Sensing area selection method and device and sensing area classification method and device in sparse crowd sensing

Abstract

The invention discloses a sensing area selection method and device and a sensing area classification method and device in sparse crowd sensing, and relates to the field of data classification, and the method comprises the steps: obtaining sparse spatio-temporal data at a t moment in a target area, and carrying out the preprocessing of the sparse spatio-temporal data, and obtaining a feature matrix corresponding to each data missing sub-area; and according to the feature matrix corresponding to each data missing sub-region, selecting each sub-region to be collected from all the data missing sub-regions by applying a greedy approximation algorithm. And obtaining the training data set of the data updating classification model of the to-be-collected sub-region, and retraining the classification model to obtain the classification model meeting the performance requirement. According to the method for reasonably selecting the sensing area, the area which is more valuable for the classification task can be selected, the data are collected based on the selected sensing area, the representativeness (value) of the collected data is improved, and then the classification precision of the classification model is improved.

Description

Technical Field

[0001] The present application relates to the field of data classification, and in particular to a perception area selection method, a classification method and a device in sparse crowd intelligence perception. Background Art

[0002] With the widespread application of sensor technology in mobile devices, crowd sensing has gradually become an important means of data collection and environmental monitoring in smart cities. However, recruiting a large number of participants often means increasing costs. For this reason, some researchers have proposed a novel crowd sensing paradigm "sparse crowd sensing", which only requires recruiting a small number of participants to collect data from some sensing areas, and accurately infer the data of other unperceived areas by mining the spatiotemporal correlation of the perceived data. Therefore, the new data acquisition mode of sparse crowd sensing can be used to obtain target data, such as wind speed data, urban traffic flow or air quality data, to train the corresponding target classification model. However, the classification model trained based on the existing new data acquisition mode of sparse crowd sensing still has the problem of insufficient classification accuracy. Summary of the invention

[0003] The purpose of this application is to provide a perception area selection method, a classification method and a device in sparse crowd sensing. Studies have shown that the rationality of perception area selection is a key issue in sparse crowd sensing. The rationality of perception area selection directly affects the representativeness (value) of the acquired data, thereby affecting the effect of classification model training, and further affecting the classification accuracy of the classification model. The present invention proposes a method for rationally selecting perception areas, which can select areas that are more valuable to the classification task, collect data based on the selected perception areas, improve the representativeness (value) of the collected data, and further improve the classification accuracy of the classification model.

[0004] To achieve the above objectives, this application provides the following solutions:

[0005] In a first aspect, the present application provides a method for selecting a perception area in sparse crowd sensing, comprising:

[0006] Acquire sparse spatiotemporal data at time t in a target area; the target area includes several sub-areas;

[0007] Preprocessing the sparse spatiotemporal data to obtain a feature matrix corresponding to each data missing sub-region; the feature matrix includes inferred data at a position corresponding to a data missing sub-region and the sparse spatiotemporal data; the data missing sub-region refers to a sub-region where no data is obtained;

[0008] According to the feature matrix corresponding to each data missing sub-region, the greedy approximation algorithm is applied to select the sub-regions to be collected from all data missing sub-regions.

[0009] In a second aspect, the present application provides a data classification method, comprising:

[0010] Obtain historical spatiotemporal data and corresponding classification labels in the target area to build an initial training data set;

[0011] Train the target classification model using the current training data set;

[0012] Determine whether the performance of the currently trained target classification model meets the preset performance requirements;

[0013] If so, the trained target classification model is obtained to classify the data;

[0014] If not, based on the sparse spatiotemporal data at time t in the target area, the perception area selection method in sparse crowd intelligence perception described in any one of claims 1 to 4 is executed to obtain each sub-area to be collected;

[0015] According to the data acquisition mode of crowd intelligence perception, the spatiotemporal data of each sub-area to be collected is acquired;

[0016] Adding the spatiotemporal data of the sub-area to be collected to the sparse spatiotemporal data at time t to obtain the training data at time t; the classification labels of the training data at time t are manually pre-defined;

[0017] Adding the training data at time t to the initial training data set to obtain an updated training data set;

[0018] Use the current updated training data set as the current training data set and return to step “Train the target classification model using the current training data set”.

[0019] In a third aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned perception area selection method in sparse crowd sensing or the above-mentioned data classification method.

[0020] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned perception area selection method in sparse crowd sensing or the above-mentioned data classification method.

[0021] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above-mentioned perception area selection method in sparse crowd sensing or the above-mentioned data classification method.

[0022] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0023] The present application provides a method for selecting a perception region in sparse crowd intelligence perception, a classification method and a device, wherein the sparse spatiotemporal data at time t in the target area is obtained and preprocessed to obtain a feature matrix corresponding to each data missing sub-region; according to the feature matrix corresponding to each data missing sub-region, a greedy approximation algorithm is applied to select each sub-region to be collected from all data missing sub-regions. The training data set of the classification model is updated by obtaining the data of the sub-region to be collected, and the classification model is retrained to obtain a classification model that meets the performance requirements. In the present invention, a greedy approximation algorithm is used to select the data missing sub-region, and the value of a group of data missing sub-regions is jointly evaluated, thereby reducing the information overlap within the combination of data missing sub-regions. In addition, the greedy approximation algorithm can be used to select sub-regions that are more valuable to the classification task, so that the further collected data is more valuable and the classification accuracy is higher. Therefore, the present invention proposes a method for reasonably selecting a perception region, which can select a region that is more valuable to the classification task, collect data based on the selected perception region, improve the representativeness (value) of the collected data, and thus improve the classification accuracy of the classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present application 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 application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 A schematic diagram of a flow chart of a method for selecting a perception area in sparse crowd sensing provided in Example 1 of the present application;

[0026] Figure 2 A flowchart of a data classification method provided for Examples 2 to 4 of the present application;

[0027] Figure 3 A schematic diagram of the technical concept of the data classification method provided in Examples 2 to 4 of the present application;

[0028] Figure 4 A schematic diagram of an active acquisition framework in sparse crowd sensing focusing on downstream classification tasks provided in Examples 2 to 4 of the present application. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0030] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0031] Example 1

[0032] The existing methods for selecting perception regions in sparse crowd sensing have certain shortcomings. (1) Different sub-regions have different effects on data inference. At present, when a method is found to measure the value of a sub-region, the sub-region with the highest value will be simply selected. However, in reality, a group of sub-regions needs to be selected, and there may be information overlap between different perception sub-regions, which will lead to a waste of perception cost. Simply selecting the k sub-regions with the highest value cannot maximize the perception information under limited cost. (2) The sub-region group selection problem is a subset selection problem, which is usually an NP-hard problem. Some existing studies have focused on the sub-region group selection problem and used heuristic algorithms, convex optimization and other methods to approximate the NP-hard problem of sub-region group selection, but these works have not given the performance guarantee of the proposed approximate solution. (3) The spatiotemporal data obtained by the crowd sensing system will be used in downstream tasks, such as training machine models. After clarifying the downstream tasks, using the downstream tasks to guide the selection of sub-regions will make the selected sub-regions more valuable to the downstream tasks, thereby obtaining more useful information at a lower cost. When selecting sub-regions, existing research considers the impact of sub-regions on data coverage, but does not consider the impact on downstream tasks. (4) The data samples screened by general active learning methods are often independent, so only the numerical value of the data needs to be paid attention to. However, the data that needs to be screened in sparse crowd sensing has spatiotemporal correlation. In addition to paying attention to the data numerical value of the sub-region, the spatiotemporal characteristics of the sub-region itself need to be considered. Current research has not proposed a data preprocessing method that retains spatial information when using active learning algorithms to select sub-regions in sparse crowd sensing. Based on the above analysis, this embodiment provides a perception area selection method in sparse crowd sensing, such as Figure 1 As shown, including:

[0033] Step 101, obtaining sparse spatiotemporal data at time t in a target area; the target area includes several sub-areas. When the sparse crowd-sensing data acquisition method is applied to obtain air quality training data for an air quality classification model, the sparse spatiotemporal data is air quality sparse spatiotemporal data; when the sparse crowd-sensing data acquisition method is applied to obtain traffic flow training data for a traffic flow classification model, the sparse spatiotemporal data is traffic flow sparse spatiotemporal data; when the sparse crowd-sensing data acquisition method is applied to obtain wind speed training data for a wind speed classification model, the sparse spatiotemporal data is wind speed sparse spatiotemporal data.

[0034] As an example, the target area is divided into several sub-areas by grid division in the target area, and one grid corresponds to one sub-area.

[0035] Step 102, preprocessing the sparse spatiotemporal data to obtain a feature matrix corresponding to each data missing sub-region; the feature matrix includes inferred data at a position corresponding to a data missing sub-region and the sparse spatiotemporal data; the data missing sub-region refers to a sub-region where no data is obtained.

[0036] The sparse spatiotemporal data obtained does not correspond to data in every sub-region. Some sub-regions have not collected relevant data, so it is necessary to infer the data of the sub-regions with missing data and further derive the feature matrix of the sub-regions with missing data.

[0037] Step 103, according to the feature matrix corresponding to each data missing sub-region, a greedy approximation algorithm is applied to select each sub-region to be collected from all data missing sub-regions. The sub-region to be collected is the sensing region, or is called the sensing sub-region.

[0038] As an optional implementation, in step 102, the sparse spatiotemporal data is preprocessed to obtain a feature matrix corresponding to each data missing sub-region, specifically including:

[0039] (a1) Determine a data missing sub-region based on the sparse spatiotemporal data.

[0040] When there is no missing data in any sub-region at time t, X t The data x of m sub-regions p,t composition, use To represent the set of m sub-regions. However, in sparse crowd sensing, X t is sparse, that is, there is missing data, that is, |X t | <m。

[0041] When there are missing data in a sub-region at time t, the indicator It is used to indicate whether the data in each sub-region is missing. The index i of m sub-regions p,t When i p,t = 0, indicating the data x at position p in sub-region at time t p,t missing; when i p,t =1, indicating the data x at position p in sub-region at time t p,t Not missing. The set of sub-regions with missing data at time t

[0042] (a2) According to the data missing sub-region, the sparse spatiotemporal data is represented in a matrix form to obtain a sparse spatiotemporal data matrix; in the sparse spatiotemporal data matrix, the elements at the positions corresponding to the data missing sub-regions are 0, and the elements at the positions corresponding to the data not missing sub-regions are the actually collected data.

[0043] The matrix can store the spatial information of the sub-region of the spatiotemporal data. Use a matrix to represent the sparse spatiotemporal data Xt. Each sub-region p corresponds to an element in the matrix. The data at point p is the data value of the corresponding element in the matrix. The conversion function is:

[0044] x matrix =Matrix(X t ,I t )

[0045] when i p,t =1, it means that the data at position p is not missing, and the element corresponding to position p is x p,t , otherwise the value is 0.

[0046] (a3) inferring the data of the data-missing sub-region based on the actually collected data in the sparse spatiotemporal data matrix to obtain inferred data corresponding to the data-missing sub-region.

[0047] Obtain all missing data points through linear interpolation and spatial smoothing The corresponding position data is value p .

[0048] Linear interpolation (time dimension): Using the known data from the previous time, linear interpolation is used to predict the missing value at the current time. This method smoothly continues the data trend in time.

[0049] Spatial smoothing (spatial dimension): Using the known location data around the missing value, the missing value is estimated by weighted average. The weight is usually set based on distance, and the closer the point is, the greater the influence on the estimated value.

[0050] Combining the two methods: first use temporal linear interpolation to get a preliminary estimate, and then use spatial smoothing to further optimize, improves the accuracy of data extrapolation.

[0051] (a4) constructing an auxiliary matrix for each data-missing sub-region; the auxiliary matrix includes inferred data at a position corresponding to a data-missing sub-region; in the auxiliary matrix, the elements at the position corresponding to the data-missing sub-region are inferred data, and the elements at other positions are 0.

[0052] For each missing data point Construct an auxiliary matrix matrix_aid p , in this auxiliary matrix, the position value corresponding to the missing point p is value p , and the values ​​of the other positions are all 0.

[0053] (a5) obtaining a feature matrix corresponding to each data-missing sub-region according to the auxiliary matrix and the sparse spatiotemporal data matrix.

[0054] The matrix x corresponding to the initial sparse spatiotemporal data at time t matrix The auxiliary matrix matrix_aid for the missing data subregion p p Add together to get the feature matrix x' of sub-region p p,t ;

[0055] x' p,t =Matrix(X t ,I t )+matrix_aid p

[0056] The feature matrices of all data-missing sub-regions p constitute the data set D of the sub-region to be selected. pool ;

[0057]

[0058] For data preprocessing, a two-dimensional matrix is ​​used to store sparse perception spatiotemporal data. A correspondence is established between sub-regions and matrix elements, and a feature matrix is ​​generated for each sub-region with missing data. The feature matrix saves the spatial information of the sub-region with missing data and the information of other known sub-regions. This information can assist in calculating the value of the sub-region and better select the required perception sub-region.

[0059] In step 103, the greedy approximation algorithm process for selecting the sub-region of the t-th time slot is:

[0060] (1) Let b represent the budget in the tth time slot, D pool is the set of feature matrices corresponding to the sub-regions to be selected. Build a Bayesian convolutional neural network model fBCNN Complete the downstream classification task. The convolutional neural network is responsible for extracting the spatial information of spatio-temporal data. The parameters ω of the Bayesian neural network satisfy a distribution p(ω), rather than taking a specific value. The parameter distribution p(ω) can indicate the confidence level of the classification model. The more concentrated the parameter distribution is, the more confident the classification model is. The model parameters ω of the downstream classification model (target classification model) satisfy a distribution, ω ∼ p(ω). The set A of the feature matrices composed of the sub-regions to be collected is initially set to an empty set, that is The set A and D of sub-regions to be collected pool Similarly, its elements are the feature matrices corresponding one-to-one to the data missing sub-regions rather than the data missing sub-regions themselves. Because the feature matrices and the data missing sub-regions are in one-to-one correspondence, the following steps will select the feature matrices of the data missing sub-regions. Next, the feature matrix of the nth sub-region to be collected will be selected in turn, and the initial value of n is 1.

[0061] (2) Traverse the data set D of the sub-regions to be selected pool for the feature matrix x of the unselected sub-regions, x ∈ D pool \A n-1 , and calculate the score s of the set A n-1 after adding x to it under the active acquisition strategy G that takes into account both parameter uncertainty and model performance n-1 . x .

[0062] (3) Select the data x that maximizes the score s of the set A n-1 ∪{x} as the feature matrix x of the sub-region to be collected selected in this round x : n

[0063] (4) Update the set state, add x n to the candidate set A n-1 to get A n , A n = A n-1 ∪{x n}, update n, n = n + 1.

[0064] (5) Compare the number n - 1 of the feature matrices of the selected sub-regions with the budget b at time slot t. If n - 1 < b, that is, there is still budget, then continue to select the feature matrix of the next sub-region and repeat the above steps. Otherwise, find the actual sub-region set to be collected through the selected set A b : P selected = position(A b ).

[0065] ​Based on the above analysis process, it is concluded that in step 103, according to the feature matrix corresponding to each data missing sub-region, the greedy approximation algorithm is applied to select each sub-region to be collected from all data missing sub-regions, specifically including:

[0066] (b1) Each currently unselected data-missing sub-region is combined with all selected data-missing sub-regions in the current set of sub-regions to be collected; initially, the set of sub-regions to be collected is an empty set.

[0067] (b2) Calculating the score of each combination using the target classification model based on the feature matrix corresponding to each combination; the score is a score that takes into account both the parameter uncertainty score and the model performance of the target classification model.

[0068] (b3) Selecting the unselected data-missing sub-regions in the combination with the highest score and adding them to the current set of sub-regions to be collected to update the current set of sub-regions to be collected.

[0069] (b4) Determine whether the number of selected data-missing sub-regions in the current set of sub-regions to be collected has reached a preset number.

[0070] (b5) If yes, then the selected data missing sub-regions in the current set of sub-regions to be collected are the sub-regions to be collected.

[0071] (b6) If not, return to step “combine each currently unselected data-missing sub-region with all selected data-missing sub-regions in the current set of sub-regions to be collected”.

[0072] As an optional implementation, in step (b2), the score of each combination is calculated, s x =G(A n-1 ∪{x}), apply the active acquisition strategy G(D) to calculate A n-1 The score of ∪{x} measures the importance of the feature matrix set D to the downstream target classification model. The score is obtained by the weighted sum of two parts:

[0073] s x =G(D)=αG ω (D)+(1-α)G M (D)

[0074] Among them, α is a hyperparameter and its value is 0.8. ω (D) represents the parameter uncertainty score, which quantifies the impact of the current set of spatiotemporal data D on the convergence speed of the model parameters ω:

[0075]

[0076] Where D = {(xi ,y i )|i=1,2,3…b},x i Represents the feature matrix x in the feature matrix set D i ;y i Represents the feature matrix x in the feature matrix set D i The corresponding classification label; b represents the number of feature matrices in the feature matrix set D; D train refers to the original training set that does not contain D; H represents entropy; x 1:b represents x1,...,x i ,…,x b ;y 1:b represents y1,...,y i ,…,y b ; The entropy of the model parameters represents the expectation of the model prediction results, where the model prediction results follow the distribution p(y 1:b |x 1:b D train ).

[0077] G ω (D) In ​​the expression, the first term H(ω|D train ) represents the entropy of the parameters before training the model using the information of the feature matrix set D. The second term G represents the entropy of the parameters after the model is trained using the information of the feature matrix set D. The entropy of the parameters characterizes the degree of convergence of the parameters. The difference in entropy measures how much the parameters converge before and after the model is trained using the feature matrix set D. ω (D) represents the contribution of spatiotemporal data D to the convergence of model parameters ω.

[0078] Because the target classification model often has many parameters, in order to facilitate calculation, the following equivalent formula is used to calculate G ω (D):

[0079]

[0080] Among them, H(ω|y 1:b ,x 1:b ,D train ) and H(y 1:b |x 1:b ,ω,D train ) has a different meaning. 1:b ,x 1:b ,D train ) is the entropy of the parameter ω after training the model using the information of the feature matrix set D, H(y 1:b |x 1:b ,ω,D train) is the entropy of the model prediction result after training the model using the information of the feature matrix set D. Similarly, H(ω|D train ) and H(y 1:b |x 1:b ,D train ) has a different meaning. train ) is the entropy of the parameter ω before using the information of the feature matrix set D to train the model, H(y 1:b |x 1:b ,D train ) is the entropy of the model prediction results before using the information of the feature matrix set D to train the model. Because G ω (D) The final calculation results (calculation difference) of the first calculation method and the second calculation method are exactly the same. They are both calculating y 1:b The second formula is convenient for calculation, so the second formula is used to calculate the equivalent results. It can be simply understood that the entropy of the parameters is required to measure the convergence speed of the model. In order to facilitate calculation, the equivalent conversion of the calculation results is made, which is converted into the entropy of the model prediction results. The numerical value of the calculation results is equivalent.

[0081] G M (D) represents the probability marginal score, which quantifies the difficulty of the target classification model to infer the sample classification label based on the data of the current set of feature matrix set D, and represents the contribution of the data of feature matrix set D to the current performance improvement of the target classification model. When it is difficult for the target classification model to infer the label of the data based on the information of feature matrix set D, it means that the target classification model needs to learn the data of feature matrix set D to improve the performance of the model when processing related data. This value can be determined by calculating the margin between the probability values ​​estimated by the target classification model:

[0082]

[0083] Among them, f BCNN (x i ) max =p(y max |ω);f BCNN (x i ) 2-max =p(y 2-max |ω).

[0084] Among them, f BCNN is the target classification model, which is a Bayesian convolutional neural network model. max and 2-max is the target classification model for the data x i The labels with the highest and second highest predicted probability values.

[0085] Therefore, in step (b2), according to the feature matrix corresponding to each combination, the target classification model is applied to calculate the score of each combination, specifically including:

[0086] (b21) For each combination, the entropy of the model parameters of the target classification model before being trained by the feature matrix corresponding to the combination is calculated to obtain a first model parameter entropy; the feature matrix corresponding to each combination has a corresponding preset classification label. In other words, for each combination, the entropy of the model prediction result of the target classification model before being trained by the feature matrix corresponding to the combination is calculated to obtain the entropy of the first model prediction result.

[0087] (b22) For each combination, the entropy of the model parameters of the target classification model after not being trained by the feature matrix corresponding to the combination is calculated to obtain the second model parameter entropy. In other words, for each combination, the entropy of the model prediction result of the target classification model after not being trained by the feature matrix corresponding to the combination is calculated to obtain the entropy of the second model prediction result.

[0088] (b23) calculating a model parameter uncertainty score based on the first model parameter entropy and the second model parameter entropy. In other words, calculating a model parameter uncertainty score based on the entropy of the first model prediction result and the entropy of the second model prediction result.

[0089] (b24) For each combination, the feature matrix corresponding to the combination is input into the target classification model to obtain the highest prediction probability result and the second highest prediction probability result for each data missing sub-region in the combination.

[0090] (b25) calculating the probability marginal score corresponding to each combination according to the highest predicted probability result and the second highest predicted probability result of each data missing sub-region in the combination.

[0091] (b26) Calculating the score of each combination based on the model parameter uncertainty score and the probability marginal score corresponding to each combination.

[0092] In this embodiment, on the one hand, an active acquisition function G that takes into account parameter uncertainty and model performance is used to measure the value of a group of data missing sub-regions. When measuring the value of a data missing sub-region, the changes in model parameter uncertainty and model performance before and after collecting the data missing sub-region are considered, and the data missing sub-region group that can best improve the convergence speed and model performance of the model is selected for perception. On the other hand, the greedy approximation of the selection of data missing sub-regions requires an evaluation of each combination of data missing sub-regions to select the most valuable group of data missing sub-regions. This is an NP-hard problem that needs to be solved using an approximate method. Because the active acquisition strategy G is a sub-module function, the selection of the data missing sub-region group is approximated using a greedy point selection method. The greedy approximation of the sub-module function can achieve a 1-1 / e approximation, which approximately solves the NP-hard problem while ensuring accuracy.

[0093] The perception region selection method of this embodiment can solve the shortcomings of the existing perception region selection method:

[0094] (1) Multi-point joint evaluation: Using an active collection strategy that takes into account parameter uncertainty and model performance to measure the value of a group of data-missing sub-regions can effectively reduce the problem of information overlap within the combination caused by selecting the top k data-missing sub-regions with the highest single-point evaluation scores as the group of sub-regions to be collected. In this application, the value of a group of data-missing sub-regions is jointly evaluated, thereby reducing the information overlap within the combination of data-missing sub-regions.

[0095] (2) Performance guarantee of approximate implementation of NP-hard problems: Because the active acquisition function that takes into account both parameter uncertainty and model performance is a submodular function, there is a theoretical guarantee for approximating the NP-hard problem of combining sub-regions with missing data. The greedy approximation of the submodular function is implemented by the 1-1 / e approximation.

[0096] (3) Consider downstream tasks when selecting data-missing sub-regions: When measuring the value of data-missing sub-regions, the active collection strategy that takes into account parameter uncertainty and model performance considers the impact of data-missing sub-regions on the convergence speed and performance improvement of downstream classification models. Therefore, the selected data-missing sub-regions are the most valuable for downstream tasks, which improves the targeted selection of data-missing sub-regions and can obtain more valuable data under a limited budget.

[0097] (4) Data preprocessing when applying active learning methods in crowd sensing: By processing sparse data in sparse crowd sensing, a set of data missing sub-regions and their corresponding feature matrices are obtained, and the feature matrices are used to construct the data set to be selected. While constructing the data set to be selected for the active learning method, the spatial information of the data missing sub-regions and the data of the remaining known sub-regions are retained. This provides an efficient and effective data preprocessing method for applying active learning methods in the sparse crowd sensing framework.

[0098] The present application also provides an application scenario, which applies the above-mentioned perception area selection method in sparse group intelligence perception. Specifically: the perception area selection method in sparse group intelligence perception provided in this embodiment can be applied in the air quality data classification scenario. The scenario includes a data acquisition link, an air quality classification model training link and an air quality classification link. The data acquisition link is used to obtain air quality data in the target area. The perception area selection method in sparse group intelligence perception provided in this embodiment belongs to the data acquisition link. Specifically, in the process of acquiring air quality data based on sparse group intelligence perception, the perception area is reasonably selected to obtain more valuable training data that is more conducive to air quality classification.

[0099] The present application also provides an application scenario, which applies the above-mentioned perception area selection method in sparse group intelligence perception. Specifically: the perception area selection method in sparse group intelligence perception provided in this embodiment can be applied in the traffic flow data classification scenario. The scenario includes a data acquisition link, a traffic flow classification model training link and a traffic flow classification link. The data acquisition link is used to obtain traffic flow data in the target area. The perception area selection method in sparse group intelligence perception provided in this embodiment belongs to the data acquisition link. Specifically, in the process of acquiring traffic flow data based on sparse group intelligence perception, the perception area is reasonably selected to obtain more valuable training data that is more conducive to traffic flow classification.

[0100] The present application also provides an application scenario, which applies the above-mentioned perception area selection method in sparse group intelligence perception. Specifically: the perception area selection method in sparse group intelligence perception provided in this embodiment can be applied in the wind speed data classification scenario. The scenario includes a data acquisition link, a wind speed classification model training link and a wind speed classification link. The data acquisition link is used to obtain wind speed data in the target area. The perception area selection method in sparse group intelligence perception provided in this embodiment belongs to the data acquisition link. Specifically, in the process of acquiring wind speed data based on sparse group intelligence perception, the perception area is reasonably selected to obtain more valuable training data that is more conducive to wind speed classification.

[0101] Example 2

[0102] like Figures 2 to 4 As shown, this embodiment provides a data classification method, which is specifically an air quality data classification method. The air quality data classification method includes:

[0103] Step 201, obtain the historical spatiotemporal data and corresponding classification labels in the target area, and construct an initial training data set. The historical spatiotemporal data refers to the historical spatiotemporal data of air quality before time t; the classification labels include excellent, good and poor air quality. The classification labels are pre-labeled by human experts based on the collected air quality data.

[0104] Step 202: Train the target classification model using the current training data set.

[0105] Step 203, determining whether the performance of the currently trained target classification model meets the preset performance requirements.

[0106] Step 204: If yes, a trained target classification model is obtained to classify the data. In this embodiment, the target classification model is an air quality classification model, and specifically a Bayesian convolutional neural network model is used.

[0107] Step 205: If not, then based on the sparse spatiotemporal data at time t in the target area, the perception area selection method in sparse crowd intelligence perception described in Example 1 is executed to obtain each sub-area to be collected.

[0108] Step 206: According to the data acquisition mode of sparse crowd intelligence perception, the spatiotemporal data of each sub-area to be collected is acquired, specifically, the air quality data of the sub-area to be collected.

[0109] The crowd intelligence perception platform takes the task location P selected , time requirements, and types of perceived data are published to the perception employee. The perception employee selects the task, performs the task, and uploads the data. The crowd intelligence perception platform pushes the task to the perception employee or the employee can browse the optional tasks and choose to participate. The perception employee confirms acceptance of the task and obtains detailed requirements for performing the task (location of the data to be perceived, time requirements, and types of perceived data). The perception employee uses the perception device to collect data at the time and place specified in the task. The perception employees upload the collected data to the crowd-intelligence perception platform.

[0110] Step 207, adding the spatiotemporal data of the sub-area to be collected to the sparse spatiotemporal data at time t, to obtain training data at time t, specifically air quality training data; the classification labels of the training data at time t are defined by manual annotation.

[0111] Update the data at time t X t Submit it to human experts for labeling to the corresponding label Yt .

[0112] Step 208: Add the training data at time t to the initial training data set to obtain an updated training data set. train =D train ∪{(X t ,Y t )}.

[0113] Step 209, using the current updated training data set as the current training data set, and returning to the step of "training the target classification model using the current training data set".

[0114] Among them, in step 205, based on the sparse spatiotemporal data at time t in the target area, the perception area selection method in sparse crowd intelligence perception described in Example 1 is executed to obtain each sub-area to be collected. Specifically:

[0115] Step 101, obtaining sparse spatiotemporal data at time t in a target area, specifically air quality data at time t; the target area includes several sub-areas.

[0116] Step 102, preprocessing the sparse spatiotemporal data to obtain a feature matrix corresponding to each data-missing sub-region; the feature matrix includes air quality inference data at a position corresponding to a data-missing sub-region and the sparse spatiotemporal data; the data-missing sub-region refers to a sub-region where no air quality data is obtained.

[0117] Step 103 : According to the feature matrix corresponding to each data missing sub-region, a greedy approximation algorithm is applied to select sub-regions to be collected from all data missing sub-regions.

[0118] As an optional implementation, in step 102, the sparse spatiotemporal data is preprocessed to obtain a feature matrix corresponding to each data missing sub-region, specifically including:

[0119] (a1) Determine a data missing sub-region based on the sparse spatiotemporal data.

[0120] (a2) According to the data missing sub-region, the sparse spatiotemporal data is represented in a matrix form to obtain a sparse spatiotemporal data matrix; in the sparse spatiotemporal data matrix, the elements at the positions corresponding to the data missing sub-regions are 0, and the elements at the positions corresponding to the data not missing sub-regions are the actually collected air quality data.

[0121] (a3) inferring the data of the data-missing sub-region based on the actually collected data in the sparse spatiotemporal data matrix to obtain the air quality inference data corresponding to the data-missing sub-region.

[0122] (a4) constructing an auxiliary matrix for each data missing sub-region; the auxiliary matrix includes air quality inference data at a position corresponding to a data missing sub-region; in the auxiliary matrix, the elements at the position corresponding to the data missing sub-region are air quality inference data, and the elements at other positions are 0.

[0123] (a5) obtaining a feature matrix corresponding to each data missing sub-region according to the auxiliary matrix and the sparse spatiotemporal data matrix.

[0124] As an optional implementation, in step 103, according to the feature matrix corresponding to each data missing sub-region, a greedy approximation algorithm is applied to select each sub-region to be collected from all data missing sub-regions, specifically including:

[0125] (b1) Each currently unselected data-missing sub-region is combined with all selected data-missing sub-regions in the current set of sub-regions to be collected; initially, the set of sub-regions to be collected is an empty set; initially, the unselected data-missing sub-regions are all data-missing sub-regions. Any unselected data-missing sub-region is combined with all selected data-missing sub-regions in the current set of sub-regions to be collected to obtain a combination, and each unselected data-missing sub-region corresponds to a combination after combination.

[0126] (b2) Based on the feature matrix corresponding to each combination, the target classification model (air quality classification model) is applied to calculate the score of each combination; the score is a score that takes into account both the parameter uncertainty score and the model performance of the target classification model.

[0127] (b3) Selecting the unselected data-missing sub-regions in the combination with the highest score and adding them to the current set of sub-regions to be collected to update the current set of sub-regions to be collected.

[0128] (b4) Determine whether the number of selected data-missing sub-regions in the current set of sub-regions to be collected has reached a preset number.

[0129] (b5) If yes, then the selected data missing sub-regions in the current set of sub-regions to be collected are the sub-regions to be collected.

[0130] (b6) If not, return to step “combine each currently unselected data-missing sub-region with all selected data-missing sub-regions in the current set of sub-regions to be collected”.

[0131] In step (b2), according to the feature matrix corresponding to each combination, the target classification model (air quality classification model) is applied to calculate the score of each combination, which specifically includes:

[0132] (b21) For each combination, the entropy of the model parameters of the target classification model before being trained by the feature matrix corresponding to the combination is calculated to obtain a first model parameter entropy; the feature matrix corresponding to each combination has a corresponding preset classification label.

[0133] (b22) For each combination, the entropy of the model parameters of the target classification model before being trained by the feature matrix corresponding to the combination is calculated to obtain a second model parameter entropy.

[0134] (b23) calculating a model parameter uncertainty score based on the first model parameter entropy and the second model parameter entropy.

[0135] (b24) For each combination, the feature matrix corresponding to the combination is input into the target classification model to obtain the highest prediction probability result and the second highest prediction probability result for each data missing sub-region in the combination.

[0136] (b25) calculating the probability marginal score corresponding to each combination according to the highest predicted probability result and the second highest predicted probability result of each data missing sub-region in the combination.

[0137] (b26) Calculating the score of each combination based on the model parameter uncertainty score and the probability marginal score corresponding to each combination.

[0138] In this embodiment, the air quality data in the target area is obtained by using the data acquisition method of sparse group intelligence perception to construct training data and train the wind speed classification model. Among them, when the air quality data in the target area is obtained by using the data acquisition method of sparse group intelligence perception, for the selection of the perception area, the present invention adopts a greedy approximation algorithm to select the perception area from all data missing sub-areas, which can effectively reduce the information overlap problem within the combination caused by selecting the top k polymer areas with single point evaluation scores as the sub-areas to be collected in the existing perception area selection method; in addition, the active collection strategy that takes into account the parameter uncertainty of the classification model and the model performance is used to measure the value (score) of each sub-area to be collected, which solves the NP-hard problem of the existing perception area selection, and when taking into account the parameter uncertainty of the classification model and the model performance, the influence of the sub-area on the convergence speed and performance improvement of the target classification model is considered. The selected sub-area is the most valuable for the classification task, which improves the pertinence of the selection of the data missing sub-area, and more valuable data can be obtained under a limited budget. Therefore, by applying the perception area selection method in the sparse group intelligence perception in this embodiment, it is possible to obtain areas that are more conducive to air quality classification, and to obtain more valuable air quality data, thereby improving the training effect of the air quality classification model, and further improving the classification accuracy during air quality classification.

[0139] Example 3

[0140] like Figures 2 to 4 As shown, this embodiment provides a data classification method, which is specifically a traffic flow data classification method. The traffic flow data classification method includes:

[0141] Step 201, obtain the historical spatiotemporal data and corresponding classification labels in the target area, and construct an initial training data set; wherein the historical spatiotemporal data refers to the historical spatiotemporal data of traffic flow before time t; the classification labels include peak traffic period, low traffic period, etc. The classification labels are pre-labeled by human experts based on the collected traffic flow data.

[0142] Step 202: Train the target classification model using the current training data set.

[0143] Step 203, determining whether the performance of the currently trained target classification model meets the preset performance requirements.

[0144] Step 204: If yes, a trained target classification model is obtained to perform data classification. In this embodiment, the target classification model is a traffic flow classification model, and specifically a Bayesian convolutional neural network model is used.

[0145] Step 205: If not, then based on the sparse spatiotemporal data at time t in the target area, the perception area selection method in sparse crowd intelligence perception described in Example 1 is executed to obtain each sub-area to be collected.

[0146] Step 206, according to the data acquisition mode of sparse crowd intelligence perception, the spatiotemporal data of each sub-area to be collected is acquired, specifically, the traffic flow data of the sub-area to be collected.

[0147] The crowd intelligence perception platform takes the task location P selected , time requirements, and types of perceived data are published to the perception employee. The perception employee selects the task, performs the task, and uploads the data. The crowd intelligence perception platform pushes the task to the perception employee or the employee can browse the optional tasks and choose to participate. The perception employee confirms acceptance of the task and obtains detailed requirements for performing the task (location of the data to be perceived, time requirements, and types of perceived data). The perception employee uses the perception device to collect data at the time and place specified in the task. The perception employees upload the collected data to the crowd-intelligence perception platform.

[0148] Step 207, adding the spatiotemporal data of the sub-area to be collected to the sparse spatiotemporal data at time t to obtain the training data at time t; the classification labels of the training data at time t are defined by manual annotation.

[0149] Update the data at time t X tSubmit it to human experts for labeling to the corresponding label Y t .

[0150] Step 208: Add the training data at time t to the initial training data set to obtain an updated training data set. train =D train ∪{(X t ,Y t )}.

[0151] Step 209, using the current updated training data set as the current training data set, and returning to the step of "training the target classification model using the current training data set".

[0152] Among them, in step 205, based on the sparse spatiotemporal data at time t in the target area, the perception area selection method in sparse crowd intelligence perception described in Example 1 is executed to obtain each sub-area to be collected. Specifically:

[0153] Step 101, obtaining sparse spatiotemporal data at time t in a target area, specifically traffic flow data at time t; the target area includes several sub-areas.

[0154] Step 102, preprocessing the sparse spatiotemporal data to obtain a feature matrix corresponding to each data-missing sub-region; the feature matrix includes traffic flow inference data at a position corresponding to a data-missing sub-region and the sparse spatiotemporal data; the data-missing sub-region refers to a sub-region where traffic flow data is not obtained.

[0155] Step 103 : According to the feature matrix corresponding to each data missing sub-region, a greedy approximation algorithm is applied to select sub-regions to be collected from all data missing sub-regions.

[0156] As an optional implementation, in step 102, the sparse spatiotemporal data is preprocessed to obtain a feature matrix corresponding to each data missing sub-region, specifically including:

[0157] (a1) Determine a data missing sub-region based on the sparse spatiotemporal data.

[0158] (a2) According to the data missing sub-region, the sparse spatiotemporal data is represented in a matrix form to obtain a sparse spatiotemporal data matrix; in the sparse spatiotemporal data matrix, the elements at the positions corresponding to the data missing sub-regions are 0, and the elements at the positions corresponding to the data not missing sub-regions are the actually collected traffic flow data.

[0159] (a3) inferring the data of the data-missing sub-region based on the actually collected data in the sparse spatiotemporal data matrix to obtain inferred traffic flow data corresponding to the data-missing sub-region.

[0160] (a4) constructing an auxiliary matrix for each data missing sub-region; the auxiliary matrix includes the traffic flow inference data at a position corresponding to the data missing sub-region; in the auxiliary matrix, the element at the position corresponding to the data missing sub-region is the traffic flow inference data, and the elements at other positions are 0;

[0161] (a5) obtaining a feature matrix corresponding to each data missing sub-region according to the auxiliary matrix and the sparse spatiotemporal data matrix.

[0162] As an optional implementation, in step 103, according to the feature matrix corresponding to each data missing sub-region, a greedy approximation algorithm is applied to select each sub-region to be collected from all data missing sub-regions, specifically including:

[0163] (b1) Each currently unselected data-missing sub-region is combined with all selected data-missing sub-regions in the current set of sub-regions to be collected; initially, the set of sub-regions to be collected is an empty set; initially, the unselected data-missing sub-regions are all data-missing sub-regions. Any unselected data-missing sub-region is combined with all selected data-missing sub-regions in the current set of sub-regions to be collected to obtain a combination, and each unselected data-missing sub-region corresponds to a combination after combination.

[0164] (b2) Based on the feature matrix corresponding to each combination, the target classification model (traffic flow classification model) is applied to calculate the score of each combination; the score is a score that takes into account both the parameter uncertainty score and the model performance of the target classification model.

[0165] (b3) Selecting the unselected data-missing sub-regions in the combination with the highest score and adding them to the current set of sub-regions to be collected to update the current set of sub-regions to be collected.

[0166] (b4) Determine whether the number of selected data-missing sub-regions in the current set of sub-regions to be collected has reached a preset number.

[0167] (b5) If yes, then the selected data missing sub-regions in the current set of sub-regions to be collected are the sub-regions to be collected.

[0168] (b6) If not, return to step “combine each currently unselected data-missing sub-region with all selected data-missing sub-regions in the current set of sub-regions to be collected”.

[0169] In step (b2), according to the feature matrix corresponding to each combination, the target classification model (traffic flow classification model) is applied to calculate the score of each combination, which specifically includes:

[0170] (b21) For each combination, the entropy of the model parameters of the target classification model before being trained by the feature matrix corresponding to the combination is calculated to obtain a first model parameter entropy; the feature matrix corresponding to each combination has a corresponding preset classification label, such as a peak period or a low peak period.

[0171] (b22) For each combination, the entropy of the model parameters of the target classification model before being trained by the feature matrix corresponding to the combination is calculated to obtain a second model parameter entropy.

[0172] (b23) calculating a model parameter uncertainty score based on the first model parameter entropy and the second model parameter entropy.

[0173] (b24) For each combination, the feature matrix corresponding to the combination is input into the target classification model to obtain the highest prediction probability result and the second highest prediction probability result for each data missing sub-region in the combination.

[0174] (b25) calculating the probability marginal score corresponding to each combination according to the highest predicted probability result and the second highest predicted probability result of each data missing sub-region in the combination.

[0175] (b26) Calculating the score of each combination based on the model parameter uncertainty score and the probability marginal score corresponding to each combination.

[0176] In this embodiment, the traffic flow data in the target area is acquired by using the data acquisition method of sparse group intelligence perception to construct training data and train the traffic flow classification model. Among them, when the traffic flow data in the target area is acquired by using the data acquisition method of sparse group intelligence perception, for the selection of the perception area, the present invention adopts a greedy approximation algorithm to select the perception area from all data missing sub-areas, which can effectively reduce the information overlap problem within the combination caused by selecting the top k polymer areas with single point evaluation scores as the sub-areas to be collected in the existing perception area selection method; in addition, the value (score) of each sub-area to be collected is measured by using an active collection strategy that takes into account the parameter uncertainty of the classification model and the model performance, which solves the NP-hard problem of the existing perception area selection, and when taking into account the parameter uncertainty of the classification model and the model performance, the influence of the sub-area on the convergence speed and performance improvement of the target classification model is considered, and the selected sub-area is the most valuable for the classification task, which improves the pertinence of the selection of the sub-area with missing data, and more valuable data can be obtained under a limited budget. Therefore, by applying the perception area selection method in the sparse group intelligence perception in this embodiment, it is possible to obtain areas that are more conducive to traffic flow classification and more valuable traffic flow data, thereby improving the training effect of the traffic flow classification model and further improving the classification accuracy when classifying traffic flow size.

[0177] Example 4

[0178] like Figures 2 to 4 As shown, this embodiment provides a data classification method, which is specifically a wind speed data classification method. The wind speed data classification method includes:

[0179] Step 201, obtain the historical spatiotemporal data and corresponding classification labels in the target area, and construct an initial training data set; wherein the historical spatiotemporal data refers to the historical spatiotemporal data of wind speed before time t; the classification labels include high wind speed, medium wind speed, low wind speed, etc. The classification labels are pre-labeled by human experts based on the collected wind speed data.

[0180] Step 202: Train the target classification model using the current training data set.

[0181] Step 203, determining whether the performance of the currently trained target classification model meets the preset performance requirements.

[0182] Step 204: If yes, a trained target classification model is obtained to perform data classification. In this embodiment, the target classification model is a wind speed classification model, and specifically a Bayesian convolutional neural network model is used.

[0183] Step 205: If not, then based on the sparse spatiotemporal data at time t in the target area, the perception area selection method in sparse crowd intelligence perception described in Example 1 is executed to obtain each sub-area to be collected.

[0184] Step 206: According to the data acquisition mode of sparse crowd intelligence perception, the spatiotemporal data of each sub-area to be collected is acquired, specifically, the wind speed data of the sub-area to be collected.

[0185] The crowd intelligence perception platform takes the task location P selected , time requirements, and types of perceived data are published to the perception employee. The perception employee selects the task, performs the task, and uploads the data. The crowd intelligence perception platform pushes the task to the perception employee or the employee can browse the optional tasks and choose to participate. The perception employee confirms acceptance of the task and obtains detailed requirements for performing the task (location of the data to be perceived, time requirements, and types of perceived data). The perception employee uses the perception device to collect data at the time and place specified in the task. The perception employees upload the collected data to the crowd-intelligence perception platform.

[0186] Step 207, adding the spatiotemporal data of the sub-area to be collected to the sparse spatiotemporal data at time t to obtain the training data at time t; the classification labels of the training data at time t are defined by manual annotation.

[0187] Update the data at time t Xt Submit it to human experts for labeling to the corresponding label Y t .

[0188] Step 208: Add the training data at time t to the initial training data set to obtain an updated training data set. train =D train ∪{(X t ,Y t )}.

[0189] Step 209, using the current updated training data set as the current training data set, and returning to the step of "training the target classification model using the current training data set".

[0190] Among them, in step 205, based on the sparse spatiotemporal data at time t in the target area, the perception area selection method in sparse crowd intelligence perception described in Example 1 is executed to obtain each sub-area to be collected. Specifically:

[0191] Step 101, obtaining sparse spatiotemporal data at time t in a target area, specifically wind speed data at time t; the target area includes several sub-areas.

[0192] Step 102, preprocessing the sparse spatiotemporal data to obtain a feature matrix corresponding to each data-missing sub-region; the feature matrix includes wind speed inference data at a position corresponding to a data-missing sub-region and the sparse spatiotemporal data; the data-missing sub-region refers to a sub-region where no wind speed data is obtained.

[0193] Step 103 : According to the feature matrix corresponding to each data missing sub-region, a greedy approximation algorithm is applied to select sub-regions to be collected from all data missing sub-regions.

[0194] As an optional implementation, in step 102, the sparse spatiotemporal data is preprocessed to obtain a feature matrix corresponding to each data missing sub-region, specifically including:

[0195] (a1) Determine a data missing sub-region based on the sparse spatiotemporal data.

[0196] (a2) According to the data missing sub-region, the sparse spatiotemporal data is represented in a matrix form to obtain a sparse spatiotemporal data matrix; in the sparse spatiotemporal data matrix, the elements at the positions corresponding to the data missing sub-regions are 0, and the elements at the positions corresponding to the data not missing sub-regions are the actually collected wind speed data.

[0197] (a3) inferring the data of the data-missing sub-region based on the actually collected data in the sparse spatiotemporal data matrix to obtain the wind speed inference data corresponding to the data-missing sub-region.

[0198] (a4) constructing an auxiliary matrix for each data missing sub-region; the auxiliary matrix includes the wind speed inference data at a position corresponding to the data missing sub-region; in the auxiliary matrix, the elements at the position corresponding to the data missing sub-region are the wind speed inference data, and the elements at other positions are 0.

[0199] (a5) obtaining a feature matrix corresponding to each data missing sub-region according to the auxiliary matrix and the sparse spatiotemporal data matrix.

[0200] As an optional implementation, in step 103, according to the feature matrix corresponding to each data missing sub-region, a greedy approximation algorithm is applied to select each sub-region to be collected from all data missing sub-regions, specifically including:

[0201] (b1) Each currently unselected data-missing sub-region is combined with all selected data-missing sub-regions in the current set of sub-regions to be collected; initially, the set of sub-regions to be collected is an empty set; initially, the unselected data-missing sub-regions are all data-missing sub-regions. Any unselected data-missing sub-region is combined with all selected data-missing sub-regions in the current set of sub-regions to be collected to obtain a combination, and each unselected data-missing sub-region corresponds to a combination after combination.

[0202] (b2) Based on the feature matrix corresponding to each combination, the target classification model (wind speed classification model) is applied to calculate the score of each combination; the score is a score that takes into account both the parameter uncertainty score and the model performance of the target classification model.

[0203] (b3) Selecting the unselected data-missing sub-regions in the combination with the highest score and adding them to the current set of sub-regions to be collected to update the current set of sub-regions to be collected.

[0204] (b4) Determine whether the number of selected data-missing sub-regions in the current set of sub-regions to be collected has reached a preset number.

[0205] (b5) If yes, then the selected data missing sub-regions in the current set of sub-regions to be collected are the sub-regions to be collected.

[0206] (b6) If not, return to step “combine each currently unselected data-missing sub-region with all selected data-missing sub-regions in the current set of sub-regions to be collected”.

[0207] Among them, in step (b2), according to the feature matrix corresponding to each combination, the target classification model (wind speed classification model) is applied to calculate the score of each combination, which specifically includes:

[0208] (b21) For each combination, the entropy of the model parameters of the target classification model before being trained by the feature matrix corresponding to the combination is calculated to obtain a first model parameter entropy; the feature matrix corresponding to each combination has a corresponding preset classification label.

[0209] (b22) For each combination, the entropy of the model parameters of the target classification model before being trained by the feature matrix corresponding to the combination is calculated to obtain a second model parameter entropy.

[0210] (b23) calculating a model parameter uncertainty score based on the first model parameter entropy and the second model parameter entropy.

[0211] (b24) For each combination, the feature matrix corresponding to the combination is input into the target classification model to obtain the highest prediction probability result and the second highest prediction probability result for each data missing sub-region in the combination.

[0212] (b25) calculating the probability marginal score corresponding to each combination according to the highest predicted probability result and the second highest predicted probability result of each data missing sub-region in the combination.

[0213] (b26) Calculating the score of each combination based on the model parameter uncertainty score and the probability marginal score corresponding to each combination.

[0214] In this embodiment, the wind speed data in the target area is obtained by using the data acquisition method of sparse group intelligence perception to construct training data and train the wind speed classification model. Among them, when the wind speed data in the target area is obtained by using the data acquisition method of sparse group intelligence perception, for the selection of the perception area, the present invention adopts a greedy approximation algorithm to select the perception area from all data missing sub-areas, which can effectively reduce the information overlap problem within the combination caused by selecting the top k polymer areas with single point evaluation scores as the sub-areas to be collected in the existing perception area selection method; in addition, the active collection strategy that takes into account the parameter uncertainty of the classification model and the model performance is used to measure the value (score) of each sub-area to be collected, which solves the NP-hard problem of the existing perception area selection, and when taking into account the parameter uncertainty of the classification model and the model performance, the influence of the sub-area on the convergence speed and performance improvement of the target classification model is considered. The selected sub-area is the most valuable for the classification task, which improves the pertinence of the selection of the data missing sub-area, and more valuable data can be obtained under a limited budget. Therefore, by applying the perception area selection method in the sparse group intelligence perception in this embodiment, it is possible to obtain an area that is more conducive to wind speed classification, and to obtain more valuable wind speed data, thereby improving the training effect of the wind speed classification model, and further improving the classification accuracy when classifying wind speeds.

[0215] Example 5

[0216] In this embodiment, a computer device is provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the above method embodiments.

[0217] Example 6

[0218] In this embodiment, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0219] Example 7

[0220] In this embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0221] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0222] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for selecting a perception area in sparse crowd sensing, characterized in that: The method for selecting a perception area in sparse crowd sensing includes: Acquire sparse spatiotemporal data at time t in a target area; the target area includes several sub-areas; Preprocessing the sparse spatiotemporal data to obtain a feature matrix corresponding to each data missing sub-region; the feature matrix includes inferred data at a position corresponding to a data missing sub-region and the sparse spatiotemporal data; the data missing sub-region refers to a sub-region where no data is obtained; According to the feature matrix corresponding to each data missing sub-region, the greedy approximation algorithm is applied to select the sub-regions to be collected from all data missing sub-regions.

2. The method for selecting a perception area in sparse crowd sensing according to claim 1, characterized in that: The sparse spatiotemporal data is preprocessed to obtain a feature matrix corresponding to each data missing sub-region, specifically including: Determine a data missing sub-region based on the sparse spatiotemporal data; According to the data missing sub-region, the sparse spatiotemporal data is expressed in a matrix form to obtain a sparse spatiotemporal data matrix; in the sparse spatiotemporal data matrix, the elements at the positions corresponding to the data missing sub-regions are 0, and the elements at the positions corresponding to the data not missing sub-regions are the actually collected data; Inferring the data of the data-missing sub-region according to the actually collected data in the sparse spatiotemporal data matrix to obtain inferred data corresponding to the data-missing sub-region; Constructing an auxiliary matrix for each data missing sub-region; the auxiliary matrix includes inferred data at a position corresponding to a data missing sub-region; in the auxiliary matrix, the element at the position corresponding to the data missing sub-region is the inferred data, and the elements at other positions are 0; A feature matrix corresponding to each data missing sub-region is obtained according to the auxiliary matrix and the sparse spatiotemporal data matrix.

3. The method for selecting a perception area in sparse crowd sensing according to claim 1, characterized in that: According to the feature matrix corresponding to each data missing sub-region, the greedy approximation algorithm is applied to select the sub-regions to be collected from all data missing sub-regions, including: Each currently unselected data-missing sub-region is combined with all selected data-missing sub-regions in the current set of sub-regions to be collected; initially, the set of sub-regions to be collected is an empty set. According to the feature matrix corresponding to each combination, the target classification model is applied to calculate the score of each combination; the score is a score that takes into account both the parameter uncertainty score and the model performance of the target classification model; Select the unselected data missing sub-regions in the combination with the highest score and add them to the current set of sub-regions to be collected to update the current set of sub-regions to be collected; Determine whether the number of selected data-missing sub-regions in the current set of sub-regions to be collected has reached a preset number; If yes, then the selected data missing sub-regions in the current set of sub-regions to be collected are the sub-regions to be collected; If not, return to the step of "combining each currently unselected data-missing sub-region with all selected data-missing sub-regions in the current set of sub-regions to be collected".

4. The method for selecting a perception area in sparse crowd sensing according to claim 3, characterized in that: According to the feature matrix corresponding to each combination, the target classification model is applied to calculate the score of each combination, including: For each combination, the entropy of the model parameters of the target classification model before being trained by the feature matrix corresponding to the combination is calculated to obtain a first model parameter entropy; the feature matrix corresponding to each combination has a corresponding preset classification label; For each combination, calculating the entropy of the model parameters of the target classification model after not being trained by the feature matrix corresponding to the combination, to obtain a second model parameter entropy; Calculating a model parameter uncertainty score based on the first model parameter entropy and the second model parameter entropy; For each combination, inputting the feature matrix corresponding to the combination into the target classification model to obtain the highest prediction probability result and the second highest prediction probability result of each data missing sub-region in the combination; Calculate the probability marginal score corresponding to each combination according to the highest predicted probability result and the second highest predicted probability result of each data missing sub-region in the combination; The score of each combination is obtained according to the model parameter uncertainty score and the probability marginal score corresponding to each combination.

5. The method for selecting a perception area in sparse crowd sensing according to claim 4, characterized in that: The score of each combination is expressed as: s=αG ω (D)+(1-α)G M (D); Among them, G ω (D)=H(y 1：b |X 1：b ,D train )-E p (ω|D train )H(y 1：b |X 1：b Oh, D train ); In the formula, s represents the score; G ω (D) represents the model parameter uncertainty score of the feature matrix set D corresponding to the combination; G M (D) represents the probability marginal score of the feature matrix set D corresponding to the combination; α represents the hyperparameter; D = {(x i ,y i )|i=1,2,3...b};x i Represents the feature matrix x in the feature matrix set D i ;y i Represents the feature matrix x in the feature matrix set D i The corresponding classification label; b represents the number of feature matrices in the feature matrix set D; H(y 1：b |x 1：b , D train ) represents the entropy of the prediction result of the first model; x 1：b represents x1, ..., x i， ..., x b ; Y 1：b represents y1, ..., yi ， ..., y b ; H(y 1：b |x 1：b ,ω,D train ) represents the entropy of the prediction result of the second model; represents the expectation of the entropy predicted by the model with respect to the model parameters, where the model parameters ω follow the distribution p(ω|D train );D train refers to the original training set that does not contain D; f BCNN (x i ) max and f BCNN (x i ) 2-max Respectively represent the feature matrix x corresponding to the i-th data missing sub-region in the combination i After being input into the target classification model, the model outputs the highest predicted probability result and the second highest predicted probability result.

6. A data classification method, characterized in that: The data classification method comprises: Obtain historical spatiotemporal data and corresponding classification labels in the target area to build an initial training data set; Train the target classification model using the current training data set; Determine whether the performance of the currently trained target classification model meets the preset performance requirements; If so, the trained target classification model is obtained to classify the data; If not, based on the sparse spatiotemporal data at time t in the target area, the perception area selection method in sparse crowd intelligence perception described in any one of claims 1 to 5 is executed to obtain each sub-area to be collected; According to the data acquisition mode of crowd intelligence perception, the spatiotemporal data of each sub-area to be collected is acquired; Adding the spatiotemporal data of the sub-area to be collected to the sparse spatiotemporal data at time t to obtain the training data at time t; the classification labels of the training data at time t are manually pre-defined; Adding the training data at time t to the initial training data set to obtain an updated training data set; Use the current updated training data set as the current training data set and return to step "Train the target classification model using the current training data set".

7. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the perception area selection method in sparse crowd sensing described in any one of claims 1 to 5 or the data classification method described in claim 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for selecting a perception area in sparse crowd sensing according to any one of claims 1 to 5 or the method for data classification according to claim 6 is implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for selecting a perception area in sparse crowd sensing according to any one of claims 1 to 5 or the method for data classification according to claim 6 is implemented.

Images