analysis method of multiple indicators such as temperature, humidity, illuminance, noise, and air quality, the accurate quantification of environmental comfort is achieved. Through the correlation analysis of multi-dimensional data such as pedestrian flow density, hazard source distribution, and equipment operating status, the accuracy and comprehensiveness of risk assessment are improved. Through the comprehensive analysis of multiple dimensions such as passenger flow congestion status, fire hazard indicators, and abnormal behaviors, the early identification of different types of potential risks is achieved.

[0142] In an alternative implementation manner,

[0143] Input the environmental assessment results into a spatio-temporal dynamic decoupler, decouple in the time dimension and space dimension by combining the attention mechanism to obtain dynamic environmental features, construct a multi-objective optimization model based on the dynamic environmental features, construct a constrained optimization problem through the particle swarm optimization algorithm, and solve the preset optimization objectives by combining the adaptive inertia weight and the dynamic convergence factor to generate a multi-objective solution result and generate an environmental regulation strategy through the fuzzy mapping algorithm, including:

[0144] Input the environmental assessment results into the spatio-temporal dynamic decoupler of the two-stream network structure, extract the decoupling features in the time dimension through the bidirectional long short-term memory network of the time stream branch, and extract the decoupling features in the space dimension through the graph convolutional network of the space stream branch;

[0145] Adopt the spatio-temporal dual attention mechanism to calculate the attention weights of the decoupled features, perform tensor multiplication operations on the attention weights and the corresponding decoupled features and add them to obtain the dynamic environmental features;

[0146] Construct a multi-objective optimization model based on the dynamic environmental features, including the optimization objective of environmental comfort, the optimization objective of energy consumption, and the optimization objective of equipment load;

[0147] Adopt the particle swarm optimization algorithm to solve the multi-objective optimization model, update the velocity and position of the particle swarm through the adaptive inertia weight and the dynamic convergence factor to obtain the multi-objective solution result;

[0148] Adopt the fuzzy mapping algorithm to perform fuzzy mapping on the multi-objective solution result to generate an environmental regulation strategy including temperature and humidity setting, dimming, ventilation frequency, and equipment start-stop timing.

[0149] Input the environmental assessment results into the spatio-temporal dynamic decoupler with a two-stream structure, and perform feature extraction through two branches of the time stream and the space stream respectively. In the time stream branch, use the bidirectional long short-term memory network to perform temporal analysis on the environmental assessment results, capture the temporal evolution characteristics of the environmental state from both the forward and reverse directions. The network saves historical information through memory units, and at the same time screens important temporal features through the forgetting gate mechanism, and finally extracts the decoupled features reflecting the temporal change law of the environmental state.

[0150] In the space stream branch, use the graph convolutional network to process the spatial correlation information in the environmental assessment results. First, construct the environmental states of different regions of the terminal into a graph structure, where the nodes represent regions and the edges represent the spatial correlation relationships between regions. Through multi-layer graph convolutional operations, spatial features are extracted layer by layer to capture the mutual influence and transmission effects between different regions, and the decoupled features representing the spatial distribution characteristics of the environmental state are obtained.

[0151] Calculate the attention weights for the obtained time - dimension decoupled features and space - dimension decoupled features respectively. In the time dimension, determine the importance of features at each time point according to the correlation between the current environmental state and the historical state. In the space dimension, allocate attention weights based on the influence degree of different regions on the target region. Perform a tensor multiplication operation on the calculated attention weights and the corresponding decoupled features, and add the weighted features in the two dimensions to generate environmental features containing spatio - temporal dynamic information.

[0152] Construct a multi - objective optimization model based on the dynamic environmental features. The model simultaneously considers the optimization objectives of environmental comfort, energy consumption, and equipment load. The environmental comfort objective focuses on optimizing the matching degree of parameters such as temperature and humidity with the comfort interval; the energy consumption objective pays attention to the energy utilization efficiency of various devices; the equipment load objective balances the operating loads of each device to avoid local overload.

[0153] Use an improved particle swarm algorithm to solve the multi - objective optimization model. Represent different environmental regulation schemes by setting the initial positions and velocities of the particle swarm. In the iterative optimization process, adaptively adjust the inertia weight of the particles according to the dynamic changes of the environmental state, and at the same time use a dynamic convergence factor to guide the search direction of the particles. Find the optimal solution set that meets multiple optimization objectives through multiple rounds of iteration.

[0154] Input the result set obtained from the optimization solution into the fuzzy mapping algorithm for processing. Based on the pre - established fuzzy rule base, convert the numerical optimization results into specific regulation instructions. The fuzzy mapping process considers the continuity of environmental parameters and the feasibility of equipment adjustment, and generates an environmental regulation strategy including temperature and humidity set values, lighting dimming parameters, fresh air ventilation frequency, and the start - stop sequence of various devices.

[0155] Exemplarily, taking the environmental regulation of the airport waiting area as an example, input the environmental assessment results of this area into the spatio - temporal dynamic decoupler. Through analysis, the time - flow branch finds that the temperature change in this area shows obvious periodic characteristics, with the temperature rising slowly in the morning and gradually decreasing after reaching the peak in the afternoon. The space - flow branch identifies a significant temperature gradient between the area near the glass curtain wall and the inner area, and also discovers the temperature transfer effect between adjacent waiting areas.

[0156] Through the calculation of the attention mechanism, it is found that the time - feature weight is higher in the afternoon, which is related to the change in solar radiation intensity. In the space dimension, the area near the glass curtain wall obtains a higher attention weight, indicating that this area is the key point of environmental regulation. Based on the weighted dynamic environmental features, an optimization model is constructed with the goals of improving passenger comfort, reducing air - conditioning energy consumption, and balancing equipment load.

[0157] Through repeated iteration, the particle swarm optimization algorithm found a set of relatively optimal control schemes. After fuzzy mapping transformation, specific control strategies were generated: during the high-temperature period in the afternoon, the air outlet temperature of the air conditioner in the area near the window was reduced in advance, and at the same time, the influence of solar radiation was reduced by adjusting the sunshade; the fresh air volume was dynamically adjusted according to the change of the crowd density to maintain the best air quality; the lighting system was gradually dimmed with the change of the natural light intensity to avoid sudden changes in brightness and darkness. The start and stop of various devices were arranged at different times to avoid load peaks caused by simultaneous startup.

[0158] In this embodiment, the dual-stream network structure is used to capture the environmental change characteristics in the time dimension and the space dimension respectively, breaking through the limitation of the mutual separation of spatio-temporal feature analysis in the traditional method. The introduction of the dual attention mechanism enables the feature extraction process to adaptively focus on important time nodes and key space regions, improving the accuracy and pertinence of the environmental feature expression. Through the improved particle swarm optimization algorithm, the trade-off problem between multiple objectives is effectively solved. The design of the adaptive inertia weight and the dynamic convergence factor significantly improves the convergence efficiency and the quality of the solution of the optimization algorithm, making the generated control scheme more practical;

[0159] In the prior art, time features and space features are often separated and processed independently, unable to accurately grasp the spatio-temporal coupling law of environmental changes. The generation of control strategies lacks a global optimization concept, and often uses single-objective or fixed-weight optimization methods, making it difficult to balance multiple objectives such as comfort, energy consumption, and equipment load. The execution of control instructions lacks flexibility, and it is difficult to adapt to the complex and changeable actual environment with a preset control logic;

[0160] This embodiment significantly reduces energy waste through optimized scheduling, the regulation of environmental parameters is more accurate, it can actively adapt to environmental changes and maintain the best state, the equipment runs more stably, avoiding load peaks and frequent start and stop. While ensuring the comfortable experience of passengers, it realizes the effective control of operating costs, providing an innovative technical path for the intelligent environmental management of large public buildings.

[0161] Figure 3 This is a comparison chart of the equipment load balance degree of the intelligent monitoring method for terminal building environment with multi-source data fusion combined with deep learning in the embodiment of the present invention, showing the comparison results of this technical solution and two existing technologies in terms of equipment load balance.

[0162] In the figure, the horizontal axis represents 8 different device types, and the vertical axis represents the load balance degree index (between 0 and 1, the closer to 1, the more balanced the load distribution). This technical solution (diamond mark) shows excellent load balance performance for all device types, with a numerical range between 0.89 and 0.94, and an average of 0.92, especially reaching the highest 0.94 for the humidifying equipment.

[0163] In contrast, the balance index of the traditional load balancing algorithm (rectangular marker) ranges from 0.81 to 0.86, with an average of only 0.84, which is 8.7% lower than this solution; the round-robin scheduling algorithm (circular marker) performs even worse, with an index ranging from 0.65 to 0.75 and an average of only 0.7, which is 23.9% lower than this solution. The data also shows that the device load balance of this technical solution fluctuates minimally among different devices (with a standard deviation of only 0.016), while the standard deviations of the load balancing algorithm and the round-robin scheduling algorithm are 0.019 and 0.033 respectively, indicating that this solution can optimize the loads of various devices more consistently. Particularly significantly, for the shading device, the balance degree of this solution is 0.92, while that of the round-robin scheduling algorithm is only 0.65, an increase of 41.5%; for the fresh air device, this solution reaches 0.93, an increase of 14.8% compared to 0.81 of the load balancing algorithm. Figure 3 It fully proves that through spatio-temporal dynamic decoupling and multi-objective optimization, this technical solution can accurately capture the load characteristics and mutual influences of different devices and achieve more intelligent load distribution.

[0164] In an alternative embodiment,

[0165] The particle swarm optimization algorithm is used to solve the multi-objective optimization model. By updating the velocity and position of the particle swarm through an adaptive inertia weight and a dynamic convergence factor, the multi-objective solution results include:

[0166] Encode the target decision variables to be optimized as the position vector of the particle, initialize the velocity vector and position vector of the particle swarm, and calculate the multi-objective fitness value of each particle;

[0167] Sort the particle swarm based on the fitness value, select a preset number of particles with high fitness values to construct an elite particle swarm, and record the position vector of the particle with the highest fitness value as the current global optimal position vector;

[0168] In the iterative optimization process, dynamically calculate the adaptive inertia weight according to the current iteration number, and update the dynamic convergence factor through an exponential function;

[0169] For each particle to be updated, randomly select three different elite particles from the elite particle swarm, add the difference between the position vectors of the first elite particle and the other two elite particles multiplied by the scaling factor to obtain a differential evolution candidate position vector; perform a Lévy flight search on the global optimal position vector, and add the global optimal position vector and the search step determined by the step factor and the Lévy index to obtain a Lévy search candidate position vector;

[0170] Calculate the fitness values of two candidate position vectors, compare them with the current optimal fitness value, select the optimal position vector to update the global optimal solution, and update the velocity and position of the particle swarm based on the updated global optimal position vector, adaptive inertia weight, and dynamic convergence factor until the preset convergence condition is met, and output the decision variable corresponding to the optimal position vector as the multi-objective solution result.

[0171] Encode the target decision variables for environmental control, and convert control parameters such as temperature set value, humidity set value, fresh air volume, and lighting brightness into components of the particle position vector. Randomly generate an initial particle swarm, where the position vector of each particle represents a set of possible control strategies, and at the same time assign a random initial velocity vector to each particle. Calculate the fitness values under multiple optimization objectives for each particle, including environmental comfort index, energy consumption index, and equipment load index.

[0172] Sort the particle swarm according to the calculated fitness values, and select several particles with the highest fitness values to form an elite particle swarm. The elite particles represent the relatively optimal control schemes discovered currently. Find the particle with the highest fitness value among all particles, and record its position vector as the current global optimal position vector, which serves as the reference direction for the group search.

[0173] During the optimization iteration process, calculate the inertia weight value according to the number of iterations completed currently. A larger inertia weight is adopted at the beginning of the iteration to expand the search range, and the inertia weight is gradually reduced as the iteration progresses to enhance the local search ability. At the same time, calculate the dynamic convergence factor through an exponential function, which is used to adjust the speed at which the particles approach the optimal solution.

[0174] For each particle to be updated, randomly select three different elite particles from the elite particle swarm. Use the position vector of the first elite particle as the benchmark, and add the product of the difference between the position vectors of the other two elite particles and the scaling factor to generate a differential evolution candidate position vector.

[0175] Conduct a Lévy flight search on the global optimal position vector. Determine the search step size according to the preset step size factor and Lévy index, and add the step size value to the global optimal position vector to obtain a Lévy search candidate position vector.

[0176] Calculate the fitness values of the differential evolution candidate position vectors and the Levy search candidate position vectors respectively, and compare these fitness values with the current optimal fitness value. Select the position vector with the highest fitness value to update the global optimal solution. According to the updated global optimal position vector, combined with the calculated adaptive inertia weight and dynamic convergence factor, update the velocities and positions of the entire particle swarm. Repeat the above process until the preset convergence condition is reached, such as reaching the maximum number of iterations or the improvement amplitude in consecutive iterations is less than the threshold. Finally, decode the global optimal position vector into specific environmental control parameters as the solution to the multi-objective optimization problem.

[0177] Exemplarily, taking the optimization control of the air conditioning system in the waiting area of a certain terminal building as an example, encode control parameters such as the temperature setpoint, air supply volume, and fresh air ratio into the particle position vector. Initialize a population containing 100 particles, and each particle represents a set of possible air conditioning control strategies. Calculate the fitness values of each particle under three objectives: comfort, energy conservation, and equipment load.

[0178] After fitness calculation and sorting, select the top 20 high-fitness particles to form an elite particle swarm. The control strategy corresponding to the particle with the highest fitness is: temperature set value of 24 degrees, air supply volume of 80%, and fresh air ratio of 30%. This scheme is recorded as the current global optimal solution.

[0179] At the 50th iteration, the calculated inertia weight is 0.6 and the dynamic convergence factor is 0.8. For a particle to be updated, randomly select three particles from the elite swarm. Assume that the corresponding temperature set values are 24 degrees, 23.5 degrees, and 24.2 degrees respectively. The candidate temperature set value obtained through differential evolution operation is 23.8 degrees. At the same time, another candidate temperature set value of 24.1 degrees is obtained through Levy flight search.

[0180] Calculate the fitness values of these two candidate schemes and find that the scheme with a temperature set of 23.8 degrees can achieve better comprehensive effects. Accordingly, update the global optimal solution and drive other particles to move in this direction. After 200 iterations, the algorithm converges to the optimal control strategy: temperature set value of 23.8 degrees, air supply volume of 75%, and fresh air ratio of 35%.

[0181] In this embodiment, through the construction mechanism of the elite particle swarm, the high-quality solutions in the population are retained, providing a more reliable evolution direction. The design of the adaptive inertia weight enables the algorithm to automatically adjust the proportion of global search and local search during the iteration process, improving the search efficiency. The introduction of the dynamic convergence factor effectively controls the convergence speed of the particles, avoiding the problems of premature convergence or excessive divergence;

[0182] In the existing technology, the traditional particle swarm algorithm is prone to fall into the local optimum and it is difficult to find the global optimal solution in the complex solution space. The single evolution operation leads to insufficient population diversity, which affects the convergence performance of the algorithm. It uses a simple linear weight method to handle multiple optimization objectives, which makes it difficult to effectively balance the relationship between multiple objectives such as environmental comfort, energy efficiency and equipment load.

[0183] This embodiment combines the differential evolution strategy and Levy flight search. Differential evolution generates new search directions through vector difference operations between elite particles, which enhances the exploration ability of the population. Levy flight search helps the algorithm escape the local optimal trap through long-range jump characteristics, improves the global search ability, and can obtain better solutions under the same number of iterations. It has achieved significant improvements in solution quality, convergence efficiency, and computational stability, providing more reliable and efficient technical support for the optimization of the terminal environment control system.

[0184] Figure 4 This is a convergence performance comparison diagram of the terminal environment intelligent monitoring method combining multi-source data fusion with deep learning in an embodiment of the present invention, showing the convergence performance comparison of five different multi-objective optimization algorithms at different iteration times. The color smoothly changes from blue (low convergence rate) through purple, magenta to red (high convergence rate). The more the color leans towards red, the higher the convergence rate.

[0185] This technical solution (adaptive inertia weight and dynamic convergence factor improved particle swarm algorithm) appears as the brightest red area in the heat map, especially at 10 iterations, it reached a convergence rate of 0.91, which is much higher than other algorithms with the same number of iterations. At 50 iterations, the convergence rate reached 0.99, tied for the highest with the standard PSO algorithm, but this technical solution can obviously achieve the same convergence effect with fewer iterations.

[0186] The standard PSO algorithm shows good convergence performance. From the heat map, we can see that its color gradually transitions from lighter red to darker red. The convergence rate is 0.85 at the initial 10 iterations and 0.99 at the final 50 iterations, which shows that although the standard PSO has good convergence performance, it requires more iterations to achieve the same effect as this technical solution.

[0187] The differential evolution algorithm shows a transition area from blue to light purple in the heat map, indicating that its convergence rate is generally low. Even after 50 iterations, its convergence rate is only 0.88, the weakest among all algorithms. This may be related to the limitations of the differential evolution algorithm in dealing with multi-objective optimization problems.

[0188] The genetic algorithm shows a gradual change from blue-violet to pink in the heat map, converging from an initial 0.72 to a final 0.95. This indicates that although the genetic algorithm converges slowly in the initial stage, it can achieve good results as the number of iterations increases. However, it is still lower than the proposed technical solution and the standard PSO.

[0189] The simulated annealing algorithm shows a transition from purple to pink in the heat map, and its convergence performance is between that of the standard PSO and the genetic algorithm. It shows a relatively stable improvement in convergence performance from an initial 0.78 to a final 0.96.

[0190] The proposed technical solution can achieve a convergence rate that other algorithms require more iterations to reach at a low number of iterations, which is of great significance for real-time control systems and can find a better solution within a limited computing time.

[0191] By combining an adaptive inertia weight adjustment, a dynamic convergence factor, an elite particle swarm strategy, and a Lévy flight search mechanism, the proposed technical solution not only improves the convergence speed of the algorithm but also enhances the ability to jump out of local optima. In complex optimization problems such as the air conditioning system in the terminal waiting area that need to balance comfort, energy consumption, and equipment load, it can find an optimal control strategy that balances various objectives in a more efficient way, providing strong algorithm support for intelligent building environment regulation.

[0192] In the second aspect of the embodiments of the present invention,

[0193] A kind of electronic device is provided, including:

[0194] A processor;

[0195] A memory for storing instructions executable by the processor;

[0196] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0197] In the third aspect of the embodiments of the present invention,

[0198] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0199] The present invention can be a method, a device, a system, and / or a computer program product. The computer program product can include a computer-readable storage medium on which computer-readable program instructions for executing various aspects of the present invention are loaded.

[0200] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Multisource data fusion combined with deep learning for intelligent monitoring method of terminal building environment, characterized in that Including: Collecting multi-source perception data in the terminal building, performing time-series alignment and spatial registration preprocessing, generating a standardized perception data matrix, and segmenting the perception data matrix; Inputting the perception data matrix into a dual-stream neural cognitive computing framework, extracting multi-scale spatio-temporal features through an adaptive dynamic convolution module, determining long-term and short-term dependencies between different types by combining a deformable attention mechanism to obtain perception stream features, constructing an environmental knowledge reasoning module based on a memory-enhanced neural network, mining the causal relationship of environmental changes by combining historical experience knowledge and a recursive neural reasoning unit to obtain cognitive stream features, and fusing the perception stream features and the cognitive stream features to obtain a fused feature representation; Inputting the fused feature representation into a terminal building environment state evaluation model to generate an environmental evaluation result including the terminal building environmental comfort index, safety risk level, and emergency event warning information; Inputting the environmental evaluation result into a spatio-temporal dynamic decoupler, decoupling in the time dimension and the space dimension by combining an attention mechanism to obtain dynamic environmental features, constructing a multi-objective optimization model based on the dynamic environmental features, constructing a constrained optimization problem through a particle swarm optimization algorithm, combining an adaptive inertia weight and a dynamic convergence factor to solve a preset optimization objective, generating a multi-objective solution result, and generating an environmental regulation strategy through a fuzzy mapping algorithm; Sending the environmental regulation strategy to the terminal building environmental control system for execution.

2. The method according to claim 1, wherein Collecting multi-source perception data in the terminal building, performing time-series alignment and spatial registration preprocessing, and segmenting the perception data matrix includes: Collecting multi-source data in the terminal building through temperature and humidity sensors, air quality sensors, video surveillance devices, and noise sensors; Synchronously aligning the multi-source data according to timestamps, registering the sensing data at different positions based on spatial coordinate mapping, and generating a perception data matrix in a unified format through standardization processing; Adopting a sliding time window technique, setting an initial sliding window according to a preset time step, and dynamically segmenting the perception data matrix in combination with the data change trend.

3. The method according to claim 1, wherein Inputting the perception data matrix into a dual-stream neural cognitive computing framework, and extracting multi-scale spatio-temporal features through an adaptive dynamic convolution module, and determining long-term and short-term dependencies between different types by combining a deformable attention mechanism to obtain perception stream features includes: Decoupling the input perception data matrix in the spatio-temporal dimension to separately obtain a time series matrix and a spatial feature matrix; Performing multi-scale decomposition on the time series matrix through wavelet transform to extract time features at different frequencies; inputting the spatial feature matrix into a feature adaptive network to generate dynamic convolution kernel weights, and extracting spatial features based on the dynamic convolution kernel weights to obtain spatial local features; and performing tensor fusion on the time features and the spatial local features to obtain a spatio-temporal feature matrix; Calculate the feature similarity based on the spatio-temporal feature matrix, and dynamically construct an association graph structure. The nodes of the association graph represent the feature vectors of different types of sensing data. Determine the connection weights between nodes through an adaptive threshold, combine the connection weights with the learnable position offsets, dynamically adjust the attention calculation positions, and generate a fused feature sequence; Divide the fused feature sequence into short-term segments and long-term segments according to the time span, extract local temporal change features from the short-term segments, and extract long-term change trend features from the long-term segments; input the local temporal change features and the long-term change trend features into a memory gated unit to generate a perceptual flow feature.

4. The method according to claim 1, wherein Construct an environmental knowledge reasoning module based on a memory-augmented neural network, combine historical experience knowledge and a recursive neural reasoning unit to mine the causal relationship of environmental changes to obtain a cognitive flow feature, and fuse the perceptual flow feature and the cognitive flow feature to obtain a fused feature representation, including: Construct a historical experience knowledge graph containing a concept layer, a relationship layer, and an instance layer based on the input perceptual data matrix, where the concept layer stores the concept mapping information of environmental elements, the relationship layer stores semantic association rules, and the instance layer stores instance feature vectors; Input the environmental state feature vector into a memory-augmented neural network, calculate the similarity score based on the node information in the historical experience knowledge graph, extract the historical experience knowledge node with the highest similarity score, calculate the attention weight coefficient, and generate an attention weight matrix; Perform importance weighting on environmental variables according to the attention weight matrix, construct a directed acyclic graph in combination with the historical experience knowledge graph, and train a recursive neural reasoning unit; Obtain a state transition matrix through the forward propagation of the trained recursive neural reasoning unit to generate a causal evolution feature, and generate an attribution relationship feature matrix through backpropagation; Intervene in the target environmental variable in the control group scenario, calculate the environmental state change amount to determine the causal intensity coefficient, and perform weighted fusion on the causal evolution feature and the attribution relationship feature matrix based on the causal intensity coefficient to obtain a cognitive flow feature; Perform feature fusion on the perceptual flow feature extracted from the perceptual data matrix and the cognitive flow feature to obtain a fused feature representation.

5. The method according to claim 1, characterized in that, Input the fused feature representation into the terminal building environmental state evaluation model to generate an environmental evaluation result including the terminal building environmental comfort index, safety risk level, and emergency event warning information, including: Input the fused feature representation into the environmental state evaluation model, calculate the PMV-PPD comfort evaluation index of temperature, humidity, illuminance, noise, and air quality index in the fused feature representation to obtain the environmental comfort index, calculate the deviation degree of the crowd density, hazard source distribution, and equipment operation status from the preset threshold to judge the safety risk level, calculate the abnormality degree of the passenger flow congestion condition, fire hazard index, and abnormal behavior detection result to identify the warning event type and risk degree, and generate emergency event warning information; Combine the environmental comfort index, safety risk level, and emergency event warning information, and output the terminal building environmental evaluation result.

6. The method according to claim 1, characterized in that, Input the environmental assessment results into a spatio-temporal dynamic decoupler, decouple in the time dimension and space dimension by combining with an attention mechanism to obtain dynamic environmental features, construct a multi-objective optimization model based on the dynamic environmental features, construct a constrained optimization problem through a particle swarm optimization algorithm, and solve the preset optimization objectives by combining an adaptive inertia weight and a dynamic convergence factor to generate a multi-objective solution result and generate an environmental regulation strategy through a fuzzy mapping algorithm, including: Input the environmental assessment results into the spatio-temporal dynamic decoupler of the dual-stream network structure, extract the decoupling features in the time dimension through the bidirectional long short-term memory network of the time stream branch, and extract the decoupling features in the space dimension through the graph convolutional network of the space stream branch; Adopt a spatio-temporal dual attention mechanism to calculate the attention weights of the decoupling features, perform tensor multiplication operations on the attention weights and the corresponding decoupling features and add them to obtain dynamic environmental features; Construct a multi-objective optimization model based on the dynamic environmental features, including an environmental comfort optimization objective, an energy consumption optimization objective, and a device load optimization objective; Adopt a particle swarm optimization algorithm to solve the multi-objective optimization model, update the velocity and position of the particle swarm through an adaptive inertia weight and a dynamic convergence factor to obtain a multi-objective solution result; Adopt a fuzzy mapping algorithm to perform fuzzy mapping on the multi-objective solution result to generate an environmental regulation strategy including temperature and humidity setting, dimming, ventilation frequency, and device start-stop timing sequence.

7. The method according to claim 6, characterized in that Adopt a particle swarm optimization algorithm to solve the multi-objective optimization model, update the velocity and position of the particle swarm through an adaptive inertia weight and a dynamic convergence factor to obtain a multi-objective solution result, including: Encode the target decision variables to be optimized into the position vectors of the particles, initialize the velocity vectors and position vectors of the particle swarm, and calculate the multi-objective fitness values of each particle; Sort the particle swarm based on the fitness values, select a preset number of particles with high fitness values to construct an elite particle swarm, and record the position vector of the particle with the highest fitness value as the current global optimal position vector; In the iterative optimization process, dynamically calculate the adaptive inertia weight according to the current iteration number, and update the dynamic convergence factor through an exponential function; For each particle to be updated, randomly select three different elite particles from the elite particle swarm, add the differences between the position vectors of the first elite particle and the other two elite particles multiplied by a scaling factor to obtain a differential evolution candidate position vector; perform a Lévy flight search on the global optimal position vector, and add the global optimal position vector and the search step determined by the step factor and the Lévy index to obtain a Lévy search candidate position vector; Calculate the fitness values of the two candidate position vectors, compare them with the current optimal fitness value to select the optimal position vector to update the global optimal solution, and update the velocity and position of the particle swarm based on the updated global optimal position vector, adaptive inertia weight, and dynamic convergence factor until the preset convergence condition is met, and output the decision variables corresponding to the optimal position vector as the multi-objective solution result.

8. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images