Air quality monitoring station site selection recommendation method based on active learning

Abstract

The invention discloses an air quality monitoring station site selection recommendation method based on active learning, and the method comprises the steps: firstly, carrying out the estimation of information amount, and enabling an air quality monitoring station site selection site to meet the following two basic standards at the same time: employing a Monte Carlo technology to calculate the uncertainty of an unmonitored position, employing the Euclidean distance to carry out the similarity estimation, quantizing the maximized influence; the uncertainty and the influence are integrated into information amount; meanwhile, the information amount of various air pollutants is integrated; and finally, by recommending the positions of the air quality monitoring stations in batches, solving a position set with the maximum information amount. According to the method, the calculation cost is reduced, the accuracy of the air quality inference graph is improved to the maximum extent, and the requirements of inference of various air pollutants can be met at the same time.

Description

Technical Field

[0001] The present invention belongs to environmental monitoring technology, and in particular to an air quality monitoring station site selection recommendation method based on active learning. Background Art

[0002] Air quality monitoring stations are a key means of obtaining air quality information, but their current number is insufficient to meet the demand for fine-grained air quality data. Therefore, with the development of social economy and the improvement of public demand for air quality, the deployment of new air quality monitoring stations has become an inevitable trend. The layout of new air quality monitoring stations should be able to maximize the accuracy of air quality inference maps, thereby providing more reliable data support for environmental management and policy making. However, due to the high construction and maintenance costs of air quality monitoring stations, the site selection of each monitoring station must be comprehensively and carefully considered. There are three main problems in previous studies on the site selection recommendation of air quality monitoring stations. First, these studies only rely on the uncertainty of the location itself [1][2] or the impact of the location on other unmonitored locations [3] to make site selection recommendations, resulting in poor performance on some data sets. Second, these studies adopt a single model when recommending the location of air quality monitoring stations, that is, after determining the recommendation priority of a location, the inference model needs to be retrained. This not only leads to high computational costs but also increases the risk of overfitting of the inference model [4]. Finally, air quality monitoring stations usually need to monitor multiple air pollutants, but existing studies only consider a single pollutant when recommending the location of monitoring stations, which obviously does not meet actual needs. Therefore, it is urgent to develop a more comprehensive and effective method to optimize the site selection recommendation method for air quality monitoring stations.

[0003] References:

[0004] [1]Hsieh HP,Lin SD,Zheng Y.Inferring air quality for stationlocation recommendation based on urban big data[C] / / Proceedings of the 21thACM SIGKDD international conference on knowledge discovery and datamining.2015:437-446.

[0005] [2]Yang Y, Zheng Z, Bian K, et al.Sensor deployment recommendation for 3D fine-grained air quality monitoring using semi-supervised learning[C] / / 2018IEEE International Conference on Communications(ICC).IEEE,2018:1-6.

[0006] [3]Kang Y,Chen J,Cao Y,et al.A higher-order graph convolutionalnetwork for location recommendation of an air-quality-monitoring station[J].Remote Sensing,2021,13(8):1600.

[0007] [4]Yin C, Qian B, Cao S, et al.Deep similarity-based batch mode activelearning with exploration-exploitation[C] / / 2017IEEE international conferenceon data mining(ICDM).IEEE,2017:575-584. Summary of the invention

[0008] In order to maximize the accuracy of air quality inference maps by optimizing the site selection of monitoring stations, the present invention provides an air quality monitoring station site selection recommendation method based on active learning.

[0009] The present invention provides an air quality monitoring station site selection recommendation method based on active learning, comprising the following steps:

[0010] Step 1: Estimation of the amount of information.

[0011] The site selection for an air quality monitoring station must meet two basic criteria:

[0012] (1) Basic Standard 1: Maximizing Uncertainty.

[0013] Air quality monitoring stations should be located where inference uncertainty is greatest.

[0014] The Monte Carlo dropout technique is used to calculate the uncertainty of unmonitored locations. The uncertainty of the inferred value consists of two parts: one is the epistemic uncertainty from the inference model, and the other is the random uncertainty from the data noise. To fully capture these two types of uncertainty, it is assumed that the concentration distribution of air pollutants conforms to the Gaussian distribution, and the output of the inference model is set to the average value of the air pollutant concentration. and its variance σ, rather than the labels themselves; therefore, the uncertainty of the set of recommended locations S is defined as:

[0015]

[0016] Among them, σ i is the uncertainty of position i, is i The maximum-minimum normalized value of .

[0017] (2) Basic Standard 2: Maximize impact.

[0018] Maximum impact includes two key requirements: first, the recommended locations should be highly representative, that is, have a large similarity with other unmonitored locations; if a location is similar to a location previously encountered by the inference model, the model's inference accuracy for that location is likely to be high; second, the recommended locations should have low redundancy, that is, have the lowest similarity with the locations of existing air quality monitoring stations; marking similar locations has limited contribution to improving the inference accuracy of air quality maps.

[0019] Based on the analysis of the maximum impact, accurate similarity estimation is obviously a key prerequisite for accurate impact estimation; similarity estimation is performed in the feature space learned by the inference model; in the feature space learned by the inference model, the similarity matrix A is defined using the Euclidean distance SI , as follows:

[0020]

[0021] Among them, w ij is the inverse Euclidean distance between positions i and j, is the similarity after maximum-minimum normalization, h i，d is the feature vector of position i in dimension d, where D is the feature dimension.

[0022] Based on the similarity matrix A SI , the influence of the recommended location set S is defined as formula (3), which can simultaneously meet the two requirements of maximizing influence;

[0023]

[0024] Among them, L is the set of locations with air quality monitoring stations, and U is the set of locations without air quality monitoring stations.

[0025] Specifically, if the positions in S are not representative enough or are highly redundant, the value of C(S) will be low.

[0026] (3) Integrate two basic standards.

[0027] The amount of information Q(S) is defined as the weighted average of uncertainty H(S) and influence C(S), and the formula is as follows:

[0028] Q(S)=αH(S)+(1-α)C(S), 0<α<1 (4)

[0029] Among them, α is a weight parameter.

[0030] (4) Integration of information on multiple air pollutants.

[0031] When recommending sites, the impact of various pollutants must be comprehensively considered; the information of multiple pollutants is weighted averaged to ensure that the recommended sites can achieve a balance in improving the accuracy of inference of various pollutants. The formula is as follows:

[0032]

[0033] Among them, Q r is the information content of the rth air pollutant, β r is its corresponding weight, R is the total amount of air pollutants, is the total amount of information.

[0034] Step 2: Find the set of locations with the maximum amount of information.

[0035] Deploying air quality monitoring stations at locations with the most information can maximize the accuracy of the air quality inference map; therefore, the objective function is defined as follows:

[0036]

[0037] Solution It is NP-hard; therefore, an optimization algorithm is proposed to solve this problem. The algorithm recommends the locations of air quality monitoring stations in batches until the predetermined number of recommendations is reached; in each batch, the inference model is first trained based on L, and then the uncertainty and impact of each location are extracted from the model, and its information content is calculated; then the locations are sorted from large to small according to the amount of information, and the first-ranked location is added to the set L; after updating the impact, the process is repeated until the number of locations added to the set L reaches the batch number.

[0038] The beneficial technical effects of the present invention are:

[0039] The present invention considers the uncertainty and influence of the location to calculate the amount of information, and ranks the locations from large to small according to the amount of information. The higher the ranking, the easier it is to recommend the location. In addition, the present invention adopts a batch mode to reduce the computing cost, that is, after recommending a batch of monitoring station locations, the inference model is retrained to update the amount of information of each location. More importantly, the site recommendation method performs a weighted average of the location information calculated for multiple air pollutants to generate a comprehensive amount of information, so that the recommended location can simultaneously meet the needs of inferring multiple air pollutants. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Flow chart for selecting recommended locations in each batch for the present invention. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0042] The present invention provides an air quality monitoring station site selection recommendation method based on active learning, comprising the following steps:

[0043] Step 1: Estimation of the amount of information.

[0044] Active learning aims to select the most informative instances to label in order to maximize the performance of the model.

[0045] In order to maximize the accuracy of the air quality inference map, the locations with the most information should be recommended to deploy air quality monitoring stations. Therefore, the site selection of air quality monitoring stations must meet the following two basic criteria at the same time:

[0046] (1) Basic Standard 1: Maximizing Uncertainty.

[0047] A simple intuition is that deploying air quality monitoring stations at locations with the lowest inference accuracy can maximize the accuracy of the air quality inference map. However, due to the lack of true labels for unmonitored locations, we cannot directly know the inference accuracy of unmonitored locations. Fortunately, uncertainty can represent the model's confidence in the inferred value, thereby reflecting the accuracy of the inference to a certain extent. Generally speaking, higher uncertainty means greater inference difficulty and lower accuracy. Therefore, air quality monitoring stations should be built at locations with the greatest inference uncertainty.

[0048] The Monte Carlo dropout technique (MC-Dropout) is used to calculate the uncertainty of unmonitored locations. The uncertainty of the inferred value consists of two parts: one is the epistemic uncertainty from the inference model, and the other is the random uncertainty from the data noise. To fully capture these two types of uncertainty, it is assumed that the concentration distribution of air pollutants conforms to the Gaussian distribution, and the output of the inference model is set to the average value of the air pollutant concentration. and its variance σ, rather than the labels themselves; therefore, the uncertainty of the set of recommended locations S is defined as:

[0049]

[0050] Among them, σ i is the uncertainty at position i, is i The maximum-minimum normalized value of .

[0051] (2) Basic Standard 2: Maximize impact.

[0052] Another intuition is that it would be very beneficial to deploy air quality monitoring stations at locations that have the greatest impact on other unmonitored locations. Maximum impact contains two key requirements: first, the recommended locations should have high representativeness, that is, have a large similarity with other unmonitored locations; if a location is similar to a location that the inference model has encountered before, the model's inference accuracy for that location is likely to be high; second, the recommended locations should have low redundancy, that is, have the lowest similarity to the locations of existing air quality monitoring stations; marking similar locations has limited contribution to improving the inference accuracy of air quality maps.

[0053] Based on the above analysis of the maximum impact, accurate similarity estimation is obviously the key prerequisite for accurate impact estimation. Similarity estimation is usually performed in feature space. In the present invention, similarity estimation is performed in the feature space learned by the inference model. Compared with the original feature space, this feature space is closer to the real feature space, which not only reduces some errors but also eliminates some redundant information. In the feature space learned by the inference model, the similarity matrix A is defined using the Euclidean distance SI , as follows:

[0054]

[0055] Among them, w ij is the inverse Euclidean distance between positions i and j, is the similarity after maximum-minimum normalization, h i，d is the feature vector of position i in dimension d, where D is the feature dimension.

[0056] Based on the similarity matrix A SI, the influence of the recommended location set S is defined as formula (3), which can simultaneously meet the two requirements of maximizing influence;

[0057]

[0058] Among them, L is the set of locations with air quality monitoring stations, and U is the set of locations without air quality monitoring stations.

[0059] Specifically, if the positions in S are not representative enough or are highly redundant, the value of C(S) will be low.

[0060] (3) Integrate two basic standards.

[0061] Uncertainty and impact play a vital role in the site selection recommendation of air quality monitoring stations. Therefore, the amount of information Q(S) is defined as the weighted average of uncertainty H(S) and impact C(S), as follows:

[0062] Q(S)=αH(S)+(1-α)C(S), 0<α<1 (4)

[0063] Among them, α is a weight parameter.

[0064] (4) Integration of information on multiple air pollutants.

[0065] Air quality monitoring stations are usually composed of multiple air pollutant sensors that are used to measure multiple air pollutants. When recommending a site, the impact of various pollutants must be considered comprehensively. Generally speaking, the most favorable location for inferring a certain air pollutant may not be ideal for other pollutants. Therefore, a weighted average of the information of multiple pollutants is used to ensure that the recommended location can achieve a balance in improving the accuracy of inferring various pollutants. The formula is as follows:

[0066]

[0067] Among them, Q r is the information content of the rth air pollutant, β r is its corresponding weight, R is the total amount of air pollutants, is the total amount of information.

[0068] Step 2: Find the set of locations with the maximum amount of information.

[0069] Deploying air quality monitoring stations at locations with the most information can maximize the accuracy of the air quality inference map; therefore, the objective function is defined as follows:

[0070]

[0071] Solution It is NP-hard. Fortunately, the greedy algorithm can provide an approximation of the optimal solution. Therefore, an optimization algorithm is proposed to solve this problem. The algorithm recommends the locations of air quality monitoring stations in batches until a predetermined number of recommendations is reached; in each batch, the process of selecting the recommended locations is as follows Figure 1 As shown in the figure, the inference model is first trained based on L, and then the uncertainty and impact of each position are extracted from the model, and its information content is calculated; then the positions are sorted from large to small according to the amount of information, and the first-ranked position is added to the set L; after updating the impact, the process is repeated until the number of positions added to the set L reaches the batch number.

Claims

1. A method for recommending site selection for air quality monitoring stations based on active learning, characterized in that: The following steps are involved: Step 1: Estimation of information quantity; The site selection for an air quality monitoring station must meet two basic criteria: (1) Basic Standard 1: Maximizing uncertainty; Air quality monitoring stations should be located where inference uncertainty is greatest; The Monte Carlo dropout technique is used to calculate the uncertainty of unmonitored locations. The uncertainty of the inferred value consists of two parts: one is the epistemic uncertainty from the inference model, and the other is the random uncertainty from the data noise. To fully capture these two types of uncertainty, it is assumed that the concentration distribution of air pollutants conforms to the Gaussian distribution, and the output of the inference model is set to the average value of the air pollutant concentration. and its variance σ, rather than the labels themselves; therefore, the uncertainty of the set of recommended locations S is defined as: Among them, σ i is the uncertainty of position i, is i The maximum-minimum normalized value of ; (2) Basic Standard 2: Maximize impact; The maximum impact includes two key requirements: first, the recommended locations should be highly representative, that is, have a large similarity to other unmonitored locations; if a location is similar to a location that the inference model has encountered before, the model's inference accuracy for that location is likely to be high; second, the recommended locations should have low redundancy, that is, have the lowest similarity to the locations of existing air quality monitoring stations; marking similar locations has limited contribution to improving the inference accuracy of air quality maps; Based on the analysis of the maximum impact, accurate similarity estimation is obviously a key prerequisite for accurate impact estimation; similarity estimation is performed in the feature space learned by the inference model; in the feature space learned by the inference model, the similarity matrix A is defined using the Euclidean distance SI , as follows: Among them, w ij is the inverse Euclidean distance between positions i and j, is the similarity after maximum-minimum normalization, h i，d is the feature vector of position i in dimension d, where D is the feature dimension; Based on the similarity matrix A SI , the influence of the recommended location set S is defined as formula (3), which can simultaneously meet the two requirements of maximizing influence; Among them, L is the set of locations with air quality monitoring stations, and U is the set of locations without air quality monitoring stations; Specifically, if the positions in S are not representative enough or are highly redundant, the value of C(S) will be low; (3) Integrate two basic standards; The amount of information Q(S) is defined as the weighted average of uncertainty H(S) and influence C(S), and the formula is as follows: Q(S)=αH(S)+(1-α)C(S),0<α<1 (4) Among them, α is a weight parameter; (4) Integration of information on multiple air pollutants; When recommending sites, the impact of various pollutants must be comprehensively considered; the information of multiple pollutants is weighted averaged to ensure that the recommended sites can achieve a balance in improving the accuracy of inference of various pollutants. The formula is as follows: Among them, Q r is the information content of the rth air pollutant, β r is its corresponding weight, R is the total amount of air pollutants, is the total amount of information; Step 2: Find the location set with the maximum amount of information; Deploying air quality monitoring stations at locations with the most information can maximize the accuracy of the air quality inference map; therefore, the objective function is defined as follows: Solution It is NP-hard; therefore, an optimization algorithm is proposed to solve this problem. The algorithm recommends the locations of air quality monitoring stations in batches until the predetermined number of recommendations is reached; in each batch, the inference model is first trained based on L, and then the uncertainty and impact of each location are extracted from the model, and its information content is calculated; then the locations are sorted from large to small according to the amount of information, and the first-ranked location is added to the set L; after updating the impact, the process is repeated until the number of locations added to the set L reaches the batch number.

Images