Air quality data screening method and system based on multi-factor deviation analysis

Abstract

The invention discloses an air quality data screening method and system based on multi-factor deviation analysis, and solves the problem that complex spatio-temporal data deviation is difficult to process by a traditional method. The method comprises the following steps: acquiring multi-city multi-time-dimension air pollutant monitoring values, meteorological data, topographic data and monitoring station metadata, performing data cleaning through three-level quality control, and constructing a spatio-temporal data cube; dynamically establishing a regional reference value, calculating a combined deviation, and identifying a deviation mode through a space-time DBSCAN method; a city deviation feature vector and a provincial deviation network are constructed, and a graph convolution network is used for sorting; establishing a multi-layer cause and effect graph, and diagnosing deviation root causes by adopting a PC algorithm and a Shapley value decomposition technology; and defining a Markov decision process, and dynamically optimizing a correction strategy. According to the method, the composite deviation can be accurately quantified, root causes such as regional pollution, equipment faults or meteorological interference can be distinguished, and data support is provided for regional collaborative governance.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of air quality monitoring, and particularly relates to an air quality data screening method and system based on multi-factor deviation analysis. BACKGROUND

[0002] With the acceleration of urbanization process and the continuous expansion of industrial scale in China, the accuracy and reliability of air quality monitoring data have become the core foundation of environmental governance and public health protection. However, air quality data is often affected by sensor errors, external environmental changes and other factors, causing data deviation. Firstly, monitoring data is prone to be disturbed by sensor drift, extreme weather and equipment failure, and traditional methods are difficult to dynamically adapt to multi-source heterogeneous data, resulting in insufficient or false correction of abnormal values. Secondly, the existing technology mainly adopts single pollutant threshold judgment, ignores the interaction effect of pollutants and the spatio-temporal correlation, and cannot distinguish regional pollution diffusion and equipment failure. Finally, deviation detection mainly relies on traditional statistical methods, which are difficult to deal with complex multi-dimensional and time-series data. Therefore, intelligent data screening technology based on deep learning has become an important development direction of future data processing. SUMMARY

[0003] The application aims to provide an air quality data screening method and system based on multi-factor deviation analysis, which can effectively identify cities or provinces with large deviations through multi-level data processing and multi-dimensional feature learning of air quality factors, and provide dynamic optimization and correction strategies.

[0004] To achieve the above-mentioned purpose, the application provides the following technical solutions. In a first aspect, the application provides an air quality data screening method based on multi-factor deviation analysis, comprising: Obtaining air pollutant monitoring values, meteorological data, terrain data and monitoring station metadata of multiple cities in multiple time dimensions, performing data cleaning through three-level quality control, interpolating and aligning the spatio-temporal data and constructing a spatio-temporal data cube; Dynamically establishing a regional reference value by fusing spatial weighted average observation value and time sliding window historical mean value, calculating combined deviation based on single deviation absolute value and pollutant interaction, classifying the deviation through a spatio-temporal DBSCAN method, and identifying deviation patterns; Constructing a city deviation feature vector to calculate a city comprehensive deviation index, establishing a provincial deviation network with cities as nodes and atmospheric influence coefficients as edges, and performing provincial deviation sorting through a graph convolution network; Constructing a multi-layer causal graph containing a device layer and an environment layer, identifying pollution propagation paths by using a PC algorithm, and quantifying the contribution of each factor to the total deviation by using Shapley value decomposition technology; The Markov decision process is defined, a correction strategy is selected according to the current bias feature and the environment state, and the data quality improvement is evaluated through a reward function and the strategy is optimized.

[0005] Further, the three-level quality control comprises: first-level rejection of equipment failure outliers, second-level correction of meteorological outliers, and third-level standardization processing, which converts the concentrations of various pollutants into comparable values in a standard state. The interpolation alignment of the spatio-temporal data and the construction of the spatio-temporal data cube comprise: uniformly interpolating data of different time resolutions to hourly data, using the Kriging interpolation method to generate regular grid data for sites with uneven spatial distribution, and establishing a spatio-temporal data cube with dimensions of [time x space x pollutant].

[0006] Further, the calculation formula of the absolute value of the single-item bias is: wherein, is the absolute value of the bias of the i-th factor, is the measured value, is the standard value; The calculation formula of the combined bias is: wherein, is the weight of the single-item bias, is the interaction coefficient of pollutants k and l, and λ is the weight of the interaction term. The interaction coefficient is determined through principal component analysis, which can identify the correlation between pollutants.

[0007] Further, the bias is classified and the bias pattern is identified through the spatio-temporal DBSCAN method, specifically including spatially aggregated bias, temporally synchronized bias, and randomly discrete bias. The spatially aggregated bias indicates a possible regional pollution source, the temporally synchronized bias indicates a systematic error, and the randomly discrete bias indicates a local disturbance or random fluctuation.

[0008] Further, the calculation formula of the urban comprehensive bias index is: wherein, is the feature weight of city i, which is dynamically calculated according to the entropy weight method; is the urban bias feature vector of city i; wherein, is the average bias of city i at time t; is the maximum bias of city i at time t; is the frequency of bias greater than the threshold value θ; is the spatial autocorrelation.

[0009] Furthermore, the device layer focuses on the device correlation between monitoring sites. Each monitoring point is regarded as a node in the causal graph. The nodes in the device layer represent the individual monitoring stations, and the edges represent the data association between the sites. The environment layer focuses on the impact of meteorological factors on the spread of pollutants.

[0010] Furthermore, the Markov decision process includes: State space S t : Describes the current system status, including deviation type, weather conditions and equipment status; Action space a t : represents the correction strategy selected at time t, including physical correction method, statistical interpolation method and learning correction method; Reward function r t : Measures the degree of improvement in the system state after performing an action, ,in, is the improvement in data quality after correction is performed, is the cost of choosing a certain correction strategy; Transition probability: The transition between states is determined based on the current state and the selected action, and is estimated using historical data.

[0011] In a second aspect, the present invention provides an air quality data screening system based on multi-factor deviation analysis, comprising: Data preprocessing module: used to perform multi-source data collection, three-level quality control and spatiotemporal data alignment, and construct spatiotemporal data cube; Joint deviation detection module: used to dynamically generate regional benchmark values, calculate combined deviations, and identify deviation patterns; Hierarchical sorting module: used to construct and sort the city-provincial deviation network; Root cause diagnosis module: used to construct multi-layer causal diagrams and quantify deviation contributions; Dynamic correction module: used to generate and optimize correction strategies.

[0012] Furthermore, the joint deviation detection module includes: Benchmark value generation unit: integrates the spatial weighted average observation value and the historical mean of the time sliding window, and dynamically weights the average through the Bayesian model; Interaction coefficient calculation unit: used to determine the interaction coefficient between pollutants.

[0013] Furthermore, the root cause diagnosis module includes: Causal graph construction unit: used to build a double-layer causal graph of the device layer and the environment layer; Propagation path analysis unit: uses PC algorithm to identify key pollution propagation paths; A contribution decomposition unit is configured to decompose the contribution of each factor to the total deviation using a Shapley value.

[0014] Based on the above technical solutions, the embodiments of the present application can at least produce the following technical effects: The air quality data screening method based on multi-factor deviation analysis provided by the present application combines dynamic reference values and a multi-factor weighted deviation model to accurately quantify composite deviations; through spatiotemporal DBSCAN clustering and causal diagram analysis, regional pollution, equipment failure or meteorological interference and other root causes are distinguished. The entropy weight method is used to dynamically weight the city-level deviation characteristics, and a provincial GCN network is constructed based on an atmospheric diffusion model to quantify cross-city pollution transmission, providing data basis for regional collaborative governance, realizing a "detection-cause-correction" closed-loop optimization, and providing core technical support for accurate environmental supervision. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained from the structures shown in the drawings without creative labor.

[0016] Figure 1 is a flowchart of the embodiments of the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely as follows. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the fact that those skilled in the art can realize it. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, and is not within the scope of protection claimed by the present application.

[0018] The purpose of the present application is achieved by the following technical solutions: Embodiment 1, the present application provides an air quality data screening method based on multi-factor deviation analysis, comprising the following steps: Step 1: data preprocessing Obtain multi-city, multi-time dimension air pollutant monitoring values, meteorological data, terrain data and monitoring station metadata, perform data cleaning through three-level quality control, interpolate and align the spatiotemporal data, and construct a spatiotemporal data cube.

[0019] Specifically, it is necessary to obtain the monitoring values of multiple air pollutants of multiple cities in different time dimensions, synchronously collect meteorological data and terrain data, and obtain monitoring station metadata. Different time dimensions include daily, weekly, and monthly, and multiple air pollutant monitoring values include PM2.5, PM10, SO2, NO2, CO, and O3. Meteorological data includes wind speed, wind direction, temperature, humidity, and air pressure. Monitoring station metadata includes device model, maintenance record, and surrounding environment. Three levels of quality control are implemented on the above-mentioned obtained data: Level 1: remove abnormal values caused by device failure (such as constant values); Level 2: correct abnormalities caused by weather conditions (such as PM10 during sandstorms); and Level 3: standardize processing, convert the concentrations of each pollutant to comparable values under standard conditions. Finally, spatio-temporal alignment processing is performed, the data of different time resolutions are uniformly interpolated to hourly data, the Kriging interpolation method is used for the sites with uneven spatial distribution to generate regular grid data, a spatio-temporal data cube is established, and the dimensions include [time × space × pollutant].

[0020] Step 2: Multi-factor combined bias detection A regional reference value is dynamically established by fusing spatially weighted average observation values and time sliding window historical mean values, combined bias is calculated based on single bias absolute value and pollutant interaction, and bias is classified by spatio-temporal DBSCAN method to identify bias patterns.

[0021] Specifically, first, the regional background value of each time slice (t) is dynamically calculated, and two kinds of reference values are fused by Bayesian Model Averaging (BMA): spatially weighted average observation values (after removing ±3σ outliers) and time sliding window historical mean values (same period data in the past 30 days). For each monitoring point, for each air quality factor (such as SO2, CO, NO2, O3, PM2.5, PM10), the absolute value of its single bias is calculated: wherein, is the absolute value of the bias of the i-th factor, is the measured value, is the standard value; The combined bias is the weighted sum of multiple single biases, considering the interaction between pollutants: wherein, is the weight of the single bias, is the interaction coefficient of pollutants k and l, and λ is the weight of the interaction term, is the bias value of the i-th factor at time t, the k-th index, is the bias value of the i-th factor at time t, the l-th index, and the interaction coefficient The correlation between pollutants can be identified by principal component analysis (PCA).

[0022] Finally, the bias is classified by spatio-temporal DBSCAN (density clustering) method to identify bias patterns: Spatially aggregated bias: indicates possible regional pollution sources. Spatial autocorrelation (such as Moran's index) can be used to measure it; Temporally synchronized bias: indicates systematic errors, which may be related to equipment failure or network transmission problems; Randomly dispersed bias: indicates local interference or random fluctuations, which may be caused by individual monitoring point equipment problems or short-term weather events.

[0023] Step 3: Spatio-temporal hierarchical ranking analysis Construct the city bias feature vector to calculate the city comprehensive bias index, establish a provincial bias network with cities as nodes and atmospheric influence coefficients as edges, and perform provincial bias ranking through graph convolution network.

[0024] First, calculate the city bias feature vector, and the bias feature vector of each city Including multiple indicators: Among them, is the average bias of city i at time t; is the maximum bias of city i at time t; is the frequency of bias greater than threshold θ; is the spatial autocorrelation, calculated by methods such as Moran's index.

[0025] Then calculate the city comprehensive bias index according to the weight of the feature vector: Among them, , , , is the feature weight calculated dynamically according to the entropy weight method.

[0026] Finally, perform provincial ranking and establish a provincial network, where the comprehensive bias index of each city is a node, and the edges between cities are the atmospheric influence coefficients between cities calculated based on atmospheric diffusion models (such as CALPUFF model). Use graph convolution network (GCN) to perform convolution operation on the provincial bias network: Among them, A is the adjacency matrix, representing the atmospheric influence coefficient between cities, GCN is the graph convolution operation, which propagates and calculates the comprehensive bias of each city through the adjacency relationship between nodes.

[0027] Step 4: Bias root cause diagnosis A multi-layer causal graph is constructed including a device layer and an environmental layer, a PC algorithm is used to identify the pollution propagation path, and a Shapley value decomposition technique is used to quantify the contribution of each factor to the total bias.

[0028] It should be noted that the core goal of the propagation path analysis is to identify the propagation path of the pollutants and determine which factors or variables have a significant impact on the distribution and bias of the pollutants. The device layer focuses on the device correlation between monitoring sites. Each monitoring point can be regarded as a node in the causal graph. The monitoring device may affect the accuracy of the data due to sensor failure, calibration problem, data transmission error, etc. Therefore, the nodes in the device layer will represent the monitoring stations, and the edges will represent the data correlation between these stations. The device layer analysis helps to find the bias caused by device problems, for example, the high reading of pollutants in a certain area may be due to the inaccurate calibration of the monitoring device. The environmental layer mainly focuses on the influence of meteorological factors (such as wind speed, temperature, humidity, air pressure, etc.) on the propagation of pollutants. The pollutants in the air will spread or settle with meteorological factors such as wind direction, wind speed and air temperature, affecting the concentration and distribution of pollutants. For example, strong winds may cause pollutants to spread over a large area, while high temperatures may exacerbate photochemical reactions to generate ozone. Through this causal relationship, the contribution of environmental factors to pollution bias can be inferred. A multi-layer causal graph is constructed to analyze the pollution propagation path: Among them, the device layer considers the device correlation between monitoring stations, and the environmental layer considers the influence of meteorological factors on the diffusion of pollutants.

[0029] Specifically, a causal discovery algorithm (such as the PC algorithm) is used to identify the key pollution propagation path. The PC algorithm is a causal discovery algorithm based on conditional independence test. In the initial stage, it is assumed that there is a causal relationship between all variables, and the irrelevant edges are removed step by step through conditional independence test. The conditional independence between each pair of variables is checked step by step, and the independence of each pair of variables under the control of other variables is calculated. Statistical tests (such as Pearson correlation-based or Granger causality test) are used to determine whether there is a causal relationship. Through the PC algorithm, a directed acyclic graph (DAG, Directed Acyclic Graph) can be constructed, which can help identify the path of pollution propagation. For example, through this causal relationship graph, it may be found that the abnormal change of some pollutants is caused by device failure or diffusion effect under certain meteorological conditions.

[0030] It is worth noting that the goal of contribution decomposition is to allocate the total deviation contribution to each factor or variable. Shapley value is a method derived from cooperative game theory to measure the contribution of each participant in cooperation. In our scenario, each factor (such as meteorological factors, equipment errors, etc.) can be regarded as a participant, and Shapley value can be used to decompose the contribution of these factors to the total deviation. Using Shapley value decomposition technology, the total deviation is explained, and the contribution of each factor to the total deviation is decomposed: where, is the Shapley value of the kth factor, representing the contribution of the factor to the total deviation. v(S) is the deviation explanation degree of the subset S, and N is the set of all influencing factors. By calculating the Shapley value, we can quantitatively evaluate the influence of each factor on the total deviation. Assuming there are multiple factors (such as different pollutant concentrations, meteorological factors, and monitoring equipment errors), Shapley value helps us identify which factors play a dominant role in pollutant deviation.

[0031] Step 5: Dynamic adaptive correction Define a Markov decision process to select a correction strategy based on the current deviation characteristics and environmental state, and evaluate the data quality improvement through a reward function and optimize the strategy.

[0032] Specifically, the Markov decision process includes: state space: state is the current system description, including deviation characteristics and environmental state. For example, the state can include the current pollutant concentration deviation, meteorological conditions (such as wind speed, temperature, etc.), and the state of the monitoring equipment (such as equipment failure identification).

[0033] Action space: action represents the correction strategy selected at time t. For example, the action can be to select a certain physical method (such as chemical mass balance correction), statistical method (such as EM algorithm interpolation), learning method (such as LSTM correction), etc.

[0034] Reward function: reward Measures the degree of improvement of the system state after performing a certain action. The reward can be measured by the improvement of data quality, such as: where, is the improvement of data quality after performing correction, is the cost (such as computational complexity, time consumption, etc.) of selecting a certain correction strategy.

[0035] Transition probability (Transition Probability): The transition between states is based on the current state and the selected action The transition probabilities can be estimated from historical data.

[0036] Embodiment 2, the application provides an air quality data screening system based on multi-factor bias analysis, comprising: a data preprocessing module for performing multi-source data collection, three-level quality control and spatio-temporal data alignment, and constructing a spatio-temporal data cube; a joint bias detection module for dynamically generating regional reference values, calculating combined bias and identifying bias patterns; a hierarchical sorting module for constructing a city-provincial bias network and sorting; a root cause diagnosis module for constructing a multi-layer causal graph and quantifying bias contribution; a dynamic correction module for generating and optimizing correction strategies.

[0037] In this embodiment, the joint bias detection module comprises: a reference value generation unit that fuses spatially weighted average observations and time sliding window historical mean values, and dynamically weights through Bayesian model averaging; an interaction coefficient calculation unit for determining the interaction coefficient between pollutants.

[0038] In this embodiment, the root cause diagnosis module comprises: a causal graph construction unit for establishing a two-layer causal graph of the device layer and the environment layer; a propagation path analysis unit that uses the PC algorithm to identify key pollution propagation paths; a contribution decomposition unit that uses Shapley values to decompose the contribution of each factor to the total bias.

[0039] In the implementation process, the air quality monitoring data bias analysis of the North China urban agglomeration (January 2024) is as follows: Step 1: Data preprocessing (1) Data collection: a. Cities: Shijiazhuang, Taiyuan, Beijing, Tianjin, Jinan, Zhengzhou (6 cities in total); b. Time range: January 1-31, 2024 (hourly data); c. Pollutants: PM2.5, PM10, SO2, NO2, CO, O3 (6 items in total); d. Meteorological data: wind speed, wind direction, temperature, humidity, air pressure (source: China Meteorological Administration); e. Topographic data: city elevation, surrounding mountain distribution (source: geographic information system); f. Equipment metadata: City monitoring station equipment models (Thermo Fisher, etc.), maintenance records (calibration time, fault log).

[0040] (2) Data cleaning: Level 1: Remove the constant value of PM2.5 (120 μg / m 3 ) at a certain station in Taiyuan on January 5 due to sensor failure for 12 hours. Level 2: Correct the PM10 data during the sandstorm in Shijiazhuang on January 15 (original value 800 μg / m 3 → corrected to 620 μg / m 3 according to the meteorological model). Level 3: Convert all pollutant concentrations to standard state (25°C, 1 atm).

[0041] (3) Spatio-temporal alignment: Interpolate daily / weekly data to hourly data (use linear interpolation method); Use Kriging interpolation to generate 1 km x 1 km grid data for Jinan City with sparse sites; Construct spatio-temporal cube dimension: [Time: 744 hours x Space: 6 cities x 50 grid points x Pollutants: 6 items].

[0042] Step 2: Multi-factor combined bias detection (1) Benchmark value establishment: Spatial weighted average: remove outliers (e.g. Beijing PM2.5 > 250 μg / m 3 ); Time sliding window: take the same period hourly average of the past 30 days (2023 December data); BMA fusion weight: spatial benchmark weight 0.6, temporal benchmark weight 0.4.

[0043] (2) Bias quantification: Calculate single-item bias: PM2.5 measurement value at a certain station in Shijiazhuang on January 10 at 12:00 = 98 μg / m 3 , benchmark value = 65 μg / m 3 → | δ i | = 33 μg / m 3 ; Combined bias calculation: Weight distribution: PM2.5 weight 0.3 (main control pollutant), interaction term λ = 0.2; PCA analysis shows that PM2.5 is strongly correlated with NO2 (correlation coefficient 0.7), and the interaction coefficient β_{PM2.5,NO2} = 0.5.

[0044] (4) Deviation pattern recognition: Temporal-spatial DBSCAN clustering results: Spatial clustering: All sites in Tianjin Binhai New Area had synchronous SO2 deviation (Moran's index I = 0.85, p < 0.01), pointing to industrial emission sources; Temporal synchronization: All sites in Zhengzhou had a sudden drop in NO2 concentration from 9:00 to 11:00 on January 20 (equipment network transmission failure); Random dispersion: Individual sites in Beijing had short-term O3 peaks (local barbecue interference).

[0045] Step 3: Temporal-spatial hierarchical ranking analysis (1) City-level ranking: (2) Provincial ranking (Hebei Province as an example): Shijiazhuang became a key pollution retention node due to topography (Taihang Mountain blockage) (provincial deviation contribution 32%).

[0046] Step 4: Deviation root cause diagnosis (1) Propagation path analysis: Device-level causal diagram: Taian site A failure → affecting adjacent site B through data communication protocol (edge weight = 0.9); Environmental-level causal diagram: Northwest wind (wind speed > 4 m / s) → Shijiazhuang pollution transmission to Jinan (path coefficient = 0.75); PC algorithm identifies key path: [wind direction → PM2.5 deviation] (p < 0.05).

[0047] (2) Contribution degree decomposition: Step 5: Dynamic adaptive correction (1) MDP framework: State: S t ={deviation type = spatial clustering, wind speed = 3.2 m / s, device status = normal} Action: Choose physical correction method (chemical mass balance model to reconstruct data) Reward: r t = data quality improvement ΔQ - cost C, ΔQ = deviation reduction 22 μg / m 3 (mass improvement score + 15), C = model calculation time cost (-2) Net reward: r t = 13 (2) Strategy optimization: Update action value through Q-learning: (learning rate α = 0.1, discount factor γ = 0.9).

[0048] The foregoing is considered as illustrative only of the principles of the application and the forms thereof which are demonstrated by the described embodiments. Further, those skilled in the art will readily recognize that certain of the above described embodiments can include, inter alia, alternate configurations and steps thereof known in the art. Therefore, there is no intention that this application be limited as described unless by the recitations of the claims that follow, or by appropriate words taken in the broader and / or more general sense.

Claims

1. A method for screening air quality data based on multi-factor deviation analysis, characterized in that: include: Acquire multi-city, multi-temporal air pollutant monitoring values, meteorological data, topographic data, and monitoring station metadata. Perform data cleaning through three-level quality control, interpolate and align spatiotemporal data, and construct a spatiotemporal data cube. Dynamically establish regional benchmark values ​​by integrating spatially weighted average observations with historical averages of temporal sliding windows, calculate combined deviations based on the absolute values ​​of individual deviations and pollutant interactions, and classify deviations and identify deviation patterns using the spatiotemporal DBSCAN method; Construct city deviation feature vectors to calculate the city comprehensive deviation index, establish a provincial deviation network with cities as nodes and atmospheric impact coefficients as edges, and use graph convolutional networks to sort provincial deviations. Construct a multi-layer causal diagram including the equipment layer and the environment layer, use the PC algorithm to identify the pollution propagation path, and use the Shapley value decomposition technique to quantify the contribution of each factor to the total deviation; Define a Markov decision process, select a correction strategy based on the current deviation characteristics and environment state, evaluate data quality improvement through a reward function, and optimize the strategy.

2. The air quality data screening method based on multi-factor deviation analysis according to claim 1 is characterized in that: The three-level quality control includes: first-level elimination of equipment failure outliers, second-level correction of meteorological outliers, and third-level standardization to convert the concentration of each pollutant into a comparable value under standard conditions; The interpolation and alignment of spatiotemporal data and construction of a spatiotemporal data cube include: uniformly interpolating data of different time resolutions into hourly data, using Kriging interpolation method to generate regular grid data for sites with uneven spatial distribution, and establishing a spatiotemporal data cube with dimensions including [time×space×pollutant].

3. The air quality data screening method based on multi-factor deviation analysis according to claim 1 is characterized in that: The calculation formula for the absolute value of the single deviation is: ； in, is the absolute value of the deviation of the i-th factor, is the measured value, is the standard value; The calculation formula of the combined deviation is: ； in, is the weight of the single deviation, is the interaction coefficient between pollutants k and l, λ is the weight of the interaction term, is the deviation value of the kth indicator of the i-th factor at time t, is the deviation value of the i-th factor at time t and the l-th indicator.

4. The air quality data screening method based on multi-factor deviation analysis according to claim 1 is characterized in that: The spatiotemporal DBSCAN method is used to classify deviations and identify deviation patterns, specifically including spatially clustered deviations, time-synchronized deviations, and random discrete deviations; spatially clustered deviations indicate possible regional pollution sources, time-synchronized deviations indicate systematic errors, and random discrete deviations indicate local interference or random fluctuations.

5. The air quality data screening method based on multi-factor deviation analysis according to claim 1 is characterized in that: The calculation formula of the city comprehensive deviation index is: ； in, is the characteristic weight of city i, which is dynamically calculated according to the entropy weight method; is the city deviation eigenvector of city i; ； in, is the average deviation of city i at time t; is the maximum deviation of city i at time t; is the frequency at which the deviation is greater than the threshold θ; is the spatial autocorrelation.

6. The air quality data screening method based on multi-factor deviation analysis according to claim 1 is characterized in that: The equipment layer focuses on the equipment correlation between monitoring sites. Each monitoring point is regarded as a node in the causal graph. The nodes in the equipment layer represent each monitoring station, and the edges represent the data association between sites. The environment layer focuses on the impact of meteorological factors on the spread of pollutants.

7. The air quality data screening method based on multi-factor deviation analysis according to claim 1 is characterized in that: The Markov decision process includes: State space S t : Describes the current system status, including deviation type, weather conditions and equipment status; Action space a t : represents the correction strategy selected at time t, including physical correction method, statistical interpolation method and learning correction method; Reward function r t : Measures the degree of improvement in the system state after performing an action, ,in, is the improvement in data quality after correction is performed, is the cost of choosing a certain correction strategy; Transition probability: The transition between states is determined based on the current state and the selected action, and is estimated using historical data.

8. An air quality data screening system based on multi-factor deviation analysis, used to implement the method according to any one of claims 1 to 7, characterized in that: include: Data preprocessing module: used to perform multi-source data collection, three-level quality control and spatiotemporal data alignment, and construct spatiotemporal data cube; Joint deviation detection module: used to dynamically generate regional benchmark values, calculate combined deviations, and identify deviation patterns; Hierarchical sorting module: used to construct and sort the city-provincial deviation network; Root cause diagnosis module: used to construct multi-layer causal diagrams and quantify deviation contributions; Dynamic correction module: used to generate and optimize correction strategies.

9. The air quality data screening system based on multi-factor deviation analysis according to claim 8, characterized in that: The joint deviation detection module includes: Benchmark value generation unit: integrates spatial weighted average observations and temporal sliding window historical mean values, and dynamically weights them using a Bayesian model. Interaction coefficient calculation unit: used to determine the interaction coefficient between pollutants.

10. The air quality data screening system based on multi-factor deviation analysis according to claim 8, characterized in that: The root cause diagnosis module includes: Causal graph construction unit: used to build a double-layer causal graph of the device layer and the environment layer; Propagation path analysis unit: uses PC algorithm to identify key pollution propagation paths; Contribution decomposition unit: Use Shapley value to decompose the contribution of each factor to the total deviation.

Images