A method for compensating missing data in environmental monitoring
Through multi-feature weighted embedding and bidirectional gated recursive unit processing air quality data, combining shape and time distortion loss function, the shortcomings of traditional methods in high-frequency continuous missing data processing are solved, and more accurate missing data compensation and time are achieved. Sequence fitting.
Patent Information
- Application Number
- CN202210188052.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-28
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-02-28
AI Technical Summary
Traditional missing data compensation methods cannot effectively maintain the integrity of the time series when processing high-frequency continuous missing data, resulting in deviations or errors in the analysis results.
The multi-feature weighted embedding method is adopted to encode air quality data through a fully connected neural network, and a bidirectional gated recursive unit is used to process the historical and future information at both ends of the missing data, and a shape and time distortion loss function is introduced to optimize the data compensation model.
The accuracy and prediction results of missing data compensation are improved, the shape trajectory similarity of the time series is enhanced, the information redundancy of the model is reduced, and the performance of data compensation is improved.
Smart Images

Figure CN114547985B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of environmental monitoring data processing, and in particular to a method for compensating missing data in environmental monitoring. Background Art
[0002] In data-driven research, missing data has always been a challenge for researchers. In the stage of manual data acquisition, due to the limitations of human recording and management capabilities, data storage is inevitably missing or lost. With the development of information technology, various data monitoring devices have gradually replaced manual work. However, due to systematic or accidental misoperation, these detection equipment data often have various missing values, which makes subsequent analysis and visualization quite difficult, especially in the field of environmental health sciences. As a key part of environmental monitoring and quality analysis, researchers usually use a large number of distributed sensors to obtain air quality data to quantify the concentration of air pollutants in the environment and for analysis and research such as air quality prediction or pollution warning. There are many reasons for missing information in air quality data sets, such as equipment damage, insufficient sampling frequency, hardware problems, and human errors. Incomplete data sets affect the applicability of specific analyses, such as data modeling, which usually requires a complete data matrix. The presence of missing data, no matter how rare, will bias the research results on the relationship between air pollutants and health outcomes.
[0003] In order to obtain a complete data set, researchers must decide whether to discard or interpolate (i.e. compensate) missing data. It is generally unreasonable to ignore missing data. For most time series analysis, a complete time series is a prerequisite because there may be valuable feature information in the missing data, which may affect the reasoning ability. Therefore, most researchers will choose to compensate for the missing data appropriately. For a small number of discrete missing data, the most widely used missing data compensation method is unconditional mean compensation; median compensation is another simple method that is usually suitable for highly skewed data and may produce better results than mean compensation. However, when the missing rate of data gradually increases and there are multiple consecutive missing data in the same time period, traditional statistical methods cannot be well compensated, resulting in unnecessary deviations or erroneous results in subsequent research and analysis.
[0004] In order to solve the problem that traditional data compensation methods destroy the integrity of time series, researchers have carried out relevant research. For example, the last observation is used to fill the gap of missing data to connect the data together, but this method ignores the relationship between variables over time, so its compensation ability is limited. Another example is to regard the missing value compensation problem as data generation, and use the generative adversarial network (GAN) to learn the overall distribution of the multivariate time series data set, and further use it to generate missing values for each sample. Although this method takes into account the protective relationship between variables, the accuracy of the generated model is not stable due to the input of random noise. Summary of the invention
[0005] The purpose of the present invention is to provide a method for compensating missing data in environmental monitoring to solve the problems existing in the above-mentioned prior art. A multi-feature weighted embedding method is introduced to give different weights according to the strength of the correlation between each indicator and the target indicator, and more attention is paid to information that is useful for data compensation. In order not to introduce additional information to produce unnecessary influence on data compensation, it is proposed to use a parallel encoder to process the data at both ends of the missing data, extract its historical and future time information, and apply the attention mechanism to the outputs of the two encoders at the same time, so as to concentrate the information on highly correlated outputs, so as to provide more accurate compensation for the missing data.
[0006] To achieve the above object, the present invention provides the following solution: The present invention provides a method for compensating missing data of environmental monitoring, comprising the following steps:
[0007] Obtain initial environmental monitoring data;
[0008] Analyzing the initial environmental monitoring data to calculate the correlation between each environmental pollution index and the target index, wherein the correlation is used to measure the strength of the correlation between the environmental pollution index and the target index;
[0009] Constructing a data compensation model, inputting the initial environmental monitoring data and the correlation into the data compensation model, training and optimizing the data compensation model using shape and time distortion loss functions, and obtaining an optimal data compensation model;
[0010] The optimal data compensation model is used to compensate for missing data in environmental monitoring.
[0011] Optionally, the method for analyzing the initial environmental monitoring data and calculating the correlation between each environmental pollution index and the target index is:
[0012]
[0013] Among them, X and Y are two different variables, r(·) represents the correlation coefficient between each pollution index and the target index, Cov(·) is the covariance, and σ(·) is the standard deviation.
[0014] Optionally, preprocessing the initial environmental monitoring data based on the correlation to obtain weighted sample data includes:
[0015] Sampling the initial environmental monitoring data to obtain initial sample data;
[0016] Dividing the initial sample data, and converting the data format of the divided initial sample data;
[0017] The converted initial sample data is normalized, different weights are assigned to the normalized sample data based on the relevance, and multi-feature embedding is performed using a fully connected neural network to obtain the weighted sample data.
[0018] Optionally, the initial environmental monitoring data is sampled and a sliding window method is used to obtain initial sample data.
[0019] Optionally, the data compensation model includes an input layer, an embedding layer, an encoder layer, an attention layer, a decoder layer and an output layer, wherein the input layer is used to input the normalized sample data, the embedding layer is used to weight and embed multiple features of the normalized sample data to obtain a feature vector representation; the encoder layer is used to encode the feature vector representation, obtain the intrinsic features of the sequence, and obtain a hidden state vector sequence; the attention layer is used to obtain the weighted sum of the hidden states transmitted by the encoder in each time index t to obtain an attention context vector; the decoder layer is used to decode according to the attention context vector and the decoder state vector, and is responsible for recursively generating an output sequence, and the output layer is used to output the processing results.
[0020] Optionally, the method of multi-feature embedding is:
[0021] f t =F norm (x t,1 , x t,2 , x t,3 , …, x t,N )·(a 1 , a 2 , a 3 , …, a N )
[0022] e t ={e 1 , e 2 , e 3 ,…,e M}=FC(f t )
[0023] Among them, f t is the sample processed in proportion, FC is the nonlinear fully connected layer, e t is the new input vector for multi-feature embedding, F norm is the normalization function, x t,i is the subsequence data, a i is the weight.
[0024] Optionally, the encoder layer includes two encoders, and the encoders adopt bidirectional gated neural units.
[0025] Optionally, the shape and temporal distortion loss functions are:
[0026]
[0027] in, is the shape loss function, is the time loss function, is to penalize the predicted sequence, y is the real time series, and α is a hyperparameter used to divide the weights of shape and time.
[0028] The present invention discloses the following technical effects:
[0029] The present invention provides a method for compensating missing data in environmental monitoring. Aiming at the strong and weak relationship of the correlation between the concentration changes of various pollutant indicators, it proposes to use multi-feature embedding coding, gives different degrees of attention to the correlation between different features, and uses a fully connected neural network to embed and encode multiple features of air quality to obtain a better representation of multi-feature data, thereby improving the prediction results; for the situation where high-frequency continuous missing data occurs, a double-headed sequence-to-sequence data compensation model is proposed, which is used to parallelly process the available data at both ends of the missing data, and uses two encoders to process the time input information at both ends of the missing data. Each encoder is based on a bidirectional gated recurrent unit (BiGRU) to capture the long-term dependence of historical and future information at both ends of the missing data; in order to consider the similarity of the shape trajectory of the time series and enhance the fitting performance of the model, shape and time distortion losses are introduced to improve the similarity between the missing values and the actual time series trajectory. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0031] Figure 1 A schematic diagram of a method flow chart for compensating for missing data in environmental monitoring according to an embodiment of the present invention;
[0032] Figure 2 A schematic diagram of a method for compensating for missing data in environmental monitoring according to an embodiment of the present invention;
[0033] Figure 3 This is a thermal diagram of the Pearson correlation analysis between the variables in the embodiment of the present invention;
[0034] Figure 4 A schematic diagram of a sliding window method used for data sample partitioning in an embodiment of the present invention;
[0035] Figure 5 This is a schematic diagram of the data compensation model structure in an embodiment of the present invention;
[0036] Figure 6 This is a schematic diagram of the principle of the embedding layer in an embodiment of the present invention;
[0037] Figure 7 Schematic diagram of test results of PM2.5, a characteristic value of an air pollution data set under different missing windows in an embodiment of the present invention;
[0038] Figure 8 A visualization diagram of data distribution of the original data in time and shape by the shape and time loss algorithm in an embodiment of the present invention;
[0039] Fig. 9 Schematic diagram of the degree of fitting of the shape and time loss algorithm to the original data in terms of time and shape in an embodiment of the present invention, wherein Figure (a) shows a downward trend when the gap of missing data is 3, Figure (b) shows a downward trend when the gap of missing data is 6, Figure (c) shows an upward trend when the gap of missing data is 3, and Figure (d) shows an upward trend when the gap of missing data is 6. DETAILED DESCRIPTION
[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0042] The present invention provides a method for compensating missing data of environmental monitoring. Figure 1-9 As shown, the following steps are included:
[0043] Obtain initial environmental monitoring data and analyze the initial environmental monitoring data:
[0044] The initial environmental monitoring data is analyzed to calculate the correlation between each environmental pollution indicator and the target indicator, wherein the correlation is used to measure the strength of the correlation between each environmental pollution indicator and the target indicator.
[0045] The Pearson correlation coefficient is used to measure the correlation between multiple environmental pollution indicators and target indicators, and is used to calculate the weight of different variables when implementing compensation. The calculation formula of the Pearson correlation coefficient is as follows:
[0046]
[0047] Among them, X and Y are two different variables, r(·) represents the correlation coefficient between each pollution index and the target index, Cov(·) is the covariance, and σ(·) is the standard deviation.
[0048] Since there are 6 feature variables in this data set (SO 2 ,NO 2 ,SO 3 ,CO,PM10,PM2.5), in order to measure the relevance of each variable to the target variable, in this embodiment, a Pearson correlation analysis was performed on the correlation between the variables, and the results were as follows: Figure 3 Heat map shown.
[0049] The air pollutant data monitored by the environmental monitoring stations in the target area are obtained, the historical data are visualized and analyzed, and appropriate data segments are selected to simulate the situation of missing data for supervised learning. The remaining data are used as the data set for model testing.
[0050] Data preprocessing: Divide the data samples by sliding windows, convert the data samples into supervised learning sequence format, and generate new weighted sample data based on the correlation strength embedding features.
[0051] When using deep learning models to process time series data, an important step is to extract subsequences from the original long time series as training and test samples. To generate samples, first determine the length of the output sequence (the missing data window to be simulated) and the input sequence (the length of the input sequence on the left and right ends is equal), and calculate the total length of each sample sequence. Then fix the left side at the first time index of the time series and move to the total length of a single sequence sample to find a sample until the right side of the window reaches the end of the original long time series. In order to obtain more samples and make the model universal in the case of different available data lengths for missing data compensation, the length of the input sequence is dynamically changed during the sampling process.
[0052]
[0053]
[0054] Among them, T is the original time series, m and o are the minimum lengths of the left and right input sequences, n and p are the maximum lengths of the left and right input sequences, q is the length of the output sequence, S represents all samples, and s in and out Represents input and output sequences.
[0055] Convert data format. Using the sliding window algorithm, the original sample data sequence is divided into three segments. The middle segment is a simulated missing data segment, which is used for supervised learning of the model. The two ends are available data segments. The data on the left and right ends are spliced as input data. The data in the middle segment is the true value of the output data, which is used to measure the effect of data compensation. The lengths of the three segments can be reasonably adjusted according to needs.
[0056] Divide the data set. Divide the sample data set after data format conversion into a training set and a validation set in a ratio of 8:2.
[0057] In this example, six air pollution indicators (NO, SO 2 , PM10, PM2.5, CO, SO 3 ) public dataset, taking PM2.5 as the target indicator, and selecting the data from one of the measurement stations from January to September 2017 as the training set, and the data from October to December 2017 as the test set.
[0058] Construct a data compensation model, input weighted sample data into the data compensation model, train and optimize the data compensation model using shape and time distortion loss functions, and obtain the optimal data compensation model:
[0059] Through preliminary data processing, the initial data set is divided into multiple data samples with a certain time step, that is, the original data set is a data sequence arranged in chronological order. After processing (division), multiple sub-data sequences within a certain time period are obtained. The sub-sequence is a sample in the sample data set. The sub-sequence is normalized, and then different weights are assigned to the processed data according to the analysis of the correlation coefficient. Finally, the data is fed into a fully connected neural network for multi-feature embedding to obtain a new input sequence. The specific calculation is as follows:
[0060] f t =F norm (x t,1 , x t,2 , x t,3 , …, x t,N )·(a 1 , a 2 , a 3 , …, a N )
[0061] e t ={e 1 , e 2 , e 3 ,…,e M}=FC(f t )
[0062] Among them, f t is the sample processed in proportion, FC is the nonlinear fully connected layer, e t is the new input vector for the final multi-feature embedding, x t,i (1≤i≤N) is the subsequence data, a i (1≤i≤N) is the weight (correlation coefficient).
[0063] The multi-feature embedding method of the present invention is an extensible method, which allows researchers to increase or adjust the number of multi-features and the weights of different features according to the richness and importance of supporting data, so that the model has a better prediction effect. Multi-feature embedding can supplement the information of the compensation of target features. Since the numerical changes between the various indicators collected in environmental monitoring will have a certain concomitant or positive correlation, the multi-feature embedding can enrich the model's perception of the rising or falling trend of the sequence and improve the accuracy of data compensation; and weighting each feature in the embedding process can avoid causing information redundancy and damaging the performance of the model in the multi-feature embedding process, because the correlation between the various features has a strong and weak relationship. Through the analysis of the correlation, the effective components in the combined features can be extracted, the dimensionality of the input model can be reduced, and the time dependency of the sequence can be better captured.
[0064] Construct a data compensation model. Figure 5As shown, the data compensation model includes an input layer, an embedding layer, an encoder layer, an attention layer, a decoder layer and an output layer, wherein the input layer is used to input the normalized sample data, the embedding layer is used to weight the data and embed multiple features, the encoder layer encodes the information in the input sequence, obtains the intrinsic features of the sequence, and obtains the hidden state vector sequence, the attention layer extracts highly relevant information from the input sequences processed by the two encoders, so that the data compensation process focuses more on more useful information, the decoder layer decodes the encoded vector, is responsible for recursively generating the output sequence, and the output layer outputs the result of the supplementary processing. Most of the existing deep learning methods will embed the missing data marker into the input sequence, and compensate for the missing data in the overall time series trend features. This method focuses on the time series features of the data around the missing data for more delicate compensation. Therefore, in this embodiment, two encoders are used to receive the available data at both ends of the missing data segment, obtain the hidden information state vector of the data, and then use a global attention mechanism to focus on the hidden state information obtained by the encoder, and finally use the decoder to obtain the final result. The encoder and decoder in this method are both operated using a bidirectional gated neural unit (BiGRU). The bidirectional gated neural unit models the temporal information in both the forward and backward directions. Compared with the standard gated neural unit, the bidirectional gated neural unit combines the forward hidden layer and the backward hidden layer, which can obtain the contextual correlation information in both the forward and backward directions of the input information, and can improve the performance of the recurrent network. That is, each encoder is based on the bidirectional gated recurrent unit (BiGRU), which processes the data from both sides of the missing window separately, so no additional information is needed to locate the missing gap.
[0065] The data compensation model is supervised and learned, the calculation results of the data compensation model are compared with the true values, and the data compensation model is optimized to obtain the best data compensation model.
[0066] Initialize various parameters of the model, input the data of the training set into the model for training until the network converges. In the study of time series prediction, the Euclidean loss function (also known as the mean square error MSE) is often used. However, the MSE method compares the loss of the true value and the predicted value at the same time step, which may cause fluctuation delays or different fluctuation intervals. This method introduces shape and time loss as the loss function, and decomposes the penalty of sequence shape comparison during model training and the penalty of time localization error into two items. The purpose of shape and time loss is to minimize the value of the two penalty items. The calculation formula is as follows:
[0067]
[0068] in, is the shape loss function, is the time loss function, is to penalize the predicted sequence, y is the real time series, and α is a hyperparameter used to divide the weight of shape and time. It is based on dynamic time warping (DTW) and can be expressed as follows:
[0069]
[0070]
[0071] Among them, the binary matrix It's about A i,j The warping path of and j Association A i,j =1, otherwise it is equal to 0. is the set of all valid regular paths connecting endpoint (1, 1) to (k, k). represents the cost matrix, δ is the calculation and j Euclidean distance function for similarity-difference.
[0072] The temporal loss function aims to penalize the prediction sequence And the time distortion degree of the real time series y, the specific calculation can be expressed as follows:
[0073]
[0074] where Z is the partition function, Ω is the penalty matrix
[0075] The performance on missing data is evaluated based on root mean square error (RMSE), mean absolute error (MAE), and dynamic time warping (DTW) scores.
[0076]
[0077]
[0078]
[0079] Among them, f i is the true value of the missing window at the i-th moment, is the predicted value of the missing window at the i-th moment, n is the size of the missing window, A is a binary matrix, is the cost matrix.
[0080] The trained best model was tested on the validation set, and its output compensation value was compared with the true value of the data in the validation sample set. It was verified that this method can improve the performance of missing data compensation for air quality sensors.
[0081] In order to objectively evaluate the effect of the model of this solution and the air quality index data compensation, comparative tests were carried out with the following models in this embodiment.
[0082] KNN: The goal of K-nearest neighbor (KNN) imputation is to find the k nearest neighbors to the observation with missing data and then impute based on the non-missing values in the neighbors.
[0083] MRNN: M-RNN is a multi-directional recurrent neural network that interpolates within and extrapolates across data streams. It provides a promising estimate of missing measurements by applying it to five real-world medical datasets.
[0084] In the experiment, the gaps of missing data were set to 3 and 6, and 10 available sequence data on the left and right of the missing window were selected as the input of the model for supervised learning training. A grid search was performed on the hyperparameters in the data compensation model, and the impact of 1-3 layers of encoders / decoders on performance and the performance of GRU units at different levels of 20, 50, and 100 were tested. In order to prevent overfitting of the model and speed up the convergence of training, the teacher_forcing training mechanism was adopted, and its teacher_forcing_ratio was set to 0.5; for the KNN model, the recommended settings in the Darts software package were used in this test, and the parameter settings of the MRNN model were also consistent with the recommended parameters in the original paper. After testing, the recommended parameter settings of the method used in this test are shown in Table 1:
[0085] Table 1
[0086] Hyperparameters Value Bi-EncoderLayer 1 No. of GRU Units in Encoder 50 Decoder Layer 1 No. of GRU Units in Decoder 50 Dropout 0.5 Optimizer Adam Epoch 1000 Batch Size 15
[0087] Table 2 illustrates the compensation performance of the proposed data compensation model in simulating missing data. It can be seen that the proposed model achieves better results than the other two models under the three evaluation indicators of MAE, RMSE, and DTW. For example, in the experiment of the air pollution data set, when the size of the missing window is 6, the MAE, RMSE, and DTW of the proposed model are 0.1404, 0.1437, and 0.3211, respectively, and its scores are smaller than the results of KNN and MRNN. In the test case with a missing window of 3, the results of the proposed data compensation model also maintain good performance, which shows that the robustness of the algorithm is considerable.
[0088] Table 2
[0089]
[0090] The test results of randomly selected missing areas in different cases are also plotted. Figure 7 The test results of PM2.5, the characteristic value of the air pollution data set, under different missing windows are shown. It can be seen that the data compensation model (Proposal model) provided by the present invention can better simulate the fluctuation trend of the data under different conditions, and its prediction value is also smaller than the error value of the other two models. Although MRNN reflects the fluctuation trend of the data, its error value is large, while KNN not only fails to reflect the fluctuation trend of the real data, but also gives deviated data values for the missing data.
[0091] In order to verify how well the shape and time loss algorithms fit the original data in terms of time and shape, a data set showing similar periodic changes was simulated. A total of 10,000 data values were simulated to form a complete time series, and each value was given a continuous timestamp, recorded every ten minutes. The same experimental configuration was used for the above data set, with three quarters of training and one quarter of testing.
[0092] The performance of the data compensation model provided by the present invention and the conventional MSE loss model under different data missing windows are tested from the two aspects of downward trend and upward trend. Fig. 9 It is not difficult to see that both the data compensation model and the conventional MSE loss model are correct in predicting trends. However, the conventional MSE loss model without shape and time loss will cause slight data fluctuations in the real data. This can be clearly seen in the test samples of the rising trend data. It can be seen that the shape and time loss method can effectively suppress the distortion during model calculation.
[0093] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. They should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.
Claims
1. A method for compensating missing data in environmental monitoring, characterized in that: The following steps are involved: Obtain initial environmental monitoring data; The initial environmental monitoring data is analyzed to calculate the correlation between each environmental pollution index and the target index, wherein the correlation is used to measure the strength of the correlation between the environmental pollution index and the target index: Sampling the initial environmental monitoring data to obtain initial sample data; Dividing the initial sample data, and converting the data format of the divided initial sample data; The converted initial sample data is normalized, different weights are assigned to the normalized sample data based on the correlation, and multi-feature embedding is performed using a fully connected neural network to obtain weighted sample data, wherein the multi-feature embedding method is: in, is the sample processed in proportion, FC is the nonlinear fully connected layer, is the new input vector for multi-feature embedding, F norm is the normalization function, is the subsequence data, is the weight; Construct a data compensation model, input the initial environmental monitoring data and the correlation into the data compensation model, train and optimize the data compensation model using shape and time distortion loss functions, and obtain an optimal data compensation model, wherein the data compensation model includes an input layer, an embedding layer, an encoder layer, an attention layer, a decoder layer, and an output layer, wherein the input layer is used to input the normalized sample data, and the embedding layer is used to perform weighted and multi-feature embedding on the normalized sample data to obtain a feature vector representation; the encoder layer is used to encode the feature vector representation, obtain the intrinsic features of the sequence, and obtain a hidden state vector sequence; the attention layer is used to obtain the weighted sum of the hidden states transmitted by the encoder in each time index t to obtain an attention context vector; the decoder layer is used to decode according to the attention context vector and the decoder state vector, and is responsible for recursively generating an output sequence, and the output layer is used to output the processing result, and the encoder layer includes two encoders, and the encoder adopts a bidirectional gated neural unit; The shape and time distortion loss functions are: in, is the shape loss function, is the time loss function, To penalize the prediction sequence, is a real time series, is a hyperparameter used to divide the weight of shape and time; The optimal data compensation model is used to compensate for missing data in environmental monitoring.
2. The method for compensating missing data for environmental monitoring according to claim 1, characterized in that: The method for analyzing the initial environmental monitoring data and calculating the correlation between each environmental pollution index and the target index is as follows: Among them, X and Y are two different variables, r represents the correlation coefficient between each pollution indicator and the target indicator, Cov(•) is the covariance, and σ(•) is the standard deviation.
3. The method for compensating missing data for environmental monitoring according to claim 1, characterized in that: The initial environmental monitoring data is sampled and the initial sample data is obtained using a sliding window method.
Citation Information
Cited By
An adaptive fusion interpolation method for missing groundwater level segments
CN122777884A
An adaptive fusion interpolation method for missing groundwater level segments
CN122777884B
