An automatic screening method for earthquake background noise dispersion curves

Abstract

This invention provides an automatic screening method for seismic background noise dispersion curves, comprising: collecting dispersion curves obtained by manually picking the maximum energy points from dispersion energy maps and theoretically generated dispersion curves; using an improved adaptive DBSCAN algorithm to eliminate discrete points in the dispersion curves for denoising; extracting feature parameters from the denoised dispersion curves; using the extracted feature parameters from the training set as input features for the dispersion curve dataset; classifying the dispersion curves in the training set into good, medium, and poor quality levels according to manual comparison with theoretical dispersion curves, and using the quality level as the output label for the dispersion curve dataset; training an improved KNN algorithm; inputting dispersion curves from the test set for screening; adjusting the algorithm based on the screening results; and increasing the diversity of dispersion curves in the training set. This method solves the problem of quality screening for extracted dispersion curves and ensures the reliability of the dispersion curves used in the inversion process.

Description

Technical Field

[0001] This invention relates to the field of geophysics, and more particularly to an automatic screening method for earthquake background noise dispersion curves. Background Technology

[0002] Background seismic noise imaging technology has been widely used in the field of seismology in recent years. One of the key steps in this technology is the extraction of surface wave dispersion curves. The quality of the dispersion curves greatly affects the accuracy of background noise imaging results. Therefore, the acquired fundamental dispersion curves must be screened to ensure their quality. The traditional manual screening method relies on human experience to select high-quality dispersion energy spectra and then manually pick out dispersion energy points to draw dispersion curves. However, this method is time-consuming and inefficient. Since background noise imaging generates tens of thousands of dispersion curves, traditional manual screening methods are insufficient for large-scale screening.

[0003] In existing technologies, in order to obtain reliable dispersion curves, it is usually necessary to manually select control points in CT images to help the program track the correct dispersion curves. However, in order to obtain more detailed subsurface structures, the number of densely packed seismic stations that need to be deployed in a study area increases rapidly. Since this method extracts dispersion data between any two stations, the number of dispersion curves that need to be extracted manually increases dramatically, requiring a large amount of human resources. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an automatic screening method for earthquake background noise dispersion curves. This method addresses the issue that most current automatic extraction methods for dispersion curves using dispersion energy spectra do not directly screen the dispersion curves. This invention can solve the problem of quality screening of the extracted dispersion curves, ensuring the reliability of the dispersion curves used in the inversion process.

[0005] This invention is implemented as follows:

[0006] An automatic screening method for earthquake background noise dispersion curves, the method comprising:

[0007] Step 1: Collect dispersion curves obtained by manually picking the maximum energy point from the dispersion energy map and theoretically generated dispersion curves, and divide them into training set and test set;

[0008] Step 2: Use the improved adaptive DBSCAN algorithm to eliminate discrete points in the dispersion curve for noise reduction;

[0009] Step 3: Extract feature parameters from the denoised dispersion curve. Feature parameters include: amplitude range, frequency range, trend, coherence, and contour coefficient. Among them, by adjusting the key parameters of the DBSCAN algorithm: scan radius Eps and minimum number of contained points MinPts, the MK (Mann-Kendall) trend test method is added to the DBSCAN algorithm to extract the feature parameters of the dispersion curve.

[0010] Step 4: Extract feature parameters from the training set as input features for the dispersion curve dataset. According to the manual comparison theory of dispersion curves, classify the dispersion curves in the training set into good, medium, and poor quality levels, and use the quality level as the output label of the dispersion curve dataset to train the improved KNN algorithm.

[0011] Step 5: Input dispersion curves from the test set for filtering, adjust the algorithm based on the filtering results, and increase the diversity of dispersion curves in the training set.

[0012] Furthermore, step 2, which utilizes the improved adaptive DBSCAN algorithm to eliminate discrete points in the dispersion curve, includes: first, normalizing the dispersion data, setting the minimum number of points (MinPts) within the range of [2, 10], and setting the scan radius (Eps) between [0.01-0.2]. Continuously adjusting these parameters until all dispersion data are grouped into one category with a label of 0, while discrete points are labeled -1, thus eliminating discrete points with a label of -1. Then, using a K-distance plot, appropriate MinPts and Eps are selected and adjusted again to ensure consistency and accurate extraction of the contour coefficients.

[0013] The distance graph is calculated by plotting the K-nearest neighbor distance (Euclidean distance) between each data point in dataset D and its K-th nearest neighbor data point, where the vertical axis represents the distance and the horizontal axis represents the number of points contained within that distance.

[0014] Furthermore, the method for finding the most suitable parameters MinPts and Eps using the K-distance graph is as follows:

[0015] Selection process:

[0016] (1) Take k (1 < k < n, where n is the number of data points), input the K-distance algorithm, find the inflection point, take Eps, MinPtsk+1.

[0017] (2) Input the DBSCAN algorithm, check the clustering effect, and get the number of clusters.

[0018] (3) Plot the number of clusters versus the value of k. When the value of k increases for more than 5 consecutive times while the number of clusters remains unchanged, we consider the clustering result to have entered the first stable interval, and we denote this number of clusters as the optimal number of clusters.

[0019] If the clustering results do not have a stable interval, instead look for a stable interval where the cluster number fluctuates within 1, and select the optimal k value in this interval based on the actual situation.

[0020] Once the value of k is selected, the optimal MinPts is k+1, and the optimal Eps is the distance at the inflection point.

[0021] Furthermore, the normalization employs a min-max normalization method to normalize the data to the range [0,1], wherein the min-max normalization method is as follows: Where x represents the original data and x′ represents the normalized data.

[0022] Furthermore, step 3 of the MK test includes: sequentially extracting the dispersion points x1, x2, ..., x... n Determine the interpolation function f(x) i -x j (n≥i>j≥1), calculate the s value, and determine the trend of the dispersion curve by judging the sign of s; determine the number of clusters of the dispersion curve by adjusting the DNSCAN parameter, and determine the coherence of the dispersion curve by the number of clusters, where the expression for the s value is:

[0023]

[0024] Furthermore, in step 3, the silhouette coefficient is used to determine the dispersion curve and the clustering effect. The cohesion a(i) represents the density between points within the same cluster, and the cohesion represents the density between different clusters. The cohesion is calculated as follows:

[0025]

[0026] The method for calculating the cohesion b(i) is changed to the average distance between points inside and outside the cluster. All points inside the cluster are traversed to obtain multiple cohesion values, {b1(i),b2(i),b3(i),...,bm(i)}, from which the smallest value is selected as the final result.

[0027] The clustering effect is judged by the silhouette coefficient s(i) value, including: a silhouette coefficient s(i) close to 1 indicates that the intra-cluster distance is less than the inter-cluster distance, the silhouette is obvious, and the clustering effect is good; a silhouette coefficient s(i) close to 1 indicates that the intra-cluster distance is greater than the inter-cluster distance, the silhouette is blurry, and the clustering effect is poor; the silhouette coefficient ranges from [-1, 1], and the expression for the silhouette coefficient s(i) is:

[0028]

[0029] Furthermore, the specific steps for training the improved KNN algorithm include:

[0030] Calculate the normalized characteristic parameters of a dispersion curve. Let there be j characteristic parameters in total. Then the characteristic parameter for the nth dispersion curve is x. n1 ,x n2 ,...,x ni If there are i dispersion curves, then all characteristic parameters correspond to x. 11 ,x 12 ,...,x ij ;

[0031] Construct a KNN model and select the parameter k value using cross-validation. The k value represents the number of points closest to the sample point.

[0032] Random forests are used to evaluate the importance of feature parameters, assigning weights to each feature parameter based on its percentage of importance. The Euclidean distance is then replaced with:

[0033]

[0034] g j x is the weight coefficient corresponding to the j-th feature parameter. ij For the j-th feature parameter and the i-th test sample, t ij Let be the j-th feature parameter and the i-th training sample.

[0035] The dispersion curves are divided into training and testing sets. Based on the feature parameters of the dispersion curves in the training set, corresponding quality labels are assigned to them for training. The training results are stored in the KNN model.

[0036] The test set dispersion curve is input into the trained KNN model to test the learning effect of the KNN algorithm. The training set is supplemented and the parameters in the KNN algorithm are adjusted until the accuracy of the test screening reaches the expected standard.

[0037] Since the traditional KNN algorithm uses the Euclidean distance formula and does not consider the differences in different feature dimensions of the data, it has low prediction accuracy for rare classes when the samples are imbalanced. Therefore, an improved KNN algorithm is introduced here.

[0038] Compared to the traditional KNN algorithm, the improved KNN algorithm introduces the concept of weights and uses random forests to calculate the impact of all feature parameters on the dispersion curve quality assessment. The calculation result is the weight g. j , representing the weight of the j-th feature parameter, the improved Euclidean distance formula is:

[0039]

[0040] x ij For the j-th feature parameter and the i-th test sample, t ijLet be the j-th feature parameter and the i-th training sample.

[0041] Compared with the prior art, the beneficial effects of this invention are as follows:

[0042] This invention primarily addresses the screening of dispersion curves generated from background noise data in a given region. Since dispersion curves are extracted from background noise data through pairwise cross-correlation calculations between stations, not all dispersion curves along these paths are necessary. Screening for high-quality dispersion curves allows for accurate inversion of the region's stratigraphic information. Compared to manual extraction and screening of dispersion curves, this invention is time-efficient and saves significant resources on machines and manpower.

[0043] This invention filters dispersion curves, retaining all the filtered dispersion curves. Dispersion curves with poor quality can still be used for imaging after processing, making full use of resources. For dispersion curves with poor shape, they can be used for empirical analysis to improve the extraction method.

[0044] With continuous improvement of the algorithm and continuous expansion of the training set (including the number and morphological diversity of dispersion curves), this invention can theoretically achieve a certain broad effect, that is, it will have a good effect on the screening of dispersion curves in various regions. Attached Figure Description

[0045] Figure 1 A schematic diagram of the process for screening earthquake background noise dispersion curves based on machine learning, provided in an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram illustrating the KNN algorithm training process and selection process as described in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram illustrating the distribution of seismic stations in the southern region of Iceland, as exemplified in this invention.

[0048] Figure 4 This is a schematic diagram illustrating the clustering of dispersion curves with respect to the DBSCAN method as described in this invention;

[0049] Figure 5 This is a schematic diagram of the dispersion energy corresponding to the dispersion curve in the example given in this invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0052] like Figure 1 The diagram shows an automatic screening method for earthquake background noise dispersion curves based on machine learning provided in an embodiment of the present invention. The method first collects manually extracted dispersion curve data, extracts feature parameters from the collected dispersion curves, and then assigns labels manually. Specifically, the feature parameters of the dispersion curves in this example include the amplitude range, frequency range, dispersion trend, dispersion degree, and coherence of the dispersion curves, while the labels are defined as the quality status.

[0053] See Table 1 for the characteristic parameters and labels corresponding to some dispersion curves in the examples of this invention:

[0054] Table 1

[0055]

[0056] The dispersion curve is normalized to remove noise points, as noise significantly impacts feature parameter extraction. Normalization prevents feature skew and improves training speed. Specifically, the amplitude and frequency range of the dispersion curve are limited to [0,1]. The dispersion data is then input into the DBSCAN algorithm. Since DBSCAN's clustering is not constrained by the shape of the data clusters, the key parameters, scanning radius Eps and minimum number of points MinPts, are manually adjusted. The DBSCAN algorithm then marks noise points in the dispersion curve as -1 and removes these points to denoise. Finally, a K-distance graph is used to select appropriate MinPts and Eps for further adjustment to ensure consistency and accurate extraction of contour coefficients.

[0057] Feature parameters are extracted from the denoised dispersion curve. The amplitude and frequency are based on the extracted true values, the dispersion is represented by the silhouette coefficient, and the coherence is represented by the number of clusters after DBSCAN clustering. Specifically, the more categories DBSCAN forms, the more discontinuous the dispersion curve is and the poorer the coherence is, and vice versa.

[0058] Among them, by adjusting the key parameters of the DBSCAN algorithm: the scanning radius Eps and the minimum number of contained points MinPts, the MK (Mann-Kendall) trend test method is added to the DBSCAN algorithm to extract the characteristic parameters of the dispersion curve;

[0059] Feature parameters extracted from the training set are used as input features for the dispersion curve dataset. Based on the artificial comparison theory of dispersion curves, the dispersion curves in the training set are classified into good, medium, and poor quality levels. The quality level is used as the output label of the dispersion curve dataset for training the KNN algorithm.

[0060] The dispersion curves in the input test set are used for filtering. The algorithm is adjusted according to the filtering results, and the diversity of the dispersion curves in the training set is increased.

[0061] Factors affecting the quality of dispersion curves include:

[0062] (1) Signal-to-noise ratio: depends on the data acquired by the station and the results of data preprocessing.

[0063] (2) Maximum Confidential Period: The data used should satisfy the condition that the distance between the two stations is greater than three times the wavelength of the Rayleigh wave used. Assuming the station spacing is Δkm and the wavelength is λ, Δ>3λ is required. Generally, a Rayleigh wave with a phase velocity of 4km / s is used. When the station spacing is Δkm, the maximum reliable period τ is... max It is Δ / 12.

[0064] (1) Repeatability: If the dispersion curve obtained on a path has high repeatability, it means that the path is relatively reliable and the obtained dispersion curve is also relatively reliable.

[0065] (2) Cluster analysis: Compare the dispersion curves obtained through the same or similar paths between two stations, classify the dispersion curves, and remove the dispersion curves with large differences (large dispersion).

[0066] The specific steps for removing discrete points using DBSCAN are as follows:

[0067] Input the dispersion curve, adjust the eps parameter to the maximum distance between the scattered points in the dispersion curve, adjust the MinPts parameter in the range of [2,10], and adjust the MinPts again in the range of [0.01-0.2] according to the specific clustering situation. Data with a label of -1 are marked as discrete points and directly filtered out.

[0068] The normalization method mentioned above refers to the following formula:

[0069]

[0070] Where x represents the original data and x′ represents the normalized data, the min-max normalization method will reduce the original data to the range of [0,1].

[0071] The MK test method for determining dispersion trend is as follows:

[0072] Step 1: Extract all dispersion points;

[0073] Step 2: Determine all n(n-1) / 2 interpolation functions f(x) i -x j (n≥i>j≥1), the interpolation function is defined as follows:

[0074] Step 3: Calculate the value of formula s, which is to subtract the number of negative differences from the number of positive differences. The formula is as follows:

[0075]

[0076] If s is positive, the data trend is increasing, and the label is 1; if s is negative, the data trend is decreasing, and the label is -1; if s is 0, the data has no obvious trend, and the label is 0.

[0077] The specific steps for extracting the contour coefficients are as follows:

[0078] Step 1: Calculate the cohesion a(i):

[0079]

[0080] To calculate the cohesion degree b(i), referencing the formula for cohesion, the calculation method for cohesion degree b(i) is changed to the average distance between points within and outside the cluster. Traverse all points within the cluster to obtain multiple cohesion degree values, {b1(i),b2(i),b3(i),...,bm(i)}, and select the smallest value from them as the final result.

[0081] Step 2: Calculate the profile coefficient s(i):

[0082]

[0083] After completing the above steps, we will now extract the feature parameters of all dispersion curves and perform KNN training:

[0084] Step 1: Calculate the normalized characteristic parameters of a dispersion curve. Let there be j characteristic parameters in total, then the characteristic parameter for the nth dispersion curve is x. n1 ,x n2 ,...,x ni If there are i dispersion curves, then all characteristic parameters correspond to x. 11 ,x12 ,...,x ij ;

[0085] Step 2: Divide the dispersion curves into training set and test set. Assign corresponding quality labels to the feature parameters of the dispersion curves based on the training set for training. Save the training results in the improved KNN.

[0086] See Figure 2 As shown, the specific steps for training the improved KNN algorithm are as follows:

[0087] Step 1: Construct the KNN model. First, select the key parameter N value using cross-validation. The N value represents the number of points closest to the sample point. The choice of N value greatly affects the partitioning of the sample points. Other parameters are selected according to the default settings of the KNN algorithm. The parameters in the KNN algorithm can be further adjusted based on the selection results.

[0088] Step 2: Using the random forest algorithm, calculate the percentage of influence of the feature parameters of all dispersion curves on the quality of the dispersion curve, and input the percentage as a weight coefficient into the improved Euclidean distance calculation formula.

[0089] Table 2 shows the characteristic parameters corresponding to some dispersion curves in the examples of this invention and their proportion of influence on the dispersion curve quality:

[0090] Table 2

[0091]

[0092] Step 3: Calculate the improved Euclidean distance:

[0093]

[0094] g j x is the weight coefficient corresponding to the j-th feature parameter. ij For the j-th feature parameter and the i-th test sample, t ij Let be the j-th feature parameter and the i-th training sample.

[0095] Step 4: The improved KNN algorithm utilizes an improved Euclidean distance. Once the parameters of the KNN algorithm are determined, the KNN algorithm is also determined. Input the test set dispersion curve into the trained KNN to test the learning effect of the KNN algorithm. Continuously adjust the parameters in the KNN algorithm until the accuracy of the test selection reaches the expected standard.

[0096] Note that the dispersion curves included in the training set will affect the KNN training effect. Here, we try to select dispersion curves with multiple scenarios as training set objects. This can not only reduce the number of training sets, but also ensure the accuracy of the final selection results. If the final selection effect is not ideal, find the dispersion curves that are incorrectly selected, manually assign labels, add them to the training set, and repeat the above process continuously.

[0097] The dispersion curve extracted in this example is from some stations on the southern platform of Iceland, such as... Figure 4 After labeling 50 dispersion curves as the training set, the accuracy of dispersion curve selection was only 66.6%. However, after labeling 100 dispersion curves as the training set, the accuracy of dispersion curve selection reached 96.7%. Furthermore, the accuracy will continue to increase as the quantity and quality of the training set improve.

[0098] The clustering results of DBSCAN on the dispersion curve of the background noise at Iceland South Station are as follows: Figure 5 .

[0099] The grouping training method for the dispersion curve of the background noise at Iceland South Station is shown in Table 3.

[0100] Table 3

[0101]

[0102]

[0103] The specific verification method is as follows:

[0104] Fifty dispersion curves were taken and divided into 10 test groups. Each group contained 45 curves in the training set and 5 curves in the test set.

[0105] The training set objects are labeled according to the feature parameters and input into the KNN for training, while the test set data is directly input into the trained objects. The output labels are compared with the manually assigned labels to obtain the preliminary accuracy.

[0106] In summary, as can be seen from the above examples, the method proposed in the embodiments of the present invention has high stability, can be directly applied to different regions of the same frequency band without additional training, and has high accuracy. It can replace manual extraction of dispersion curves, greatly saving time costs and reducing the subjectivity of manual extraction.

[0107] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic screening method for earthquake background noise dispersion curves, characterized in that, The method includes: Step 1: Collect dispersion curves obtained by manually picking the maximum energy point from the dispersion energy map and theoretically generated dispersion curves, and divide them into training set and test set; Step 2: Use the improved DBSCAN algorithm to eliminate discrete points in the dispersion curve for noise reduction; Step 3: Extract feature parameters from the denoised dispersion curve. Feature parameters include: amplitude range, frequency range, trend, coherence, and contour coefficient. Among them, by adjusting the key parameters of the DBSCAN algorithm: scan radius Eps and minimum number of contained points MinPts, the MK trend test method is added to the DBSCAN algorithm to extract the feature parameters of the dispersion curve. Step 4: Extract feature parameters from the training set as input features for the dispersion curve dataset. According to the manual comparison theory of dispersion curves, classify the dispersion curves in the training set into good, medium, and poor quality levels, and use the quality level as the output label of the dispersion curve dataset to train the KNN algorithm. Step 5: Input dispersion curves from the test set for filtering, adjust the algorithm based on the filtering results, and increase the diversity of dispersion curves in the training set.

2. The automatic dispersion curve screening method according to claim 1, characterized in that, Step 2, which uses the improved adaptive DBSCAN algorithm to eliminate discrete points in the dispersion curve, includes: first, normalizing the dispersion data, setting the minimum number of points (MinPts) in the range of [2, 10], and setting the scan radius (Eps) between [0.01-0.2]. Continuously adjusting the parameters until all the dispersion data are classified into one category with a label of 0, while the discrete points are labeled -1, eliminating the discrete points with a label of -1, and using the K-distance plot to select the most suitable minimum number of points (MinPts) and scan radius (Eps), and then adjusting the parameters again.

3. The automatic dispersion curve screening method according to claim 2, characterized in that, Selecting the optimal minimum number of points (MinPts) and scan radius (Eps) using a K-distance graph involves calculating the K-nearest neighbor distance between each data point in dataset D and its Kth nearest neighbor data point. Specifically, this includes: Selection process: (1) Take k, 1 < k < n, where n is the number of data points, input the K-distance algorithm, find the inflection point, and take Eps and MinPts; (2) Input the improved adaptive DBSCAN algorithm, check the clustering effect, and obtain the number of clusters; (3) Plot the number of clusters versus the value of k. When the value of k increases for more than 5 consecutive times while the number of clusters remains unchanged, the clustering result enters the first stable interval. This number of clusters is recorded as the optimal number of clusters. (4) When the clustering results do not have a stable interval, instead find a stable interval in which the cluster number fluctuates within 1, and select the best k value in this interval; (5) After selecting the k value, the optimal MinPts for that k value is k+1, and the optimal Eps is the distance at the inflection point.

4. The automatic dispersion curve screening method according to claim 2, characterized in that, The normalization adopts the min-max normalization method, which normalizes the data to the range [0,1]. The min-max normalization method is as follows: Where x represents the original data and x′ represents the normalized data.

5. The automatic dispersion curve screening method according to claim 1, characterized in that, Step 3 of the MK test method includes: sequentially taking the dispersion points x1, x2, ..., x... n Determine the interpolation function f(x) i -x j (n≥i>j≥1), calculate the s value, and determine the trend of the dispersion curve by judging the sign of s; determine the number of clusters of the dispersion curve by adjusting the DNSCAN parameter, and determine the coherence of the dispersion curve by the number of clusters, where the expression for the s value is:

6. The automatic dispersion curve screening method according to claim 1, characterized in that, In step 3, the silhouette coefficient is used to determine the dispersion curve and the clustering effect. The cohesion a(i) represents the density between points within the same cluster, and the cohesion represents the density between different clusters. The cohesion is calculated as follows: The method for calculating the cohesion b(i) is changed to the average distance between points inside and outside the cluster. All points inside the cluster are traversed to obtain multiple cohesion values, {b1(i),b2(i),b3(i),...,bm(i)}, from which the smallest value is selected as the final result. Clustering performance is judged by the silhouette coefficient s(i) value, including: if the silhouette coefficient s(i) is close to 1, the intra-cluster distance is smaller than the inter-cluster distance, the silhouette is obvious, and the clustering effect is good; When the coefficient s(i) approaches 1, the intra-class distance is greater than the inter-class distance, the contour is blurred, and the clustering effect is poor. The contour coefficient ranges from [-1, 1], and the expression for the contour coefficient s(i) is:

7. The automatic dispersion curve screening method according to claim 1, characterized in that, The specific steps for training the improved KNN algorithm include: Calculate the normalized characteristic parameters of a dispersion curve. Let there be j characteristic parameters in total. Then the characteristic parameter for the nth dispersion curve is x. n1 ,x n2 ,...,x ni If there are i dispersion curves, then all characteristic parameters correspond to x. 11 ,x 12 ,...,x ij ; Construct a KNN model and select the parameter k value using cross-validation. The k value represents the number of points closest to the sample point. Random forest is used to calculate the impact of all feature parameters on the dispersion curve quality assessment; the result is the weight g. j , representing the weight of the j-th feature parameter, the improved Euclidean distance formula is: x ij For the j-th feature parameter and the i-th test sample, t ij For the j-th feature parameter, the i-th training sample; The dispersion curves are divided into training and testing sets. Based on the feature parameters of the dispersion curves in the training set, corresponding quality labels are assigned to them for training. The training results are stored in the KNN model. The test set dispersion curve is input into the trained KNN model to test the learning effect of the KNN algorithm. The training set is supplemented and the parameters in the KNN algorithm are adjusted until the accuracy of the test screening reaches the expected standard.

Images