Indoor air quality joint detection system based on spatio-temporal graph neural network

Abstract

This invention relates to the field of indoor air quality monitoring technology and discloses a joint indoor air quality detection system integrating a spatiotemporal graph neural network. The system includes a multi-source sensor data acquisition module that collects multi-source air quality data streams by configuring multiple sensors; a spatiotemporal graph construction module that constructs spatiotemporal graph data based on the spatial location and time series of the sensors; a fusion processing module that uses a spatiotemporal graph neural network to perform feature fusion on the spatiotemporal graph data to generate an air quality fusion feature set; a prediction strategy module that determines a prediction strategy based on the remaining time at the end of the prediction period and combines it with the fusion feature set to execute air quality prediction; a dynamic adjustment module that analyzes real-time data changes, constructs pollution variability and trend, and adjusts detection control parameters accordingly; and a joint detection output module that integrates the prediction results and adjustment parameters to generate a joint indoor air quality detection report. This system provides strong support for indoor air quality assessment.

Description

Technical Field

[0001] This invention relates to the field of indoor air quality monitoring technology, specifically to an indoor air quality joint detection system that integrates spatiotemporal graph neural networks. Background Technology

[0002] With the acceleration of urbanization and changes in people's lifestyles, indoor environments have become the main places for people's daily activities, and the quality of indoor air directly affects human health and living comfort. In recent years, indoor air pollution has received increasing attention. Formaldehyde and benzene compounds released from decoration materials, cooking fumes, and increased carbon dioxide concentrations due to poor ventilation can all cause respiratory diseases, allergic reactions, and other health problems. Against this backdrop, the accuracy and real-time performance of indoor air quality testing technology are becoming increasingly important.

[0003] Traditional indoor air quality monitoring methods mostly rely on single-source sensors or fixed monitoring points, which have significant limitations. Single-source sensors collect data with only one dimension, reflecting the air quality status of a localized area and failing to comprehensively cover the pollution differences between different indoor spaces, easily leading to detection blind spots. Furthermore, existing technologies often sever the connection between time and space during data processing, analyzing pollutant concentrations only at a single point in time, ignoring the diffusion patterns of pollutants within the indoor space and their dynamic changes over time. This results in detection results that fail to reflect the true evolution of pollution.

[0004] In terms of data fusion, traditional methods often employ simple weighted averages or statistical analysis, which cannot effectively uncover the spatiotemporal correlation features hidden in sensor data. As the number of sensors increases, the data dimensionality rises sharply. When processing high-dimensional spatiotemporal data, traditional algorithms are prone to feature redundancy or loss of key information, affecting the accuracy of subsequent analysis.

[0005] In the forecasting phase, existing systems typically employ fixed forecasting models without considering the impact of forecast period length on model applicability. When the forecast period is short, the model needs to focus on short-term data fluctuations; while when the period is long, it needs to focus on long-term trend changes. Fixed models struggle to meet the forecasting needs of different scenarios, leading to discrepancies between forecast results and actual conditions.

[0006] Traditional detection systems typically use preset, fixed control parameters that cannot be dynamically adjusted based on real-time changes in indoor pollution. When pollutant concentrations suddenly increase or diffusion rates accelerate, the detection frequency and accuracy under fixed parameters cannot respond quickly enough, potentially missing the optimal time for pollution control and impacting the system's practicality and flexibility.

[0007] Existing indoor air quality detection technologies are inadequate in terms of the comprehensiveness of data acquisition, the ability to mine spatiotemporal features, the adaptability of prediction strategies, and the dynamic adjustment of detection parameters. There is an urgent need for a joint detection system that can integrate multi-source data, take into account spatiotemporal correlation, dynamically adapt to prediction needs, and adjust detection parameters in real time to improve the accuracy and reliability of indoor air quality detection. Summary of the Invention

[0008] The purpose of this invention is to provide an indoor air quality joint detection system that integrates spatiotemporal graph neural networks to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides an indoor air quality joint detection system integrating spatiotemporal graph neural networks, the system comprising:

[0010] The multi-source sensor data acquisition module is equipped with various indoor air quality sensors to collect multi-source air quality data streams.

[0011] The spatiotemporal map construction module is used to construct spatiotemporal map data based on the sensor spatial locations and time series in the multi-source air quality data stream;

[0012] The fusion processing module is used to perform feature fusion processing on the spatiotemporal map data using a spatiotemporal map neural network to generate an air quality fusion feature set.

[0013] The prediction strategy module is used to determine the prediction strategy based on the remaining time between the current time and the end time of the prediction cycle, and to perform air quality prediction according to the air quality fusion feature set.

[0014] The dynamic adjustment module is used to analyze the real-time changes of the multi-source air quality data stream, construct the pollution variability and trend, and adjust the detection and control parameters according to the pollution variability and trend.

[0015] The joint detection output module is used to generate a joint indoor air quality detection report based on the output of the prediction strategy and the adjustment result of the dynamic adjustment module.

[0016] Preferably, the multi-source sensor data acquisition module includes:

[0017] Multiple indoor air quality sensors, including temperature and humidity sensors, particulate matter sensors, and gas concentration sensors;

[0018] The noise analysis unit is used to perform noise characteristic analysis on the multi-source air quality data stream and determine the noise data denoising threshold.

[0019] The preprocessing unit is used to perform noise identification and filtering on the multi-source air quality data stream according to the noise data denoising threshold, and generate a standard multi-source air quality data stream.

[0020] Preferably, the fusion processing module includes:

[0021] The feature extraction unit is used to determine the spatiotemporal feature extraction method and feature type based on the data characteristics of the standard multi-source air quality data stream;

[0022] The correlation feature generation unit is used to extract correlation features from the standard multi-source air quality data stream according to the spatiotemporal feature extraction method and feature type to obtain a multi-source signal correlation feature set.

[0023] The weight allocation unit is used to set dynamic weight allocation rules based on sensor reliability and accuracy;

[0024] The fusion execution unit is used to perform weighted fusion of the multi-source signal correlation feature set based on the weight dynamic allocation rule to generate the air quality fusion feature set.

[0025] Preferably, the prediction strategy module includes:

[0026] The remaining duration calculation unit is used to calculate the remaining duration between the current time and the end time of the prediction period;

[0027] The prediction period determination unit is used to determine the first prediction period by matching the remaining time with a preset conversion relationship;

[0028] The strategy generation unit is used to determine the adjustment period based on the number of energy types or the number of equipment, combined with the second conversion relationship, and to calculate the second prediction period by the difference between the first prediction period and the adjustment period.

[0029] The prediction execution unit is used to control the prediction model to process the air quality fusion feature set according to the second prediction cycle and output future air quality prediction data.

[0030] Preferably, the prediction execution unit includes:

[0031] The model calling subunit is used to call a pre-trained spatiotemporal graph neural network prediction model;

[0032] The data input subunit is used to input the air quality fusion feature set and connection relationships into the spatiotemporal graph neural network prediction model;

[0033] The prediction output subunit is used to perform a prediction operation based on the second prediction period to generate the future air quality prediction data.

[0034] Preferably, the dynamic adjustment module includes:

[0035] The variation period extraction unit is used to perform frequency domain transformation on the multi-source air quality data stream and extract the variation period of pollution parameters.

[0036] The variability construction unit is used to analyze the frequency of parameter changes within the change cycle of the pollution parameters and construct the pollution variability.

[0037] The trend building unit is used to analyze the downward and upward trends of parameters at multiple adjacent sampling times and to build the trend.

[0038] The adjustment parameter generation unit is used to generate detection control parameters based on the pollution variability and trend, combined with a preset adjustment amount.

[0039] Preferably, the degree of change construction unit includes:

[0040] The vector generation subunit is used to generate a periodic vector by arranging pollution parameters according to a time series.

[0041] The difference analysis subunit is used to calculate the first-order difference sequence of the periodic vector;

[0042] The change measurement operator unit is used to calculate the degree of pollution change based on the change frequency and average change amplitude of the first-order difference sequence.

[0043] Preferably, the trend construction unit includes:

[0044] The time series modeling subunit is used to fit a time series trend line based on pollution parameters at multiple adjacent sampling times through linear regression analysis;

[0045] The trend quantification subunit is used to extract the slope feature of the time series trend line and generate the trend degree based on the value of the slope feature.

[0046] Preferably, the adjustment parameter generation unit includes:

[0047] The interference degree calculation subunit is used to determine the degree of environmental interference based on the correlation and average value of the pollution variability and trend.

[0048] A looseness construction sub-unit is used to calculate the detected looseness based on the degree and trend of environmental interference;

[0049] The feedback parameter generation subunit is used to generate feedback detection control parameters based on the difference in the detected looseness and the preset adjustment amount.

[0050] Preferably, the joint detection output module includes:

[0051] The deviation calculation unit is used to compare future air quality forecast data with preset air quality plans and generate deviation data.

[0052] The report generation unit is used to integrate the deviation data and the detection control parameters to generate the joint indoor air quality detection report.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] This indoor air quality joint detection system, which integrates spatiotemporal graph neural networks, brings multifaceted practical value through the collaborative operation of multiple modules. The multi-source sensor data acquisition module is equipped with various indoor air quality sensors, changing the traditional single-source sensor data dimension limitation. It can simultaneously acquire concentration data of different types of pollutants, covering the air quality status of different areas indoors, reducing detection blind spots, and making the acquired multi-source air quality data stream more reflective of the overall indoor environment.

[0055] The spatiotemporal map construction module generates spatiotemporal map data based on the spatial location and time series of sensors. This connects the originally scattered sensor data through spatial location correlation and time series correlation, forming a structured spatiotemporal data model. This approach breaks through the limitations of traditional data processing that separates time and space, revealing the inherent relationships between data points. This lays the foundation for subsequent feature extraction and analysis, enabling the system to more comprehensively grasp the diffusion paths of pollutants indoors and their changes over time.

[0056] The fusion processing module employs a spatiotemporal graph neural network to perform feature fusion processing on spatiotemporal graph data, fully leveraging the advantages of spatiotemporal graph neural networks in processing spatiotemporally correlated data. Compared to traditional weighted averaging or statistical analysis methods, this network can automatically uncover complex spatiotemporal features hidden within the data, effectively integrating information from different time points and spatial locations. The generated air quality fusion feature set retains more key information, avoiding feature redundancy or information loss, and providing more reliable feature support for subsequent prediction stages.

[0057] The forecasting strategy module determines the forecasting strategy based on the current time and the remaining time at the end of the forecast period, thus overcoming the limitations of traditional fixed forecasting models. Depending on the remaining time, the focus of the forecasting model can be flexibly adjusted. When the remaining time is short, the focus can be on predicting short-term data fluctuations; when the remaining time is long, the focus shifts to analyzing long-term trend changes, making the forecast results more aligned with actual needs and improving the relevance and rationality of the forecasts.

[0058] The dynamic adjustment module analyzes real-time changes in multi-source air quality data streams to construct pollution variability and trend indicators, and adjusts detection and control parameters accordingly. This enables the system to respond in real-time to dynamic changes in indoor air quality. When pollutant concentrations fluctuate abnormally or diffusion trends change, the detection and control parameters can be adjusted promptly to optimize detection frequency and accuracy. This ensures the system maintains high-efficiency detection performance under different pollution conditions, enhancing the system's flexibility and adaptability.

[0059] The combined detection output module integrates the output of the prediction strategy with the adjustment results of the dynamic adjustment module to generate a detection report. By combining predicted information with real-time adjustment parameters, the report becomes more comprehensive and detailed. The report not only reflects the current air quality status but also provides predictions of pollution trends over a future period. It also includes the basis for the system's adjustments, offering users comprehensive reference information to understand indoor air quality and take appropriate improvement measures, thus enhancing the system's practical value. Attached Figure Description

[0060] Figure 1 This is a timing diagram of the indoor air quality joint detection system fused with spatiotemporal graph neural network as described in this invention;

[0061] Figure 2 This is a schematic diagram of the working principle of the fusion processing module;

[0062] Figure 3 This is a diagram showing the data acquisition and preprocessing of temperature sensors.

[0063] Figure 4 This is a diagram showing the data acquisition and preprocessing of particulate matter sensors.

[0064] Figure 5 This is a diagram showing the data acquisition and preprocessing of gas concentration sensors.

[0065] Figure 6 This is a diagram showing the data acquisition and preprocessing of humidity sensor data.

[0066] Figure 7 Schematic diagram of the predictive execution unit;

[0067] Figure 8 This is a graph showing the results of multi-period air quality forecasts;

[0068] Figure 9 This is a schematic diagram of the working principle of the dynamic adjustment module. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Please see Figure 1 This invention provides an indoor air quality joint detection system integrating spatiotemporal graph neural networks, the system comprising:

[0071] Intelligent indoor air quality monitoring is achieved through multi-source sensor data acquisition, spatiotemporal map construction, feature fusion processing, dynamic prediction strategies, and parameter adjustment. The system collects multi-dimensional air quality data using distributed temperature and humidity sensors, particulate matter sensors, and gas concentration sensors, constructing a spatiotemporal map data structure that includes spatial topological relationships and temporal dependencies. A spatiotemporal graph neural network is used to perform feature-level fusion of multi-source heterogeneous data, establishing a dynamic correlation model between sensor nodes. Based on an adaptive prediction strategy for remaining time, multi-period air quality prediction results are generated. Simultaneously, real-time optimization and adjustment of detection parameters are achieved through pollution dynamic analysis, outputting a joint detection report of the fused prediction results and real-time monitoring data.

[0072] Example 1: See Figure 2Based on the data characteristics of standard multi-source air quality data streams, the spatiotemporal feature extraction method and feature types are determined. Correlation features are extracted from the standard multi-source air quality data streams according to the spatiotemporal feature extraction method and feature types to obtain a multi-source signal correlation feature set. Based on sensor reliability and accuracy, dynamic weight allocation rules are set, and the multi-source signal correlation feature set is weighted and fused based on the dynamic weight allocation rules to generate an air quality fusion feature set. The multi-source sensor data acquisition module adopts a distributed architecture deployed in the indoor environment, with sensor node clusters set up at diagonal locations, ventilation areas, and densely populated areas. Each cluster node integrates multiple sensor units. Physical sensors use digital temperature and humidity sensors to detect changes in ambient temperature and relative humidity. Particulate matter sensors using the laser scattering principle monitor PM1.0, PM2.5, and PM10 concentration gradients. The electrochemical gas sensor array includes a carbon dioxide sensor, a total volatile organic compound sensor, and a formaldehyde detection sensor. All sensor nodes form a wireless sensor network via the LoRaWAN protocol, transmitting raw data packets in the 868MHz band with an initial transmission interval of 60 seconds. The noise analysis unit implements a dual time-frequency domain analysis mechanism. Within the time domain window, it statistically analyzes the peak-to-valley fluctuation range of sampled values, and in the frequency domain, it uses Morlet wavelet basis functions for multi-scale decomposition. For temperature sensor data, it calculates the energy distribution of its third-order wavelet detail coefficients; for gas concentration data, it analyzes the standard deviation distribution of its approximation coefficients. Dynamic judgment criteria are set according to sensor type differences: a noise threshold model based on interquartile range is used for particulate matter sensors, while the sliding coefficient of variation method is employed for gas sensors. The system has a built-in noise characteristic database; when abnormal fluctuations exceeding the current threshold are detected, an adaptive threshold update procedure is triggered.

[0073] The preprocessing unit executes a multi-stage filtering process: first, sensor-level data cleaning is performed, compensating for zero-point drift based on equipment calibration parameters; second, an improved bilateral filtering algorithm is employed, considering both temporal neighborhood similarity and numerical differences. A weighted average filter with a 5-second time window is applied to temperature data, and a relative saturation correction algorithm is implemented for humidity data. The particulate matter data processing is unique in that it uses dynamic density clustering to eliminate brief interference pulses and performs linear interpolation compensation for abrupt changes lasting less than 0.5 seconds. Due to the response delay characteristics of gas sensor data, additional reverse prediction compensation is implemented: the current true value is estimated based on the changing trends of the three most recent sampling points, and the compensation coefficient is calculated by combining this with the sensor response curve. All processing results are ultimately converted to a standard unit system, with temperature data calibrated to an accuracy range of ±0.3℃, particulate matter concentration converted to μg / m³ and compensated for altitude, and gas concentration data subjected to nonlinear correction based on standard gas parameters.

[0074] The fusion processing module establishes a hierarchical processing framework for feature extraction. The first-level time-domain feature extraction employs an overlapping sliding window mechanism, using 10 sampling points as the basic window unit and sliding calculation with an 80% overlap rate. In the basic statistical extraction stage, the arithmetic mean and root mean square value within the window are calculated for temperature and humidity data; peak-to-peak value (the difference between the maximum and minimum values ​​of the pollution parameter within the same time window, used to characterize the fluctuation range of the parameter within the window) and coefficient of variation are extracted for particulate matter concentration data; and a weighted average of the absolute values ​​of the first-order differences is calculated for gas concentration data. The second-level frequency domain analysis includes two stages: first, Hann windowing is applied to the raw data, and then the spectral distribution is obtained through Fast Fourier Transform. Temperature and humidity data focuses on extracting the energy proportion in the 0.05-0.5Hz frequency band; particulate matter data analyzes the power of high-frequency components above 1Hz; and gas data extracts phase angle information at characteristic frequency points. The third-level spatial feature modeling uses a graph structure representation, with sensor locations as node coordinates, and constructs an adjacency matrix based on physical distance and environmental connectivity. Establish a dynamic spatial association model: when the distance between two nodes is less than 2 meters, they are set as strong connections; add virtual connection edges with a weight of 0.5 between areas sharing ventilation paths.

[0075] A multimodal feature fusion mechanism is designed for the association feature generation unit. A spatiotemporal association model is constructed based on a graph attention network: a fully connected graph structure with 32 nodes is created, with each node corresponding to 8 dimensions of temporal features and 4 dimensions of frequency features. Inter-node interaction features are calculated through a two-layer attention network. The first layer calculates the influence weights of spatially adjacent nodes, and the second layer analyzes the long-distance associations between nodes across regions. The attention network design uses a 16-dimensional query vector and a 24-dimensional key vector, employing a scaled dot product attention mechanism. The output of each attention layer is processed by a normalization layer and then residually concatenated with the original features. Finally, the association features are combined in a cascaded manner: 12 dimensions of the node's own basic features, 16 dimensions of spatially adjacent node features, and 4 dimensions of cross-regional association features, forming a 32-dimensional fixed-length spatiotemporal association feature vector.

[0076] The weighting unit establishes a dynamic weighting system. The sensor reliability scoring system is based on three dimensions: maintaining the device calibration record and the time of the five most recent calibrations, calculating the calibration interval coefficient; real-time health diagnosis to detect sensor power supply voltage fluctuations and signal integrity; and environmental adaptability indicators to assess the sensor's operational stability under specific temperature and humidity conditions. The accuracy evaluation system differentiates based on device type: temperature and humidity sensors use fixed accuracy weight coefficients; particulate matter sensors dynamically adjust based on the current concentration level: a high-accuracy mode is activated when PM2.5 exceeds 75 μg / m³; and gas sensors are converted to confidence levels based on the data change rate. The weight calculation uses an improved entropy weighting method: an evaluation matrix is ​​constructed containing 18 evaluation indicators, the dispersion of each indicator is calculated through information entropy, and the final comprehensive weight value is generated by combining subjective weight parameters. The feature weights of each sensor are updated hourly, and an immediate weight recalculation procedure is triggered when the system detects a sudden environmental change.

[0077] The fusion execution unit implements a feature pyramid merging architecture. The input layer receives a 32-dimensional associated feature vector and a sensor weight matrix, while the feature reduction layer normalizes features at different scales. In the basic feature layer, temporal features are reduced to 16 dimensions using max pooling; in the spatial feature layer, bilinear interpolation is used to expand node features to 64 dimensions; and positional encoding enhancement is implemented in the associated feature layer. A coefficient amplification mechanism is designed in the weighted fusion stage: the weights of high-confidence sensor features are increased by 50%, and low-precision channel features are smoothly compressed. Feature concatenation employs a multi-resolution fusion strategy: bottom-level features retain their original numerical precision, middle-level features undergo logarithmic transformation to compress their numerical range, and high-level features incorporate positional differential encoding. Finally, the three-layer feature output is integrated through a fully connected layer containing 128 neurons, with the ELU function as the activation function. The output layer performs feature standardization, generating a 128-dimensional air quality fusion feature vector. This feature set includes information on temporal dynamics, spatial correlation, and device characteristics. The feature set is updated every 60 seconds and stored in a circular buffer, supporting feature backtracking analysis in subsequent modules.

[0078] See Figure 3 , Figure 4 , Figure 5 and Figure 6 This paper demonstrates the data acquisition process and preprocessing effects of multi-source sensors. The system collects multi-dimensional air quality data through distributed temperature and humidity sensors, particulate matter sensors, and gas concentration sensors. The figure shows a comparison between the raw sensor data and the data after multi-stage filtering. Temperature sensor data is filtered using a weighted average with a 5-second time window; humidity data undergoes a relative saturation correction algorithm; particulate matter data uses dynamic density clustering to eliminate transient interference pulses; and gas sensor data undergoes inverse prediction compensation due to its response delay characteristics. The preprocessed data effectively eliminates noise and outliers, providing high-quality input for subsequent analysis.

[0079] Example 2: See Figure 7 The system invokes a pre-trained spatiotemporal graph neural network prediction model, inputting the air quality fusion feature set and connectivity relationships into the model. Based on the second prediction cycle, it performs prediction operations to generate future air quality forecast data. The time management unit of the prediction strategy module adopts a hierarchical time architecture design. The remaining time calculation unit obtains the Coordinated Universal Time (UTC) reference signal through a satellite time synchronization module and calculates the precise time difference between the current moment and the planned prediction termination point using a local real-time clock, achieving millisecond-level time resolution. This calculation process is automatically executed once per minute, and the calculation result is written to a time buffer register. The prediction cycle determination unit has a built-in intelligent matching engine containing a three-layer dynamic judgment structure: when the remaining time is less than 60 minutes, the short-term prediction mode is activated; between 60 and 240 minutes, the medium-term prediction mode is used; and when it exceeds 240 minutes, the long-term prediction mechanism is switched to. Each mode corresponds to preset basic configuration parameters: the default step size for the short-term mode is 5 minutes; the medium-term mode dynamically sets the step size to 10-20 minutes according to the proportion of the remaining time; and the long-term mode uses a fixed step size of 30 minutes. The matching engine implements a forward optimization strategy, setting a 10-minute overlap buffer at mode switching boundaries to avoid prediction gaps.

[0080] The strategy generation unit establishes a database of device response models. This database stores dynamic parameters for various air handling equipment: the standard response time for a single fresh air unit is 45 seconds; the response time decreases by 2% for every 100W increase in air purifier power; and the central air conditioning system has a 60-second system delay. Based on a spatial area division model, an effective device density index is calculated, compressing the adjustment cycle by 0.7% for each additional online device. The dynamic prediction cycle generator performs three steps: first, extracting the number and total power of currently active device types; second, calculating the minimum adjustment cycle based on the device response time model; and third, subtracting the adjustment cycle value from the basic prediction cycle to generate the final execution cycle. The purpose of calculating the minimum adjustment cycle is to match the actual response capability of the air handling equipment, avoiding mismatches between the prediction cycle and the device adjustment cycle that would prevent the prediction results from guiding real-time control. For example, if the prediction cycle is 5 minutes but the device response requires 60 seconds, the prediction step size needs to be calibrated using the minimum adjustment cycle to ensure that the prediction results synchronously guide the device's actions. The minimum adjustment cycle calculation logic is as follows: ① Based on the number and total power of currently activated devices, match the basic response time of a single device from the device response model database (e.g., the basic response time of a 1000W air purifier is 38 seconds, and the basic response time of a fresh air system is 45 seconds); ② Calculate the multi-device collaborative delay: when multiple devices are running simultaneously, add the collaborative delay by multiplying the number of devices by 5 seconds (e.g., adding a 10-second delay for 2 devices); ③ Minimum adjustment cycle = (maximum basic response time of a single device) + multi-device collaborative delay (e.g., the minimum adjustment cycle of 1 air purifier + 1 fresh air system = 45 seconds + 5 seconds = 50 seconds). When a device status change event is detected, the strategy reconstruction process is immediately triggered. For spatial environments with multiple independent control zones, each zone generates an independent prediction strategy and sets a prediction synchronization point.

[0081] The prediction execution unit integrates a triple model invocation mechanism. The model invocation subunit maintains a prediction model repository, storing three spatiotemporal neural network architectures trained offline: a four-layer temporal convolutional structure for short-term prediction, a hybrid module of graph convolution and long short-term memory for mid-range models, and a codec framework for long-range models. The model selector automatically loads the corresponding model parameters into the computation memory based on the prediction mode. The data input subunit performs feature reorganization: reordering the air quality fusion feature set into 32-dimensional feature vectors according to sensor node indices, while simultaneously loading a pre-generated spatial adjacency matrix. This adjacency matrix uses a three-order connection form: direct physical connections have a weight of 1, indirect ventilation associations have a weight of 0.3, and unrelated nodes have negative weights to prevent interference. In the model data channel, the feature vectors are input to the spatiotemporal convolution branch, and the adjacency matrix is ​​input to the graph computation branch.

[0082] The data prediction process employs a differentiated processing scheme. In short-term prediction mode, the system uses a rolling prediction mechanism: first, a single-step prediction is performed to obtain the result of the most recent period, which is then used as a new feature to fill the input queue before predicting the next time node, forming a closed-loop prediction chain. Mid-term prediction utilizes a multi-step parallel engine, dividing the prediction time axis into continuous time segments. Each segment starts an independent prediction thread, and data smoothing and alignment are performed at the segment junctions. Long-term prediction is achieved through a sequence generation architecture: the encoder network compresses the input features into a 128-dimensional state vector, and the decoder gradually unfolds the prediction results according to preset time steps. In the pollutant concentration prediction task, the model simultaneously outputs six indicators: temperature deviation, relative humidity percentage, PM2.5 concentration, carbon dioxide content, total volatile organic compounds, and formaldehyde concentration. The air quality index prediction module, based on a comprehensive conversion algorithm, inputs the six raw concentration values ​​into the indicator converter and outputs an air quality assessment level that conforms to national standards.

[0083] The prediction output subsystem constructs a three-dimensional data integration structure. The basic data layer stores the original prediction numerical matrix, with the time dimension corresponding to the prediction step sequence and the spatial dimension maintaining a 32-sensor node structure. The analysis layer adds a trend labeling function: calculating directional indicators for data changes over three consecutive periods, marking upward trends with positive values ​​and downward trends with negative values. The presentation layer generates visualized prediction curves, with each pollutant indicator using an independent vertical axis, and the time horizontal axis automatically scaling according to the prediction duration. To address spatial difference analysis needs, the system outputs a hotspot distribution cloud map at the end of the prediction period, identifying the three-dimensional coordinates of areas exceeding pollutant standards. Prediction data is refreshed every 5 minutes, and complete data packets are transmitted to the storage unit via a secure encrypted channel; historical prediction records are retained for at least 30 days. In the event of power outages or other abnormal situations, the prediction module automatically switches to a simplified mode, retaining only the core indicator prediction function until the system returns to normal.

[0084] See Figure 8 This paper showcases the results of multi-cycle air quality forecasts based on adaptive remaining time. The system determines the forecast strategy based on the remaining time between the current time and the end of the forecast cycle. When the remaining time is less than 60 minutes, a short-term forecast mode is activated; between 60 and 240 minutes, a medium-term forecast mode is used; and after 240 minutes, a long-term forecast mechanism is switched. The figure compares the actual monitoring values ​​with the results of the three forecast modes. Short-term forecast (5 minutes) has the highest accuracy, medium-term forecast (15 minutes) balances accuracy and forecast duration, and long-term forecast (30 minutes) mainly reflects the changing trends of pollutants. The system achieves accurate forecasts at different time scales through dynamic forecast strategies, providing decision support for air quality management.

[0085] Example 3: See Figure 9The system performs frequency domain transformation on multi-source air quality data streams to extract pollution parameter change cycles. It analyzes the frequency of parameter changes within these cycles to construct a pollution variability degree. Furthermore, it analyzes the decreasing and increasing trends of parameters at multiple adjacent sampling times to construct a trend degree. Based on the pollution variability degree and trend degree, and combined with preset adjustment values, detection and control parameters are generated. The environmental perception system of the dynamic adjustment module uses multi-dimensional signal analysis methods to construct a pollution dynamic model. The change cycle extraction unit deploys a real-time spectrum monitoring mechanism, performing a sliding window Fourier transform on the input data stream. The window width is set to 256 sampling points, and the sliding step size is 32 points. For PM2.5 concentration signals, a Blackman window function is used to suppress spectral leakage, while a flat-top window is used to enhance amplitude accuracy for carbon dioxide concentration data. The spectrum analyzer identifies frequency components with significant peaks and establishes a list of periodic components: fluctuations below 0.01Hz are marked as long-term trend terms, the 0.01-0.1Hz range is considered environmental disturbance components, and fluctuations above 0.1Hz are classified as equipment noise. The primary period determination algorithm selects frequency bands with an energy share exceeding 15% as effective periodic components and records their center frequency and half-power bandwidth. When multiple competing periods are detected, the system establishes a period priority queue and sets the component with the closest historical average period as the primary reference.

[0086] The pollution variability calculation engine implements a time-domain differential analysis process. The vector generation subunit creates a circular buffer to store the pollution parameter sequence for the most recent 30 minutes, with sampling intervals strictly aligned with the system clock pulse. The differential processor performs two levels of differential operations: first-order differential calculation calculates the absolute change between adjacent sampling points, and second-order differential analysis analyzes the fluctuation characteristics of the change itself. The change frequency detector counts the number of times the differential value exceeds a threshold per unit time, which is dynamically adjusted according to the sensor type: a base threshold of 5 μg / m³ for PM2.5 sensors and a variation threshold of 200 ppm for carbon dioxide sensors. The change amplitude estimator calculates the normalized average fluctuation intensity.

[0087]

[0088] The meanings of each character in the formula are as follows: Indicates the intensity coefficient of periodic fluctuations. This represents the total number of sampling points contained within the analysis window. This represents the index variable of the sampling point number. Indicates the first The original values ​​of pollution parameters obtained at each sampling time. Indicates the first Each sampling time (i.e.) The pollution parameter values ​​(at the previous moment). This represents the maximum value among all sampled values ​​within the analysis window. arrive The peak values ​​between This represents the minimum value among all sampled values ​​within the analysis window, i.e. arrive The valley between It represents the absolute value of the change in values ​​between adjacent sampling points, reflecting the intensity of instantaneous fluctuations.

[0089] The calculation of the pollution variability needs to be achieved by combining the frequency of change and the average magnitude of change of the first-order difference sequence: ① Frequency of change statistics: using the analysis window (e.g., 30 minutes) as the unit, statistically analyze the first-order difference sequence. The number of times the absolute change between adjacent sampling points exceeds the preset threshold (5 μg / m³ for PM2.5 and 200 ppm for carbon dioxide) is recorded as follows: This number reflects the frequency of fluctuations in pollution parameters; ② Calculation of average change amplitude: based on the above formula... The calculation result is divided by the total number of sampling points. The normalized average change amplitude is obtained and denoted as (i.e., in the formula) (The core component), this amplitude reflects the average intensity of a single fluctuation in the pollution parameter; ③ The final calculation of the degree of pollution variability: through weighted fusion. and The weighting coefficients are dynamically assigned based on the sensor type (e.g., in a PM2.5 sensor). Weight 0.4 Weight 0.6, in the carbon dioxide sensor Weight 0.3 (Weight 0.7), the calculation formula is as follows: ( Weighted by the frequency of change. ),in This refers to the degree of pollution variability; a larger value indicates that the pollution parameters fluctuate more frequently and with higher intensity.

[0090] The trend analysis system employs robust regression techniques to process parameter sequences. The time series modeling sub-unit uses a sliding time window, with the window width maintained as an integer multiple of the detected contamination cycle. The linear fitter uses a Tukey double-weighting function to suppress outliers, calculating the optimal fitted line for each data point within the window. The slope feature extractor performs a three-stage processing: the original slope value is converted to an angle value through an arctangent transformation, then compressed to the 0-1 interval using a sigmoid function, and finally multiplied by the direction sign to preserve trend polarity. The trend stability analysis module monitors the slope change rate over three consecutive windows, triggering a trend reliability downgrade mechanism when fluctuations exceed a preset tolerance. For parameter sequences with significant periodicity, the autocorrelation coefficient of the fitting residuals is additionally calculated to correct the final trend score.

[0091] A fuzzy logic decision-making system is constructed using an adjustment parameter generator. The interference degree calculation subunit is designed with a three-dimensional feature space: the pollution change degree coordinate axis is divided into three fuzzy sets (low, medium, and high), and the trend degree coordinate axis is set into three regions (negative, stable, and positive). The environmental interference degree output is divided into five levels. The fuzzy rule base contains 27 decision rules, such as "when the change degree is high and the trend is negative, the interference degree increases to level four." The loosening degree construction subunit uses an incremental calculation method: the basic loosening degree is mapped from the interference degree level, and then a correction term for the trend degree change rate is superimposed. For continuously deteriorating environmental conditions, the system introduces a time accumulation factor, automatically increasing the loosening degree by 0.1 units every 5 minutes.

[0092] The calculation method for the trend change rate correction term is as follows:

[0093] ① Calculate the autocorrelation of the data: Use the Pearson autocorrelation coefficient. The formula is

[0094]

[0095] in Let be the pollution parameter value at the i-th sampling time. The average value of n sampling points. The value of the pollution parameter at the (i-1)th sampling time;

[0096] ② Calculate the first-order lag difference term: ;

[0097] ③ Trend change rate correction term (when When it is determined to be a significant autocorrelation, this correction item is activated; when At that time, the correction item );

[0098] ④ Corrected trend degree ,in This is the trend value without any correction.

[0099] The feedback adjustment system implements multi-channel parameter control. The sampling frequency regulator adjusts the sampling rhythm according to the looseness level: a 60-second interval is maintained under normal conditions, shortened to 40 seconds for level three looseness, and a 20-second rapid sampling mode is activated for level five. The data verification intensity controller dynamically configures verification rules: the basic mode only verifies the reasonableness of the data range, while the advanced mode adds trend continuity and statistical distribution checks. The sensor operating parameter regulator manages the equipment's operating status: in high looseness environments, it activates the automatic cleaning cycle of the particulate matter sensor and extends the preheating time of the gas sensor by 10%. Adjustment commands are distributed through a priority queue; critical parameters are executed immediately, while auxiliary parameters undergo a gradual adjustment process. All adjustment operations are logged in detail, including adjustment time, original parameters, target values, and execution results, forming a complete adjustment traceability chain. The system evaluates the adjustment effect every 10 minutes and automatically initiates the parameter recovery procedure when the environmental conditions stabilize.

[0100] The environmental status visualization interface presents the dynamic adjustment process. A 3D trend graph displays the pollution change, trend, and loosening curves for the past hour, using color gradients to represent numerical values. A real-time spectrum graph shows the frequency distribution of the main periodic components, highlighting frequency bands marked as environmental disturbances. The parameter adjustment panel lists all active control items, including adjustable parameters such as sampling interval, calibration rules, and operating modes. The historical record query function supports retrieving adjustment logs by time range and generates a time distribution statistical chart of adjustment operations. The interface automatically pops up when the system detects a sustained high-interference state.

[0101] Example 4: The trend degree construction unit uses a time series segmented modeling method to process pollution parameter data. The system is configured with a circular data buffer to store the concentration values ​​of the most recent 20 sampling points, with the sampling interval synchronized with the sensor acquisition cycle. The linear regression analyzer performs robust fitting on the data in the buffer: first, it calculates the median value of all data points, excludes outliers that deviate from the median value by more than three standard deviations, and then uses an iterative weighted least squares method to fit the optimal trend line. The slope calculation module implements a dynamic compensation mechanism; when significant autocorrelation is detected in the data, a difference term with a lag of one order is automatically added as a correction factor. The trend intensity quantification process includes a three-step transformation: the original slope value is first multiplied by the sampling interval to convert it into a rate of change per unit time, then compressed to the [-1,1] interval using a hyperbolic tangent function, and finally the absolute value is taken to generate a non-directional trend degree in the range of 0-1. The system configures differentiated trend sensitivity parameters for different types of pollutants; the trend threshold for particulate matter concentration is set to 0.15, and the judgment standard for gas concentration is 0.08.

[0102] The environmental disturbance assessment system constructs a multi-dimensional feature fusion model. The principal component analysis engine receives standardized pollution change and trend data and calculates the eigenvalues ​​and eigenvectors of their covariance matrix. The first principal component direction reflects the overall intensity of environmental change, while the second principal component characterizes the balance between the rate and magnitude of change. The disturbance degree calculator maps the principal component scores to a three-dimensional classification space: the first dimension distinguishes between periodic fluctuations and random disturbances, the second dimension identifies gradual and abrupt patterns, and the third dimension determines the persistence of change. The system maintains a disturbance pattern knowledge base containing feature templates for 12 typical environmental scenarios and determines the matching pattern of the current state through nearest neighbor search.

[0103] Determining the degree of environmental disturbance requires simultaneously considering the correlation between pollution variability and trend, as well as their average value. Average value calculation: ① Average pollution variability: Using a 5-minute sliding time window, calculate the pollution variability for all values ​​within the window. ( , The arithmetic mean of the number of samples within the window. The formula is ② Average trend score: Using the same 5-minute sliding window, calculate the trend score for all values ​​within the window. ( , The arithmetic mean of the number of samples within the window. The formula is ③ Correlation calculation: The correlation between the degree of pollution variability and the degree of trend is characterized by the covariance matrix. , A positive value indicates a positive correlation between the two (the trend is more obvious when pollution fluctuations are frequent), and a negative value indicates a negative correlation; ④ Interference level fusion calculation: , , Together, they serve as input data for the principal component analysis engine (rather than just inputting standardized pollution variability and trend), among which... and Used to characterize the basic strength of environmental disturbances Used to characterize the correlation features of interference; in principal component analysis, the score of the first principal component is determined by... (Weight 0.3) (Weight 0.3) (Weight 0.4) The weighted average is used to construct the final result based on the score of the first principal component, which maps the degree of environmental interference (e.g., a score of 0-0.3 indicates low interference, 0.3-0.6 indicates medium interference, and 0.6-1.0 indicates high interference).

[0104] The loosening degree calculation module employs a hybrid architecture combining fuzzy inference and reinforcement learning. The fuzzy controller takes standardized disturbance degree, trend change rate, and historical stability index as input variables and outputs a preliminary loosening degree suggestion. The reinforcement learning agent evaluates the long-term regulatory benefits under different loosening degrees using a Q-network. The network state space contains 10 environmental feature dimensions, and the action space defines 5 levels of loosening degree regulation policies. The experience replay buffer stores the most recent 1000 regulation records, including environmental state, executed actions, and result feedback data. The policy optimizer performs offline training every 24 hours, updating the Q-network weight parameters. The actual loosening degree calculation combines the fuzzy output and the network suggestion value; when the difference between the two exceeds a threshold, an expert arbitration mechanism is activated.

[0105] Table 1: Correspondence between Typical Environmental Conditions and Loosening Adjustment

[0106] Interference mode type Trend characteristics Variation level Principal component scores Recommended looseness Periodic fluctuations Oscillating upward middle (0.72,0.15) 0.45 Equipment start-up and shutdown disturbances steep drop recovery high (0.31,0.89) 0.68 Continued deterioration linear descent high (0.95,0.02) 0.82 steady state Horizontal fluctuations Low (0.05,0.11) 0.15 Random noise Disorder jump Extremely high (0.18,0.97) 0.73

[0107] The parameter adjustment execution system implements a hierarchical control strategy. The core parameter adjuster directly interfaces with the sensor hardware, allowing real-time modification of sampling frequency, signal gain, and filtering parameters. For particulate matter sensors, pulse mode is activated when the looseness exceeds 0.5, shortening the sampling interval from 60 seconds to 30 seconds; for gas sensors, high-precision range is switched when the looseness reaches 0.7, sacrificing response speed for measurement accuracy. The auxiliary parameter controller manages the data preprocessing process: simple moving average filtering is used under low looseness conditions, while Kalman filtering is switched to under high looseness conditions. The system status monitor continuously tracks the adjustment effect, triggering a re-evaluation process of the adjustment strategy when the actual measured value deviates from the expected improvement curve by more than a tolerance threshold.

[0108] A feedback mechanism is used to construct a closed-loop optimization system. The performance evaluation module calculates the rate of change of key indicators before and after adjustment, including the improvement in data stability and the reduction in outliers. The adjustment effect quantifier converts multi-dimensional evaluation results into a unified benefit score for updating the strategy knowledge base. A historical adjustment case storage system retains successful adjustment records in typical scenarios, forming a triplet database containing environmental features, adjustment actions, and effect scores. When a new environmental pattern is detected, the system initiates a similar case retrieval function, prioritizing historically effective strategies. An adaptive learning algorithm analyzes long-term adjustment data, automatically adjusting the weight distribution of the fuzzy rule base and the reward function parameters of the Q-network.

[0109] The system features a visual interface that dynamically displays the adjustment process. The environmental status dashboard simultaneously displays the original measurement curves and the adjusted data stream, using different colors to distinguish processing stages. The trend analysis chart shows the changes in disturbance level, trend level, and looseness over the past 6 hours, and marks the time points of significant adjustment events. The adjustment decision dashboard lists all currently effective parameter modifications, including modification time, original value, target value, and actual execution results. The system provides an adjustment history playback function, supporting browsing the complete environmental status evolution process and corresponding adjustment measures along a timeline. When a persistently high-interference environment is detected, the interface automatically pops up strategy optimization suggestions and displays a comparison chart of the expected effects of alternative adjustment schemes.

[0110] An anomaly handling system ensures the reliability of the adjustment process. Sensor fault detectors monitor hardware status indicators, automatically marking nodes as unreliable data sources when signal drift or abnormal responses are detected. The adjustment safety monitor sets safety boundaries for parameter modifications to prevent over-adjustment from causing system instability. The emergency recovery module automatically saves a snapshot of the current adjustment strategy in the event of a system crash or power outage, prioritizing the restoration of critical parameter settings upon restart. The version control system maintains a record of adjustment algorithm iterations, supporting rapid rollback to historical stable versions.

[0111] The system maintenance module enables automated management. The calibration reminder predicts the next calibration time and provides advance notice based on sensor usage time and accuracy degradation curves. The resource optimizer analyzes and calculates load distribution, performing strategy pre-training and data archiving operations during idle periods. A remote diagnostic interface allows authorized engineers to view real-time adjustment logs and submit parameter adjustment suggestions. A self-test program automatically runs hardware diagnostic tests daily to verify the response characteristics of each sensor channel. A configuration wizard guides administrators in setting environmental parameters, including basic information such as space volume, ventilation conditions, and typical pollution source characteristics.

[0112] A data synchronization mechanism ensures collaborative operation among multiple devices. The main controller periodically broadcasts an environmental status summary, including the current looseness value and the identification results of major interference patterns. After receiving broadcast information from the devices, each node adjusts its local strategy according to its role: monitoring nodes prioritize data acquisition quality optimization, while execution nodes prioritize control response speed. A clock synchronization protocol keeps the time error of each device within 50 milliseconds, ensuring the consistency of adjustment actions. A conflict resolution algorithm handles disagreements in adjustment suggestions from multiple nodes, employing a weighted voting mechanism to determine the final execution plan.

[0113] The status report system generates structured documents. The execution summary section outlines the current environmental quality level and key control measures. The trend analysis section displays time-series charts and statistical characteristics of key parameters. The control record table details all parameter modifications and their execution context. The recommended actions section proposes equipment maintenance and system upgrade plans based on continuous monitoring results. Report output supports multiple formats, including PDF documents, structured JSON data, and visual HTML pages. The document generator has a built-in template engine, allowing users to customize report content and style. All reports are automatically archived and indexed for full-text search, supporting quick retrieval by time range, environmental status, or control type.

[0114] Example 5: The joint detection output module establishes a multi-source data integration framework. The deviation calculation unit implements a hierarchical comparison strategy and configures a differentiated evaluation standard system: absolute difference is used to calculate temperature parameters, relative deviation percentage is used to assess humidity values, and segmented proportional deviation method is used for particulate matter concentration (absolute deviation is used for values ​​below 50 μg / m³, and relative deviation is used for values ​​above 50 μg / m³). The planned air quality baseline values ​​are stored in a multi-dimensional array, with each dimension containing three reference levels: national standard limits, industry recommended values, and user-defined targets. The deviation evaluator performs a synchronous comparison operation, simultaneously calculating the deviation of three types of data—current monitoring values, short-term forecast values, and medium- to long-term forecast values—from the baseline. The deviation results generate a deviation coefficient in the 0-1 range through a standardization converter. This coefficient integrates the actual measured deviation magnitude and the duration of the deviation. The system establishes a deviation trend prediction model, deducing future trend directions based on the deviation changes over the past three hours.

[0115] The report generation engine constructs a dynamic document architecture. The core framework adopts a hierarchical tree structure: the root node records the overall air quality rating, first-level branches contain environmental parameter reports, and second-level nodes store detailed data records. The spatial distribution model divides the monitoring area into 0.5m × 0.5m grid cells, with each cell storing the thermal value of pollutant concentration. The time axis processing module creates a dual-time dimension system: the physical time axis marks actual monitoring points by minute, and the prediction time axis constructs virtual time markers using prediction step sizes. The data association establishes a four-fold index relationship: the parameter type index connects different pollutant datasets, the spatial location index associates regional measurements, the time index links historical and prediction data, and the equipment index binds to the status of actuators.

[0116] The visualization system enables collaborative display of multiple views. The main monitoring view adopts an instrument panel layout, with a 3D spatial heat map displayed in the central area, overlaid with sensor node location markers; the right panel displays real-time monitoring curves, supporting simultaneous display of the changing trends of six parameters; the lower left is set up as a prediction comparison area, displaying actual values ​​and predicted curves in a semi-transparent overlay; the lower right is configured with an adjustment parameter panel, dynamically scrolling to display the most recently activated control strategies. The time navigator supports status backtracking at any point in time, including a complete historical monitoring value sequence and records of executed adjustment parameters. The spatial analysis tool provides a cross-sectional view function, which can arbitrarily capture a spatial plane to display a profile of pollutant distribution. The equipment status view independently displays the working mode and remaining life prediction of each actuator.

[0117] The control command generator implements an intelligent decision-making process. The strategy matching engine queries the rule knowledge base to search for corresponding solutions based on the current deviation type and degree. When particulate matter concentration deviates slightly, the equipment list filtering algorithm is triggered: prioritizing effective purification equipment covering the area exceeding the standard, excluding units under maintenance or already at maximum load. The timing optimizer plans the equipment start-up and shutdown sequence: purifiers with fast response times are started first, and high-power equipment is turned on during off-peak hours to avoid circuit overload. The runtime prediction model calculates the time required to reach the target value, taking into account factors such as equipment purification efficiency, space volume, and pollutant diffusion rate. The multi-device collaborative solution implements load balancing detection to ensure that the total system power consumption does not exceed the circuit safety threshold.

[0118] The data management platform constructs a full lifecycle storage solution. The raw database employs a tiered storage design: the real-time layer caches data from the last two minutes, the operation layer retains daily report information, and the archive layer compresses and stores historical records. Storage formats are optimized based on data type: monitoring values ​​are stored in a time-series database, configuration parameters are saved in a document database, and spatial information is stored in a graph database. A secure storage mechanism implements a dual-node hot standby strategy, automatically switching to the backup node in case of primary node failure. Data cleanup strategies follow multi-level retention rules: raw monitoring values ​​are retained for 15 days, processed feature data for 90 days, and statistical reports for three years. All transmitted data is end-to-end encrypted, and a dynamic key rotation mechanism protects communication security.

[0119] The report output system supports multimodal generation schemes. The basic report format includes a text summary section, data statistics tables, and trend curves. Figure 3Essential elements. The enhanced visualization report adds 3D spatial animation, dynamically demonstrating the pollutant diffusion path and concentration evolution process. The simplified mobile version extracts key indicator summaries and automatically adapts to mobile screen sizes to display core information. The machine-readable version outputs in JSON-LD format, containing complete structured data and semantic annotation information. The document release process is configured with multi-level approval rules; reports of major exceedance events require confirmation from two levels of administrators. The version management system automatically records report modification history, ensuring data traceability. All output reports embed digital watermarks, including tamper-proof identifiers such as generation time, device number, and operator information.

[0120] The anomaly handling module incorporates a fault-tolerant mechanism. A data integrity monitor detects input stream anomalies: when three consecutive predicted values ​​exceed a reasonable range, a self-checking procedure for the prediction model is triggered; when real-time monitoring data is missing for more than five cycles, the system automatically switches to a backup signal source. The report generation fault tolerance mechanism includes checkpoints: a document structure validator verifies the integrity of report nodes, and a data consistency detector checks the logical relationships between values. In emergency situations such as power outages, the system automatically saves a snapshot of the current state to non-volatile memory, allowing report generation to resume after power is restored. Critical operations include a secondary confirmation process, such as modifying parameter baseline values ​​and executing device start / stop commands.

[0121] The equipment maintenance module integrates status early warning functionality. The health assessment system analyzes historical sensor calibration data and usage frequency to establish equipment reliability growth curves. The lifespan prediction model, based on cumulative operating time and workload parameters, provides early warnings of equipment replacement needs three months in advance. The maintenance plan optimizer coordinates equipment maintenance cycles, avoiding simultaneous downtime for multiple critical pieces of equipment. The spare parts management system automatically records consumable usage and issues purchase requests before the filter replacement cycle ends. The maintenance record subsystem details the operation content, replaced parts, and verification results for each service, forming a complete equipment history file. The remote diagnostic interface allows equipment manufacturers to access and view operating logs, submit maintenance suggestions, and provide technical support solutions.

[0122] The system's coordinated control achieves a closed-loop environmental optimization. The control command distributor compiles the adjustment scheme into equipment control codes: the air purification equipment receives switch commands, airflow setpoints, and countdown parameters; the fresh air system acquires the supply airflow setpoints and target temperature and humidity values; and the central air conditioning controller reads the temperature control curve setting scheme. The execution status feedback loop continuously monitors the actual control effect: when the execution deviation continuously exceeds the set threshold, a recalculation process for the adjustment parameters is triggered. The abnormal interruption control mechanism immediately activates the emergency plan when the equipment reports a fault signal: stopping the operation of the problematic equipment, reallocating the load to the standby unit, and updating the control strategy section of the report. The system periodically evaluates control effectiveness and generates an adjustment strategy optimization suggestion document, which is added as an attachment to the main report.

[0123] The user interaction system provides multi-channel services. The voice assistant receives natural language query commands and can perform tasks such as data retrieval for specific time periods or regional comparative analysis. The message push service automatically sends early warning information to preset contacts during major pollution events. The self-service analysis workstation supports a drag-and-drop interface, allowing users to customize the arrangement of data views and analysis dimensions. A multi-user collaborative mode with a permission management system is implemented: administrators have full control, technicians can modify system parameters, and observers are only allowed to view real-time data. User behavior logs record key operational events, including report export times, view switching records, and parameter adjustment operations. The help system has a built-in contextual guidance function, dynamically displaying relevant operation prompts based on the current interface status.

[0124] The data archiving process employs an intelligent compression algorithm. Raw monitoring data is incrementally stored, recording only values ​​where changes exceed a preset threshold. Preprocessed feature data undergoes lossy compression, optimizing storage space within a set accuracy loss range. Analysis reports are stored using version differential technology; the base template retains the complete version, while subsequent updates only store the changed content. The archiving index system constructs a multi-dimensional query structure, supporting searches by time window, spatial location, pollution type, and device status. Long-term storage media utilizes an automated tiered mechanism: data within three months is stored on solid-state storage, data within one year is transferred to high-speed hard drives, and historical data is transferred to a tape library. The data destruction process adheres to international information security standards, performing three overwrite / erase operations when physical storage media becomes obsolete.

[0125] The environmental scenario modes support customized configuration. The system pre-configures six typical scenario templates, including a home mode, a meeting room mode, an industrial workshop mode, and a medical cleanroom mode. When switching modes, the corresponding benchmark system and evaluation criteria are automatically loaded. The scenario transfer learning function records user adjustment behavior and automatically applies optimized parameter configurations in similar scenarios. The cross-scenario comparison tool can display the difference curves of evaluation indicators under different scenarios side-by-side. The scenario recommendation engine analyzes historical environmental data and suggests the operating mode most suitable for the current space characteristics. The scenario parameter export function supports configuration template sharing, allowing users to migrate optimized mode settings among themselves.

[0126] The peripheral system interfaces are designed with standardization in mind. The device access layer adopts industrial IoT communication protocols, supporting common industrial bus protocols such as Modbus and BACnet. Third-party system interface adapters can output data from this system to building management platforms or smart home control systems. The open API interface follows RESTful specifications, providing twelve core data access and command execution services. The external data import module supports access to related information systems such as weather forecasts and traffic flow. The system upgrade and maintenance channel adopts a differential update mechanism, and the component-based deployment allows for the individual updating of specific functional modules without service interruption. The interface performance monitor periodically checks the transmission quality of each communication channel and automatically generates connection quality reports to be attached to the main system operation log.

[0127] The output quality control system implements full-process verification. A data traceability mechanism records the processing of each value: conversion records from raw to standard values, feature extraction parameter configuration, prediction model version information, etc. A rationality validator scans the numerical logic of the output report, including temporal consistency, spatial distribution rationality, and equipment status matching. A format specification tool ensures consistency in data content and presentation accuracy across different output formats. The version control system locks the report release status, prohibiting unauthorized modifications to issued reports. The metadata management system automatically annotates environmental condition parameters used to generate the report, including atmospheric pressure, electromagnetic environment, and other factors that may affect measurement accuracy. An integrity check procedure verifies the existence of all required sections and data elements before final output.

[0128] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0129] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A joint indoor air quality detection system integrating spatiotemporal graph neural networks, characterized in that, include: The multi-source sensor data acquisition module is equipped with various indoor air quality sensors to collect multi-source air quality data streams. The spatiotemporal map construction module is used to construct spatiotemporal map data based on the sensor spatial locations and time series in the multi-source air quality data stream; The fusion processing module is used to perform feature fusion processing on the spatiotemporal map data using a spatiotemporal map neural network to generate an air quality fusion feature set. The prediction strategy module is used to determine the prediction strategy based on the remaining time between the current time and the end time of the prediction cycle, and to perform air quality prediction according to the air quality fusion feature set. The dynamic adjustment module is used to analyze the real-time changes of the multi-source air quality data stream, construct the pollution variability and trend, and adjust the detection and control parameters according to the pollution variability and trend. The joint detection output module is used to generate a joint indoor air quality detection report based on the output of the prediction strategy and the adjustment result of the dynamic adjustment module.

2. The indoor air quality joint detection system integrating spatiotemporal graph neural networks as described in claim 1, characterized in that, The multi-source sensor data acquisition module includes: Multiple indoor air quality sensors, including temperature and humidity sensors, particulate matter sensors, and gas concentration sensors; The noise analysis unit is used to perform noise characteristic analysis on the multi-source air quality data stream and determine the noise data denoising threshold. The preprocessing unit is used to perform noise identification and filtering on the multi-source air quality data stream according to the noise data denoising threshold, and generate a standard multi-source air quality data stream.

3. The indoor air quality joint detection system integrating spatiotemporal graph neural networks as described in claim 2, characterized in that, The fusion processing module includes: The feature extraction unit is used to determine the spatiotemporal feature extraction method and feature type based on the data characteristics of the standard multi-source air quality data stream; The correlation feature generation unit is used to extract correlation features from the standard multi-source air quality data stream according to the spatiotemporal feature extraction method and feature type to obtain a multi-source signal correlation feature set. The weight allocation unit is used to set dynamic weight allocation rules based on sensor reliability and accuracy; The fusion execution unit is used to perform weighted fusion of the multi-source signal correlation feature set based on the weight dynamic allocation rule to generate the air quality fusion feature set.

4. The indoor air quality joint detection system integrating spatiotemporal graph neural networks as described in claim 1, characterized in that, The prediction strategy module includes: The remaining duration calculation unit is used to calculate the remaining duration between the current time and the end time of the prediction period; The prediction period determination unit is used to determine the first prediction period by matching the remaining time with a preset conversion relationship; The strategy generation unit is used to determine the adjustment period based on the number of energy types or the number of equipment, combined with the second conversion relationship, and to calculate the second prediction period by the difference between the first prediction period and the adjustment period. The prediction execution unit is used to control the prediction model to process the air quality fusion feature set according to the second prediction cycle and output future air quality prediction data.

5. The indoor air quality joint detection system integrating spatiotemporal graph neural networks as described in claim 4, characterized in that, The prediction execution unit includes: The model calling subunit is used to call a pre-trained spatiotemporal graph neural network prediction model; The data input subunit is used to input the air quality fusion feature set and connection relationships into the spatiotemporal graph neural network prediction model; The prediction output subunit is used to perform a prediction operation based on the second prediction period to generate the future air quality prediction data.

6. The indoor air quality joint detection system integrating spatiotemporal graph neural networks as described in claim 1, characterized in that, The dynamic adjustment module includes: The variation period extraction unit is used to perform frequency domain transformation on the multi-source air quality data stream and extract the variation period of pollution parameters. The variability construction unit is used to analyze the frequency of parameter changes within the change cycle of the pollution parameters and construct the pollution variability. The trend building unit is used to analyze the downward and upward trends of parameters at multiple adjacent sampling times and to build the trend. The adjustment parameter generation unit is used to generate detection control parameters based on the pollution variability and trend, combined with a preset adjustment amount.

7. The indoor air quality joint detection system integrating spatiotemporal graph neural networks as described in claim 6, characterized in that, The degree of change construction unit includes: The vector generation subunit is used to generate a periodic vector by arranging pollution parameters according to a time series. The difference analysis subunit is used to calculate the first-order difference sequence of the periodic vector; The change measurement operator unit is used to calculate the degree of pollution change based on the change frequency and average change amplitude of the first-order difference sequence.

8. The indoor air quality joint detection system integrating spatiotemporal graph neural networks as described in claim 6, characterized in that, The trend construction unit includes: The time series modeling subunit is used to fit a time series trend line based on pollution parameters at multiple adjacent sampling times through linear regression analysis; The trend quantification subunit is used to extract the slope feature of the time series trend line and generate the trend degree based on the value of the slope feature.

9. The indoor air quality joint detection system integrating spatiotemporal graph neural networks as described in claim 6, characterized in that, The adjustment parameter generation unit includes: The interference degree calculation subunit is used to determine the degree of environmental interference based on the correlation and average value of the pollution variability and trend. A looseness construction sub-unit is used to calculate the detected looseness based on the degree and trend of environmental interference; The feedback parameter generation subunit is used to generate feedback detection control parameters based on the difference in the detected looseness and the preset adjustment amount.

10. The indoor air quality joint detection system integrating spatiotemporal graph neural networks as described in claim 1, characterized in that, The joint detection output module includes: The deviation calculation unit is used to compare future air quality forecast data with preset air quality plans and generate deviation data. The report generation unit is used to integrate the deviation data and the detection control parameters to generate the joint indoor air quality detection report.

Images