A similar typhoon path search method based on multimodal fusion

Abstract

This invention relates to the field of data processing technology and discloses a method for searching similar typhoon paths based on multimodal fusion. The method includes: acquiring and preprocessing a typhoon dataset to generate candidate typhoon trajectory data; using dynamic time warping to search for similar typhoons to the target typhoon in each candidate typhoon trajectory data; using dynamic time warping combined with Bézier curve parameterization to search for similar typhoons to the target typhoon in each candidate typhoon trajectory data; using dynamic time warping combined with path bifurcation identification to search for similar typhoons to the target typhoon in each candidate typhoon trajectory data; using dynamic time warping combined with a long short-term memory network model and a random forest model to search for similar typhoons to the target typhoon in each candidate typhoon trajectory data; and finally, performing decision-level fusion on the searched similar typhoons to generate the optimal similar typhoon. This method improves the accuracy and effectiveness of similar typhoon search by considering multiple influencing factors and multimodal fusion.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a method for searching similar typhoon paths based on multimodal fusion. Background Technology

[0002] Existing methods for searching similar typhoons are mainly based on Dynamic Time Warping (DTW). DTW uses DTW distance to determine the similarity of typhoons, without fully considering factors such as wind speed, air pressure, and the season in which the typhoon occurs, resulting in poor search performance for similar typhoons. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides a method for searching similar typhoon paths based on multimodal fusion. By fusing the results obtained from multiple methods, this method solves the problem of poor similar typhoon search performance caused by existing methods, thereby improving the search performance for similar typhoons.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0005] The method for searching similar typhoon paths based on multimodal fusion includes the following steps:

[0006] S1. Obtain the typhoon dataset, which includes several candidate typhoon data, and each candidate typhoon data includes a header record and a best path data record;

[0007] S2. Preprocess the candidate typhoon data in the typhoon dataset to generate a candidate typhoon trajectory dataset, which includes several candidate typhoon trajectory data.

[0008] S3. Use dynamic time warping to align the trajectory data of each candidate typhoon with the target typhoon, obtain the trajectory data of the candidate typhoon with the optimal alignment path with the target typhoon, and obtain N groups of first similar typhoons or 2N groups of second similar typhoons.

[0009] S4. Construct Bézier curves for the trajectory data of the target typhoon and the candidate typhoons in the 2N groups of second similar typhoons, and perform parameterized comparison to obtain the N groups of third similar typhoons.

[0010] S5. Based on the target typhoon, perform path bifurcation identification on the 2N groups of second similar typhoons to obtain N groups of fourth similar typhoons.

[0011] S6. Use the Long Short-Term Memory Network model and the Random Forest model respectively to filter the similar typhoons of the target typhoon in each candidate typhoon trajectory data, and perform weighted fusion based on 2N groups of second similar typhoons according to the set proportion weight to obtain N groups of fifth similar typhoons.

[0012] S7. Perform decision-level fusion on the N groups of first similar typhoons, N groups of third similar typhoons, N groups of fourth similar typhoons, and N groups of fifth similar typhoons. By assigning different weights, obtain the fusion score and finally obtain the N groups of optimal similar typhoons of the target typhoon.

[0013] The present invention has the following beneficial effects:

[0014] The proposed method for searching similar typhoon paths based on multimodal fusion combines dynamic time warping, dynamic time warping combined with Bézier curve parameterization, dynamic time warping combined with path bifurcation identification, and dynamic time warping combined with long short-term memory and random forest models to search for similar typhoons to the target typhoon. Finally, the four types of similar typhoons searched are fused using multimodal methods. Compared with existing similar typhoon search methods, this method uses multiple methods to identify similar typhoons and fuses them to obtain the final result. It can leverage the advantages of various methods, make up for the shortcomings of a single method, and consider all factors, resulting in high search accuracy. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the similar typhoon path search method based on multimodal fusion proposed in this invention.

[0016] Figure 2 This is a schematic diagram showing the results of searching for similar typhoons of the target typhoon using the dynamic time warping method in the embodiment.

[0017] Figure 3 This is a schematic diagram showing the results of using the dynamic time warping method combined with Bézier curve parameterization to search for similar typhoons to the target typhoon in the embodiment.

[0018] Figure 4 This is a schematic diagram showing the results of using the dynamic time warping method combined with path bifurcation to identify similar typhoons to the target typhoon in the embodiment.

[0019] Figure 5 This is a schematic diagram showing the results of searching for similar typhoons to the target typhoon using the dynamic time warping method combined with LSTM and RF models in the embodiment.

[0020] Figure 6 This is a schematic diagram showing the results of multimodal fusion of similar typhoons found in the search. Detailed Implementation

[0021] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0022] like Figure 1 As shown, the similar typhoon path search method based on multimodal fusion includes the following steps S1-S7:

[0023] S1. Obtain the typhoon dataset, which includes several candidate typhoon data, and each candidate typhoon data includes a header record and a best path data record.

[0024] Specifically, the header record includes the typhoon's classification symbol, international designation, number of rows of track data, tropical cyclone serial number (including tropical depression), tropical cyclone number, tropical cyclone termination record, number of hours between each track row, English name of the tropical cyclone, and date of formation.

[0025] Specifically, the optimal path data records include the time, intensity marker, latitude, longitude, minimum central pressure, 2-minute average maximum wind speed near the center (MSW, m / s), and 2-minute average wind speed (m / s) corresponding to the typhoon's location.

[0026] In this embodiment, in the original data txt file, each candidate typhoon data includes a header record and an optimal path data record. The fields in the header record, from left to right, are: typhoon classification symbol, international number, number of rows in the path data record, tropical cyclone number including tropical depression, my country's number for tropical cyclones, tropical cyclone termination record, number of hours between each path row, English name of the tropical cyclone, and date the dataset was created. The fields in the optimal path data record, from left to right, are: time corresponding to the typhoon location, intensity marker, latitude, longitude, minimum central pressure, 2-minute average maximum wind speed near the center, and 2-minute average wind speed.

[0027] S2. Preprocess the candidate typhoon data in the typhoon dataset to generate a candidate typhoon trajectory dataset, which includes several candidate typhoon trajectory data.

[0028] In this embodiment, the typhoon dataset can be downloaded from the Tropical Cyclone Resource Center of the China Meteorological Administration. Data preprocessing is performed on several candidate typhoon data points within the dataset, primarily including data format conversion, data attribute extraction and integration. The specific operations are as follows:

[0029] Specifically, step S2 includes S21-S22:

[0030] S21. In the header record and the best path data record, obtain the time corresponding to the typhoon location, the English name of the tropical cyclone, latitude, longitude, minimum central pressure, and 2-minute average maximum wind speed near the center. At the same time, extract the year of occurrence of the typhoon from the time corresponding to the typhoon location, add the year of occurrence of the typhoon to the English name of the tropical cyclone, and generate a unique identification number for each typhoon.

[0031] In this embodiment, the English name of a tropical cyclone is the typhoon name. Since typhoon names will appear repeatedly in historical data, in order to uniquely identify a typhoon, a combination of the typhoon name and the year in which the typhoon occurred is used to uniquely identify a typhoon, thereby obtaining a unique identification number for the typhoon.

[0032] S22. After converting the longitude and latitude coordinates of each typhoon into Web Mercator coordinates, normalization is performed to finally generate several candidate typhoon trajectory data.

[0033] In this embodiment, for the conversion and standardization of the typhoon latitude and longitude projection system, common similarity algorithms are generally applicable to the similarity calculation of planar trajectories. In order to adapt to the algorithm requirements, the latitude and longitude coordinates are converted into Web Mercator coordinates; at the same time, for the convenience of calculation, the Mercator coordinates are normalized.

[0034] S3. Use dynamic time warping to align the trajectory data of each candidate typhoon with the target typhoon, obtain the trajectory data of the candidate typhoon with the optimal alignment path with the target typhoon, and obtain N groups of first similar typhoons or 2N groups of second similar typhoons.

[0035] In this embodiment, Dynamic Time Warping (DTW) is a method for measuring the similarity between two time series, particularly suitable for comparing series of different lengths or those with temporal distortion. In typhoon trajectory similarity analysis, DTW can effectively handle the problem of time misalignment of trajectory points caused by different typhoon movement speeds. The core idea of ​​the DTW method is to find the optimal alignment between two series, minimizing the total distance after alignment. Therefore, the process of using dynamic time warping to obtain typhoons similar to the candidate typhoon trajectory data is as follows:

[0036] Specifically, step S3 includes S31-S37:

[0037] S31. Calculate the time series of each candidate typhoon trajectory data and the target typhoon, i.e.:

[0038]

[0039] in, This represents the time series of data for a candidate typhoon trajectory. This represents the time series of the target typhoon. , , , These represent the 1st, 2nd, and 3rd times of the time series of a candidate typhoon trajectory data. , No. Data points, , , , These represent the 1st, 2nd, and 3rd times in the time series of the target typhoon. , No. Data points.

[0040] S32. Calculate the Euclidean distance from any data point in the time series of each candidate typhoon trajectory data to any data point in the time series of the target typhoon, and generate a distance matrix, specifically:

[0041]

[0042] in, Represents the distance matrix. express The set of real numbers of order 1, The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon Euclidean distance of data points This represents the element in the distance matrix, i.e., the first element of the time series of candidate typhoon trajectory data. The time series from the data point to the target typhoon The Euclidean distance between the data points.

[0043] S33. Based on the distance matrix, obtain the minimum path cost of the Euclidean distance between all data points of the time series of each candidate typhoon trajectory data and all data points of the time series of the target typhoon, i.e.:

[0044]

[0045] in, The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon The minimum path cost based on the Euclidean distance between data points. This indicates taking the minimum value. The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon The minimum path cost based on the Euclidean distance between data points. The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon The minimum path cost based on the Euclidean distance between data points. The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon The minimum path cost of the Euclidean distance between data points.

[0046] Finally, the cumulative distance matrix is ​​generated. .

[0047] S34. Initialize the cumulative distance matrix, i.e.:

[0048]

[0049]

[0050] in, This means that when no data points in the time series of a candidate typhoon trajectory are aligned with the data points in the time series of the target typhoon trajectory, the minimum path cost is set to 0. This represents the first data point of the time series of a candidate typhoon trajectory. The first data point in the time series of the target typhoon The distance during alignment, i.e., the starting point of the DTW path. This represents the Euclidean distance between the first data point of the time series of candidate typhoon trajectory data and the first data point of the time series of target typhoon.

[0051] S35. Set the boundary padding conditions for the cumulative distance matrix, namely:

[0052]

[0053] in, This represents the first time series of a candidate typhoon trajectory data. Data points Matching with empty sequences, i.e., time series without a corresponding target typhoon. Data points Matching was attempted, but the reference trajectory was empty, so a match could not be found. This indicates the first time series of the target typhoon. Data points Matching with empty sequences, i.e., time series without corresponding candidate typhoon trajectory data. Data points Matching was attempted, but the target trajectory was empty, so no match could be found.

[0054] S36. Based on the initialized cumulative distance matrix and boundary filling conditions, generate the dynamic time-warped distance between each candidate typhoon trajectory data and the target typhoon, i.e.:

[0055]

[0056] in, Represents a time series of data from a candidate typhoon trajectory. Time series of target typhoons The generated dynamic time-warped distance, Time series representing candidate typhoon trajectory data Time series with the target typhoon Minimum path cost for alignment.

[0057] S37. Sort the dynamic time-normalized distances between each candidate typhoon trajectory data and the target typhoon from smallest to largest, and take the top N or 2N candidate typhoon trajectory data as similar typhoons to the target typhoon, to obtain N groups of first similar typhoons or 2N groups of second similar typhoons.

[0058] S4. Construct Bezier curves for the trajectory data of the target typhoon and the candidate typhoons in the second group of 2N similar typhoons, and perform parameterized comparison to obtain the third group of N similar typhoons.

[0059] In this embodiment, a Bézier curve construction method is introduced. Specifically, based on the 2N groups of second-similar typhoons obtained by filtering candidate typhoon trajectory data using a dynamic time warping method, Bézier curves are constructed between the target typhoon and each similar typhoon in the 2N groups. Then, through parameterized comparison, N groups of third-similar typhoons are generated, as detailed below:

[0060] Specifically, step S4 includes S41-S46:

[0061] S41. Extract the trajectory point sequence of each similar typhoon in the 2N groups of the second similar typhoons, and obtain the trajectory point number of each similar typhoon, i.e.:

[0062]

[0063] in, This represents a sequence of trajectory points along a similar typhoon. Indicates the trajectory point number. , , These represent trajectory points with serial numbers 0, 1, and 2 respectively. The location of the trajectory points.

[0064] S42. Based on the trajectory point number of each similar typhoon, perform an indexing operation on its corresponding candidate typhoon trajectory data to obtain the corresponding latitude and longitude, and generate several typhoon trajectories.

[0065] S43. Extract the key control points of each typhoon trajectory using interpolation or sparsification methods, namely:

[0066]

[0067] in, , , , These represent the 0th, 1st, and 3rd typhoon trajectories, respectively. , No. Key control points.

[0068] S44. Based on the key control points of each typhoon trajectory, generate the Bézier curves for each typhoon trajectory, i.e.:

[0069]

[0070] in, Represents the parameter in the 0-1 interval. Indicates parameters The Bezier curve of the typhoon's trajectory.

[0071] S45. Generate the Bezier curve of the target typhoon's trajectory.

[0072] In this embodiment, the process of constructing the Bézier curve of the target typhoon's trajectory is the same as the process of constructing the Bézier curve of the trajectories of similar typhoons.

[0073] S46. Compare the coordinates, curvature differences, and maximum offset of each key control point on the Bézier curve of each typhoon trajectory with the Bézier curve of the target typhoon trajectory, and generate N groups of third similar typhoons.

[0074] S5. Based on the target typhoon, perform path bifurcation identification on the 2N groups of second similar typhoons to obtain the N groups of fourth similar typhoons.

[0075] In this embodiment, the bifurcation point is where the trajectory of a group of typhoons changes significantly. Therefore, by comparing the target typhoon with each of the similar typhoons in the second group of 2N similar typhoons to see if they bifurcate, that is, if the trajectory of a similar typhoon clearly bifurcates from that of the target typhoon at a certain stage (i.e., the direction of movement deviates significantly and lasts for a certain length), then it is considered that the two typhoons are no longer similar in their subsequent paths and should be eliminated, thereby obtaining the similar typhoons of the target typhoon. The specific operation is as follows:

[0076] Specifically, step S5 includes S51-S56:

[0077] S51. Obtain the trajectory point positions of each similar typhoon in the 2N groups of the second similar typhoons, that is:

[0078]

[0079] in, This indicates the trajectory of a typhoon that is similar to another typhoon. , , , , The trajectory points of similar typhoons are numbered 0, 1, 2, and 3 respectively. , The location of the trajectory points.

[0080] S52. Obtain the location of the target typhoon's trajectory points, i.e.:

[0081]

[0082] in, This indicates the trajectory of the target typhoon. , , , , The numbers 0, 1, 2, and 3 represent the trajectory points of the target typhoon. , The location of the trajectory points.

[0083] S53, Define the window length as The target typhoon and each similar typhoon are traversed within a sliding window. The movement direction vectors of the target typhoon and each similar typhoon within each sliding window are calculated, i.e.:

[0084]

[0085]

[0086] in, Indicates that similar typhoons occurred in the [number]th [year]. The movement direction vector within each sliding window The sequence number of the trajectory point of a similar typhoon is: The location of the trajectory points, Indicates the target typhoon is in The movement direction vector within each sliding window The trajectory point number of the target typhoon is indicated as follows: The location of the trajectory points.

[0087] In this embodiment, a sliding window (such as 3 points) is used to calculate the overall movement direction (vector) of the target typhoon and similar typhoons within the window.

[0088] S54. Standardize the movement direction vectors of the target typhoon and each similar typhoon within each sliding window to obtain the standardized movement direction vectors of the target typhoon and each similar typhoon within each sliding window, i.e.:

[0089]

[0090]

[0091] in, , These represent similar typhoons in the [number]th [year]. The unit vector of the movement direction vector within each sliding window, and the target typhoon in the [number]th [phase]. The unit vector of the movement direction vector within each sliding window. This indicates that the standardized similar typhoons are in the first The movement direction vector within each sliding window This indicates the standardized target typhoon in the 1st month. The movement direction vector within a sliding window.

[0092] S55. Based on the standardized target typhoon and the movement direction vectors of each similar typhoon within each sliding window, calculate the angle between the movement direction vectors of the target typhoon and each similar typhoon within each sliding window, i.e.:

[0093]

[0094] middle, Indicates the first The angle between the movement direction vectors of the target typhoon and each similar typhoon within a sliding window. This represents the inverse cosine function.

[0095] In this embodiment, for each window, it is necessary to calculate the angle between the movement direction vectors of the target typhoon and similar typhoons in order to determine whether a bifurcation has occurred.

[0096] S56. Within several consecutive windows, determine whether the angle between the moving direction vectors of the target typhoon and each similar typhoon in each window is greater than or equal to a set threshold. If so, the trajectory of the similar typhoon has significantly bifurcated and is removed. Otherwise, the trajectory of the similar typhoon has not bifurcated and is similar to the target typhoon. Finally, N groups of fourth similar typhoons are generated.

[0097] In this embodiment, if the angle calculated for each window in a continuous window segment (such as three consecutive windows) is greater than a set threshold (such as 45 degrees), the typhoon trajectory is considered to have deviated. That is, if there is no significant bifurcation (no consecutive windows deviating beyond the threshold), the two typhoon trajectories are considered to be consistent and are judged to be similar; if there is a significant bifurcation, the trajectories are considered to have separated and are judged to be dissimilar, and the typhoon will be removed.

[0098] S6. Using the Long Short-Term Memory Network model and the Random Forest model respectively, similar typhoons to the target typhoon are screened from the candidate typhoon trajectory data. Based on the 2N groups of second similar typhoons, weighted fusion is performed according to the set proportion weight to obtain N groups of fifth similar typhoons.

[0099] In this embodiment, the Long Short-Term Memory (LSTM) model is a special type of RNN unit used to process sequential data. Its core mechanism is gating, including an input gate, a forget gate, and an output gate. Let the input sequence be:

[0100]

[0101] in, , , , They represent the 1st, 2nd, 3rd, and 4th respectively. The input vector at each time step, express 3D real space.

[0102] Let the first The hidden state at each time step is , No. The cell state at each time step is , express If the LSTM unit is in a real space, then at time step... The update formula is:

[0103] Forgotten Gate:

[0104]

[0105] in, Indicates the first The forget gate output at each time step This represents the weight matrix input to the forget gate. This represents the hidden state weight matrix of the forget gate. The bias term representing the forget gate. This represents the sigmoid activation function.

[0106] Input Gate:

[0107]

[0108] in, Indicates the first The input gate outputs at each time step (determine how much new information to write at the current time). This represents the input weight matrix of the input gate. This represents the hidden state weight matrix of the input gate. This represents the bias term of the input gate.

[0109] Candidate state:

[0110]

[0111] in, Indicates the first The new candidate cell states generated at each time step This represents the hyperbolic tangent activation function. The input weight matrix represents the candidate state. The hidden state weight matrix represents the candidate state. The bias term represents the candidate state.

[0112] Cell status update:

[0113]

[0114] in, Indicates the first Cell state updated at each time step (long-term memory). Indicates the first Cell state at each time step This indicates element-wise multiplication.

[0115] Output gate:

[0116]

[0117] in, Indicates the first Output gate value at each time step This represents the input weight matrix of the output gate. This represents the hidden state weight matrix of the output gate. This represents the bias term of the output gate.

[0118] Hidden status update:

[0119]

[0120] in, Indicates the first The hidden state at each time step (which is also the final output of the LSTM unit).

[0121] Random Forest (RF) is an ensemble learning method consisting of multiple decision trees used for classification or regression. Let the training dataset be... for:

[0122]

[0123] in, , , Indicates the 1st, 2nd, and 3rd The input feature vector of each sample, , , respectively representing the 1st, 2nd, and 3rd The target label corresponding to each sample The total number of samples.

[0124] The construction process of Random Forest (RF) is as follows:

[0125] 1) Bagging sampling: Sampling with replacement from the dataset. Next, generate subsets. , , ..., . Indicates the first Sub-training sets (from) Random sampling with replacement , The total number of decision trees in a random forest.

[0126] 2) Construct each decision tree : Randomly select a subset of features at each node, and choose the best splitting features and threshold to minimize a certain loss function (such as Gini exponent or MSE).

[0127] Therefore, based on the working principles of Random Forest and Long Short-Term Memory models, similar typhoons to the target typhoon are screened. Then, the results of DTW, LSTM, and RF are initially fused to obtain the fused screening results of the three models. The specific operation process is as follows:

[0128] Specifically, step S6 includes S61-S64:

[0129] S61. Extract the trajectory data of each candidate typhoon, the latitude and longitude of the target typhoon, the corresponding time, the 2-minute average maximum wind speed near the center, the minimum central pressure, and the month in which the typhoon occurred, and encode them to generate the feature sequences of each candidate typhoon and the feature sequence of the target typhoon, namely:

[0130]

[0131]

[0132] in, Represents a set of feature sequences. , , , They represent the 1st, 2nd, and 3rd respectively. , No. Feature sequences, Indicates encoding operation. Represents normalized latitude and longitude information. The unique hot code representing the month. Indicates the lowest central pressure. This indicates the average maximum wind speed near the center over a 2-minute period. This indicates the time corresponding to each latitude and longitude of the typhoon.

[0133] S62. Use the feature sequence of the target typhoon as the label data, and the feature sequence of each candidate typhoon trajectory data as the raw data. Use the raw data to be labeled to train the Long Short-Term Memory Network Model and the Random Forest Model respectively, and obtain the trained Long Short-Term Memory Network Model and Random Forest Model.

[0134] S63. Input the trajectory data of each candidate typhoon into the trained Long Short-Term Memory Network Model and Random Forest Model again to identify similar typhoons, and generate similar typhoons identified by the Long Short-Term Memory Network Model and the Random Forest Model respectively.

[0135] S64. Set the proportional weights, and then sort the similar typhoons identified by the Long Short-Term Memory Network Model, the similar typhoons identified by the Random Forest Model, and the 2N groups of second similar typhoons after weighted fusion. Output the top N similar typhoons to obtain the N groups of fifth similar typhoons.

[0136] Specifically, the formula for weighted fusion is:

[0137]

[0138] in, Indicates the fusion result. , , , representing the proportional weights of the second most similar typhoon in the 2N groups, the proportional weights of the similar typhoons identified by the Long Short-Term Memory Network model, and the proportional weights of the similar typhoons identified by the Random Forest model, respectively. This indicates the second most similar typhoon in group 2N. This represents similar typhoons identified by the Long Short-Term Memory network model. This represents similar typhoons identified by the random forest model.

[0139] In this embodiment, , , The values ​​are 0.5, 0.3, and 0.2.

[0140] S7. Perform decision-level fusion on the N groups of first similar typhoons, N groups of third similar typhoons, N groups of fourth similar typhoons, and N groups of fifth similar typhoons. By assigning different weights, obtain the fusion score and finally obtain the N groups of optimal similar typhoons of the target typhoon.

[0141] Specifically, step S7 includes S71-S73:

[0142] S71. Using a trial-and-error method, assign different weights to the N groups of first-similar typhoons, N groups of third-similar typhoons, N groups of fourth-similar typhoons, and N groups of fifth-similar typhoons.

[0143] In this embodiment, firstly, all typhoon IDs that have appeared are collected from the similarity results of various methods, and the target typhoon itself is removed to avoid self-comparison. Secondly, in order to achieve multimodal typhoon similarity analysis, this invention integrates multiple similarity calculation models and assigns different weights. The weights are assigned using a trial-and-error method, and the weight of the Dynamic Time Warping (DTW) method is set to 0.35. The weight of the result of joint modeling of DTW and Bezier parametric curves is 0.35. The weight of the joint result of DTW and path bifurcation point identification is 0.15. The weight of the result based on the fusion model of Long Short-Term Memory Network (LSTM) and Random Forest (RF) combined with DTW is 0.15.

[0144] S72. For the first, third, fourth, and fifth similar typhoons in the N groups, determine whether each similar typhoon has a distance metric value. If so, proceed to step S73; otherwise, proceed to step S74.

[0145] S73. Convert the distance metric values ​​of each similar typhoon in the N groups of first, third, fourth, and fifth similar typhoons into similarity values, respectively:

[0146]

[0147] in, Represents the similarity value. This represents the distance metric.

[0148] S74. Multiply the similarity value of each similar typhoon in the N groups of first, third, fourth, and fifth similar typhoons by its corresponding weight to obtain the fusion score of each similar typhoon.

[0149] S75. Sort the fusion scores of the similar typhoons from high to low, and take the top N typhoons as the path candidates most similar to the target typhoon, thus obtaining the N groups of optimal similar typhoons to the target typhoon.

[0150] In this embodiment, for steps S72-S75, the initial fusion score of all similar typhoons is 0. Subsequently, the results of each similarity analysis method are iterated through. For each similar typhoon, it is checked whether it appears in the method and a distance metric is obtained. If it exists, the distance value is converted into a similarity value. The similarity scores are multiplied by their corresponding weights and added to the typhoon's fusion score. If a typhoon does not appear in the results of a certain method, a smaller default similarity (e.g., 0.01) is used in the weighted calculation. Finally, the fusion scores of all similar typhoons are converted into a list, sorted from highest to lowest score, and the top N typhoons with the highest fusion scores are returned as the path candidates most similar to the target typhoon, thus obtaining the N optimal similar typhoons to the target typhoon.

[0151] In addition, to verify the effectiveness of the proposed multimodal fusion-based similar typhoon path search method, the following experiments were conducted:

[0152] First, data processing is performed. The original data (typhoon dataset) is the CMA best path dataset, which is a txt file named according to the year. This file contains data on all typhoons that occurred in that year. Each typhoon dataset includes two parts: a header record and a best path data record. Partial data is shown in Tables 1 and 2:

[0153] Table 1 Header Record Table

[0154]

[0155] In Table 1, the classification identifier for typhoons consists of 5 characters, such as "66666" indicating optimal path data; the international designation consists of 4 characters, representing the last two digits of the year followed by a two-digit number, such as "0000" indicating the last two digits of the year are 00 and the designation is 00; the number of rows in the path data record consists of 3 characters, such as "049" indicating 49 rows; the sequence number for tropical cyclones, including tropical depressions, consists of 4 characters, such as "0001" indicating the sequence number is 0001; the Chinese designation for tropical cyclones consists of 4 characters, such as "0000" indicating the designation is 0000; and the tropical cyclone termination record consists of 1 character, where 0 represents... The numbers indicate dissipation, 1 indicates moving out of the responsibility area of ​​the Western Pacific Typhoon Committee, 2 indicates merging, and 3 indicates quasi-stationary. The number of hours between each path is 1 character, and it is 6 hours before 2017. For cases that made landfall starting in 2017, there are 3-hour encrypted records, and the rest are still 6 hours. The English name of a tropical cyclone can be up to 20 characters long. For example, "Carmen" indicates the English name of a tropical cyclone. In addition, "(-1)n" can be added after the name to indicate the subcenter and its sequence number. The date the dataset was formed is 8 characters long. For example, "20110729" indicates that it was formed on July 29, 2011.

[0156] Table 2 Optimal Path Data Record Table

[0157]

[0158] In Table 2, the time corresponding to the typhoon location is expressed as YYYY year, MM month, DD day, HH hour (UTC). Therefore, "1949011300" represents 00:00 on January 13, 1949. The intensity is marked as the average wind speed from 2 minutes before the exact time to the exact time, which can be referenced in the national standard "Tropical Cyclone Classification" (GB / T). (19201-2006), where 0 indicates weaker than a tropical depression (TD) or unknown level, 1 indicates a tropical depression (TD, 10.8-17.1 m / s), 2 indicates a tropical storm (TS, 17.2-24.4 m / s), 3 indicates a severe tropical storm (STS, 24.5-32.6 m / s), 4 indicates a typhoon (TY, 32.7-41.4 m / s), 5 indicates a severe typhoon (STY, 41.5-50.9 m / s), 6 indicates a super typhoon (SuperTY, ≥51.0 m / s), 9 indicates transformation, with the first mark indicating transformation completed; MSW represents the 2-minute average maximum wind speed near the center (MSW, m / s), where MSW=15 indicates a 2-minute average maximum wind speed near the center of 15 m / s, MSW=20 indicates a 2-minute average maximum wind speed near the center of 20 m / s, and WSW=0 indicates missing measurement.

[0159] Then, the data such as the time corresponding to the typhoon location, the English name of the tropical cyclone (typhoon name), latitude and longitude, minimum central pressure, and 2-minute average maximum wind speed near the center were extracted and integrated from the original data. The latitude and longitude were then converted to Web Mercator projection and normalized. Some of the data are shown in Table 3.

[0160] Table 3. Normalized Typhoon Data (Candidate Typhoon Track Data)

[0161]

[0162] Therefore, based on the trajectory data of each candidate typhoon, different methods are used to search for similar typhoons to the target typhoon, and the following operations are performed respectively:

[0163] 1) The Dynamic Time Warping (DTW) method is used to search for similar typhoons. Specifically, given a target typhoon (e.g., Carmen_1949), several typhoons similar to it are searched in the candidate typhoon trajectory data. The results are as follows: Figure 2 As shown, where Figure 2 The red curve represents the target typhoon, while the other five colored curves represent similar typhoons obtained through DTW search. Figure 2 As can be seen, the obtained similar typhoons have highly similar paths and trajectories to the target typhoon.

[0164] 2) The DTW method combined with Bézier curve parameterization is used to search for similar typhoons. The search results are as follows: Figure 3 As shown, where Figure 3 The red curve represents the target typhoon, while the other five colored curves represent similar typhoons obtained through the DTW method combined with Bézier curve parameterization search. Figure 3 As can be seen, the obtained similar typhoons have highly similar paths and trajectories to the target typhoon.

[0165] 3) The DTW method combined with path bifurcation identification is used to search for similar typhoons, and the search results are as follows: Figure 4 As shown, where Figure 4 The red curve represents the target typhoon, while the other five colored curves represent similar typhoons obtained through the DTW method combined with path bifurcation identification. Figure 4 As can be seen, the obtained similar typhoons have highly similar paths and trajectories to the target typhoon.

[0166] 4) The DTW method combined with LSTM and RF models is used to search for similar typhoons. The search results are as follows: Figure 5 As shown, where Figure 5 The red curve represents the target typhoon, while the other five colored curves represent similar typhoons found using the DTW method combined with LSTM and RF models. Figure 5 As can be seen, the obtained similar typhoons have highly similar paths and trajectories to the target typhoon.

[0167] Finally, the similar typhoons found by the above four methods are merged, and the fusion result is as follows: Figure 6 As shown, where Figure 6 The red curve represents the target typhoon, while the other five colored curves represent the fusion results of similar typhoons found using the four methods described above. Figure 6 As can be seen, the obtained similar typhoon fusion results are highly similar to the target typhoon's path trajectory.

[0168] Meanwhile, DTW distance and curve distance are used as evaluation metrics for this invention. The DTW distance and curve distance in the multimodal fusion search results are shown in Table 4:

[0169] The DTW distance calculation formula is as follows:

[0170]

[0171] in, For point and Distance (such as Euclidean distance). This represents a valid alignment path. , Representing the time series of candidate typhoon trajectory data respectively Track point index and time series of target typhoons The trajectory point index in the data. This indicates taking the minimum value.

[0172] The Bézier curve distance is defined as the integral of the distance between two curves over the interval [0, 1]. The calculation formula is:

[0173]

[0174] in, Indicates in the parameter Integrating the entire curve above corresponds to the curve's progression from the starting point to the ending point. A Bézier curve representing the trajectory of a target typhoon; The Bézier curve representing candidate typhoon trajectory data. Indicates the same parameter Below, the Euclidean distance between the two curves.

[0175] Therefore, the smaller the DTW distance value, the closer the shapes and trends of the two trajectories are, and the more similar they are. The smaller the curve distance value, the closer the parameters of the two curves are, and the more similar their shapes are.

[0176] Table 4. DTW distance and curve distance in multimodal fusion search results.

[0177]

[0178] As shown in Table 4, DTW distance and curve distance are roughly positively correlated. DTW and curve distance measure similarity at different levels; one is more local matching, and the other is more global shape. By fusing multimodal search results, taking into account both point matching and overall shape, similar typhoons can be screened more comprehensively and accurately.

[0179] In summary, the multimodal fusion-based similar typhoon path search method proposed in this invention combines dynamic time warping, dynamic time warping combined with Bézier curve parameterization, dynamic time warping combined with path bifurcation identification, and dynamic time warping combined with long short-term memory and random forest models to search for similar typhoons to the target typhoon. Finally, the four types of similar typhoons searched are fused using multimodal methods. Compared with existing similar typhoon search methods, this method uses multiple methods to identify similar typhoons and fuses them to obtain the final result. It can leverage the advantages of various methods, make up for the shortcomings of a single method, and considers all factors, resulting in high search accuracy.

[0180] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0181] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A similar typhoon path search method based on multimodal fusion, characterized in that, Includes the following steps: S1. Obtain the typhoon dataset, which includes several candidate typhoon data, and each candidate typhoon data includes a header record and a best path data record; S2. Preprocess the candidate typhoon data in the typhoon dataset to generate a candidate typhoon trajectory dataset, which includes several candidate typhoon trajectory data. S3. Use dynamic time warping to align the trajectory data of each candidate typhoon with the target typhoon, obtain the trajectory data of the candidate typhoon with the optimal alignment path with the target typhoon, and obtain N groups of first similar typhoons or 2N groups of second similar typhoons. S4. Construct Bézier curves for the trajectory data of the target typhoon and the candidate typhoons in the 2N groups of second-similar typhoons, and perform parameterized comparisons to obtain the N groups of third-similar typhoons, specifically: Based on the key control points of each typhoon's trajectory, Bézier curves for each typhoon's trajectory are generated, namely: in, Represents the parameter in the 0-1 interval. Indicates parameters The Bezier curve of the typhoon's trajectory; Generate the Bézier curve of the target typhoon's trajectory; By comparing the coordinates, curvature differences, and maximum offset of each key control point on the Bézier curve of each typhoon trajectory with the Bézier curve of the target typhoon trajectory, N groups of third similar typhoons are generated. S5. Based on the target typhoon, perform path bifurcation identification on the 2N groups of second similar typhoons to obtain N groups of fourth similar typhoons, specifically: Obtain the trajectory point positions of the target typhoon and each second similar typhoon, and perform a window sliding operation on the target typhoon and each similar typhoon, and calculate the movement direction vector of the target typhoon and each similar typhoon in each sliding window; After standardizing the movement direction vector, the angle between the movement direction vector of the target typhoon and each similar typhoon within each sliding window is calculated. Within several consecutive windows, by comparing the included angle with a set threshold, similar typhoons that have not branched are filtered out, and N groups of fourth similar typhoons are generated. S6. Use the Long Short-Term Memory Network model and the Random Forest model respectively to filter the similar typhoons of the target typhoon in each candidate typhoon trajectory data, and perform weighted fusion based on 2N groups of second similar typhoons according to the set proportion weight to obtain N groups of fifth similar typhoons. S7. Perform decision-level fusion on the N groups of first similar typhoons, N groups of third similar typhoons, N groups of fourth similar typhoons, and N groups of fifth similar typhoons. By assigning different weights, obtain the fusion score and finally obtain the N groups of optimal similar typhoons of the target typhoon.

2. The method for searching similar typhoon paths based on multimodal fusion according to claim 1, characterized in that, The header record includes the typhoon's classification symbol, international designation, number of rows of track data, tropical cyclone serial number (including tropical depression), tropical cyclone number, tropical cyclone termination record, number of hours between each track row, English name of the tropical cyclone, and date of formation.

3. The method for searching similar typhoon paths based on multimodal fusion according to claim 2, characterized in that, The optimal path data record includes the time, intensity marker, latitude, longitude, minimum central pressure, 2-minute average maximum wind speed near the center, and 2-minute average wind speed corresponding to the typhoon's location.

4. The similar typhoon path search method based on multimodal fusion according to claim 3, characterized in that, Step S2 specifically includes: S21. In the head record and the best path data record, obtain the time corresponding to the typhoon location, the English name of the tropical cyclone, latitude, longitude, minimum central pressure, and 2-minute average maximum wind speed near the center. At the same time, extract the year of occurrence of the typhoon from the time corresponding to the typhoon location, add the year of occurrence of the typhoon to the English name of the tropical cyclone, and generate a unique identification number for each typhoon. S22. After converting the longitude and latitude coordinates of each typhoon into Web Mercator coordinates, normalization is performed to finally generate several candidate typhoon trajectory data.

5. The method for searching similar typhoon paths based on multimodal fusion according to claim 4, characterized in that, Step S3 specifically includes: S31. Calculate the time series of each candidate typhoon trajectory data and the target typhoon, i.e.: in, This represents the time series of data for a candidate typhoon trajectory. This represents the time series of the target typhoon. , , , These represent the 1st, 2nd, and 3rd times of the time series of a candidate typhoon trajectory data. , No. Data points, , , , These represent the 1st, 2nd, and 3rd times in the time series of the target typhoon. , No. One data point; S32. Calculate the Euclidean distance from any data point in the time series of each candidate typhoon trajectory data to any data point in the time series of the target typhoon, and generate a distance matrix, specifically: in, Represents the distance matrix. express The set of real numbers of order 1, The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon Euclidean distance of data points This represents the element in the distance matrix, i.e., the first element of the time series of candidate typhoon trajectory data. The time series from the data point to the target typhoon Euclidean distance of each data point; S33. Based on the distance matrix, obtain the minimum path cost of the Euclidean distance between all data points of the time series of each candidate typhoon trajectory data and all data points of the time series of the target typhoon, i.e.: in, The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon The minimum path cost based on the Euclidean distance between data points. This indicates taking the minimum value. The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon The minimum path cost based on the Euclidean distance between data points. The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon The minimum path cost based on the Euclidean distance between data points. The first time series of candidate typhoon trajectory data The time series from the data point to the target typhoon The minimum path cost of the Euclidean distance between data points; Finally, the cumulative distance matrix is ​​generated. ; S34. Initialize the cumulative distance matrix, i.e.: in, This means that when no data points in the time series of a candidate typhoon trajectory are aligned with the data points in the time series of the target typhoon trajectory, the minimum path cost is set to 0. This represents the first data point of the time series of a candidate typhoon trajectory. The first data point in the time series of the target typhoon The distance during alignment, i.e., the starting point of the DTW path. The Euclidean distance is represented by the first data point in the time series of the candidate typhoon trajectory data to the first data point in the time series of the target typhoon. S35. Set the boundary padding conditions for the cumulative distance matrix, namely: in, This represents the first time series of a candidate typhoon trajectory data. Data points Matching with empty sequences, i.e., time series without a corresponding target typhoon. Data points Matching was attempted, but the reference trajectory was empty, so a match could not be found. This indicates the first time series of the target typhoon. Data points Matching with empty sequences, i.e., time series without corresponding candidate typhoon trajectory data. Data points Matching was attempted, but the target trajectory was empty, so no match could be found. S36. Based on the initialized cumulative distance matrix and boundary filling conditions, generate the dynamic time-warped distance between each candidate typhoon trajectory data and the target typhoon, i.e.: in, Represents a time series of data from a candidate typhoon trajectory. Time series of target typhoons The generated dynamic time-warped distance, Time series representing candidate typhoon trajectory data The Time series of data points and target typhoon The Minimum path cost between data points; S37. Sort the dynamic time-normalized distances between each candidate typhoon trajectory data and the target typhoon from smallest to largest, and take the top N or 2N candidate typhoon trajectory data as similar typhoons to the target typhoon, to obtain N groups of first similar typhoons or 2N groups of second similar typhoons.

6. The method for searching similar typhoon paths based on multimodal fusion according to claim 5, characterized in that, Step S4 specifically includes: S41. Extract the trajectory point sequence of each similar typhoon in the 2N groups of the second similar typhoons, and obtain the trajectory point number of each similar typhoon, i.e.: in, This represents a sequence of trajectory points along a similar typhoon. Indicates the trajectory point number. , , These represent trajectory points with serial numbers 0, 1, and 2 respectively. The position of the trajectory point; S42. Based on the trajectory point number of each similar typhoon, perform an indexing operation on its corresponding candidate typhoon trajectory data to obtain the corresponding latitude and longitude, and generate several typhoon trajectories. S43. Extract the key control points of each typhoon trajectory using interpolation or sparsification methods, namely: in, , , , These represent the 0th, 1st, and 3rd typhoon trajectories, respectively. , No. One key control point; S44. Based on the key control points of each typhoon trajectory, generate the Bézier curves for each typhoon trajectory, i.e.: in, Represents the parameter in the 0-1 interval. Indicates parameters The Bezier curve of the typhoon's trajectory; S45. Generate the Bézier curve of the target typhoon's trajectory; S46. Compare the coordinates, curvature differences, and maximum offset of each key control point on the Bézier curve of each typhoon trajectory with the Bézier curve of the target typhoon trajectory, and generate N groups of third similar typhoons.

7. The method for searching similar typhoon paths based on multimodal fusion according to claim 6, characterized in that, Step S5 specifically includes: S51. Obtain the trajectory point positions of each similar typhoon in the 2N groups of the second similar typhoons, that is: in, This indicates the trajectory of a typhoon that is similar to another typhoon. , , , , The trajectory points of similar typhoons are numbered 0, 1, 2, and 3 respectively. , The position of the trajectory point; S52. Obtain the location of the target typhoon's trajectory points, i.e.: in, This indicates the trajectory of the target typhoon. , , , , The numbers 0, 1, 2, and 3 represent the trajectory points of the target typhoon. , The position of the trajectory point; S53, Define the window length as The target typhoon and each similar typhoon are traversed within a sliding window. The movement direction vectors of the target typhoon and each similar typhoon within each sliding window are calculated, i.e.: in, Indicates that similar typhoons occurred in the [number]th [year]. The movement direction vector within each sliding window The sequence number of the trajectory point of a similar typhoon is: The location of the trajectory points, Indicates the target typhoon is in The movement direction vector within each sliding window The trajectory point number of the target typhoon is indicated as follows: The position of the trajectory point; S54. Standardize the movement direction vectors of the target typhoon and each similar typhoon within each sliding window to obtain the standardized movement direction vectors of the target typhoon and each similar typhoon within each sliding window, i.e.: in, , These represent similar typhoons in the [number]th [year]. The unit vector of the movement direction vector within each sliding window, and the target typhoon in the [number]th [phase]. The unit vector of the movement direction vector within each sliding window. This indicates that the standardized similar typhoons are in the first The movement direction vector within each sliding window This indicates the standardized target typhoon in the 1st month. The movement direction vector within each sliding window; S55. Based on the standardized target typhoon and the movement direction vectors of each similar typhoon within each sliding window, calculate the angle between the movement direction vectors of the target typhoon and each similar typhoon within each sliding window, i.e.: middle, Indicates the first The angle between the movement direction vectors of the target typhoon and each similar typhoon within a sliding window. Represents the inverse cosine function; S56. Within several consecutive windows, determine whether the angle between the moving direction vectors of the target typhoon and each similar typhoon in each window is greater than or equal to a set threshold. If so, the trajectory of the similar typhoon has significantly bifurcated and is removed. Otherwise, the trajectory of the similar typhoon has not bifurcated and is similar to the target typhoon. Finally, N groups of fourth similar typhoons are generated.

8. The method for searching similar typhoon paths based on multimodal fusion according to claim 7, characterized in that, Step S6 specifically includes: S61. Extract the trajectory data of each candidate typhoon, the latitude and longitude of the target typhoon, the corresponding time, the 2-minute average maximum wind speed near the center, the minimum central pressure, and the month in which the typhoon occurred, and encode them to generate the feature sequences of each candidate typhoon and the feature sequence of the target typhoon, namely: in, Represents a set of feature sequences. , , They represent the 1st, 2nd, and 3rd respectively. Feature sequences, Indicates encoding operation. Represents normalized latitude and longitude information. The unique hot code representing the month. Indicates the lowest central pressure. This indicates the average maximum wind speed near the center over a 2-minute period. This indicates the time corresponding to each latitude and longitude of the typhoon; S62. Use the feature sequence of the target typhoon as the label data and the feature sequence of each candidate typhoon trajectory data as the raw data. Use the labeled raw data to train the Long Short-Term Memory Network Model and the Random Forest Model respectively to obtain the trained Long Short-Term Memory Network Model and the Random Forest Model. S63. Input the trajectory data of each candidate typhoon into the trained Long Short-Term Memory Network Model and Random Forest Model respectively to identify similar typhoons, and generate similar typhoons identified by the Long Short-Term Memory Network Model and similar typhoons identified by the Random Forest Model respectively. S64. Set the proportional weights, and then sort the similar typhoons identified by the Long Short-Term Memory Network Model, the similar typhoons identified by the Random Forest Model, and the 2N groups of second similar typhoons after weighted fusion. Output the top N similar typhoons to obtain the N groups of fifth similar typhoons.

9. The method for searching similar typhoon paths based on multimodal fusion according to claim 8, characterized in that, The formula for weighted fusion is: in, Indicates the fusion result. , , , representing the proportional weights of the second most similar typhoon in the 2N groups, the proportional weights of the similar typhoons identified by the Long Short-Term Memory Network model, and the proportional weights of the similar typhoons identified by the Random Forest model, respectively. This indicates the second most similar typhoon in group 2N. This represents similar typhoons identified by the Long Short-Term Memory network model. This represents similar typhoons identified by the random forest model.

10. The method for searching similar typhoon paths based on multimodal fusion according to claim 9, characterized in that, Step S7 specifically includes: S71. Using a trial-and-error method, assign different weights to the N groups of first-similar typhoons, N groups of third-similar typhoons, N groups of fourth-similar typhoons, and N groups of fifth-similar typhoons. S72. For the first, third, fourth, and fifth similar typhoons in the N groups, determine whether each similar typhoon has a distance measurement value. If so, proceed to step S73; otherwise, proceed to step S74. S73. Convert the distance metric values ​​of each similar typhoon in the N groups of first, third, fourth, and fifth similar typhoons into similarity values, respectively: in, Represents the similarity value. Represents the distance metric; S74. Multiply the similarity value of each similar typhoon in the N groups of first similar typhoons, N groups of third similar typhoons, N groups of fourth similar typhoons, and N groups of fifth similar typhoons by their corresponding weights to obtain the fusion score of each similar typhoon. S75. Sort the fusion scores of the similar typhoons from high to low, and take the top N typhoons as the path candidates most similar to the target typhoon, thus obtaining the N groups of optimal similar typhoons to the target typhoon.

Images