Typhoon multi-task estimation method based on prior identity orientation and physical invariance

Abstract

The present application belongs to the technical field of real-time estimation of typhoon, and particularly relates to a typhoon multi-task estimation method based on prior identity orientation and physical invariance. The present application comprises: constructing an estimation network model; using an identity token initialization module to extract multi-modal features from time-series multi-channel infrared data and associated data, and obtaining an initial identity token based on the multi-modal features; using a task-dependent flow learning module to obtain special identity tokens for four tasks of pressure, wind speed, inner core wind circle and outer core wind circle; using an associated information bridge module to obtain a shared identity token based on the initial identity token; splicing the shared identity token, the special identity tokens of different tasks and the multi-modal features to obtain spliced features; and inputting the spliced features into an identity orientation estimation module to obtain estimated wind speed, pressure, inner core wind circle radius and outer core wind circle radius. The present application avoids the concept drift problem existing in simultaneous learning of multiple tasks, and guides the model to more accurately and effectively estimate multiple attributes of typhoon.

Description

Technical Field

[0001] This invention belongs to the field of typhoon real-time estimation technology, and in particular relates to a typhoon multi-task estimation method based on prior identity guidance and physical invariance. Background Technology

[0002] Typhoon estimation utilizes meteorological data, including satellite data, to estimate the current attributes of a typhoon in real time. Real-time estimation of typhoon intensity and wind circle size is essential for preventing weather disasters.

[0003] With the continuous development of artificial intelligence technology, many studies have proposed various deep learning methods to simultaneously estimate multiple attributes of typhoons, thereby helping to mitigate the disasters caused by strong typhoons. Currently, most deep learning-based methods, while achieving simultaneous estimation of typhoon wind speed and wind circle, do not fully consider the distribution offset problem inherent in multi-task typhoon estimation, resulting in insufficient generalization ability for learning unknown typhoons. Furthermore, existing deep learning methods often rely solely on data-driven representation learning, neglecting the distributional constraints of typhoon-related physical priors on model representations. Therefore, this invention proposes a typhoon multi-task estimation method based on prior identity guidance and physical invariance. Summary of the Invention

[0004] The purpose of this invention is to solve the problems of existing technical solutions being unable to effectively cope with distribution shifts and lacking physical prior constraints, and to propose a typhoon multi-task estimation method based on prior identity guidance and physical invariance.

[0005] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0006] A typhoon multi-task estimation method based on prior identity guidance and physical invariance includes:

[0007] Acquire time-series multi-channel infrared data and associated data;

[0008] Construct an estimation network model, including an identity token initialization module, a task dependency flow learning module, an association information bridge module, and an identity-oriented estimation module;

[0009] Using the identity token initialization module, multimodal features are extracted from time-series multi-channel infrared data and associated data, and an initial identity token is obtained based on the multimodal features;

[0010] Using the task-dependent stream learning module, special identity tokens for four tasks—pressure, wind speed, core wind circle, and outer core wind circle—are obtained based on development data. The development data refers to the typhoon's intensity level and duration of existence at the previous moment.

[0011] Using the associated information bridge module, a shared identity token is obtained based on the initial identity token;

[0012] By concatenating shared identity tokens, special identity tokens for different tasks, and multimodal features, concatenated features for different tasks are obtained.

[0013] By inputting the spliced ​​features of different tasks into the identity-guided estimation module, the estimated wind speed, pressure, core wind circle radius, and outer core wind circle radius are obtained.

[0014] Furthermore, the identity token initialization module extracts multimodal features from temporal multi-channel infrared data and associated data, and obtains an initial identity token based on the multimodal features, including:

[0015] The identity token initialization module includes two parts: an encoding network and a normal distribution sampling. The encoding network includes two deep convolutional networks, an associative encoder, and a self-attention long short-term memory network.

[0016] The coding network is used to extract multimodal features from temporal multichannel infrared data and associated data;

[0017] Normal distribution sampling is used to statistically analyze the mean and variance of multimodal features. A normal distribution is defined using the mean and variance, and distribution features of a specified dimension are sampled from the normal distribution as the initial identity token.

[0018] Furthermore, the task-dependent flow learning module obtains special identity tokens for four tasks—pressure, wind speed, core wind circle, and outer core wind circle—based on development data, including:

[0019] The task-dependent flow learning module includes a development encoder and two prior approximators. The development encoder includes two identical multilayer perceptrons, and the prior approximators include two graph convolutions and a self-attention long short-term memory network.

[0020] The typhoon's intensity level and duration data are input into two multilayer perceptrons. Each multilayer perceptron obtains two distribution modulation features. The two distribution modulation features are used as coefficients and biases, respectively, to transform the initial identity token. The two multilayer perceptrons obtain special identity tokens for wind speed and pressure, respectively.

[0021] The special identity tokens for the two tasks of wind speed and pressure are input into the prior approximator to extract the special identity tokens for the two tasks of the core wind circle and the outer core wind circle.

[0022] Furthermore, the initial identity token is transformed using the following formula:

[0023] The special identity token for either the core wind circle or the outer core wind circle task = gamma × initial identity token + bata;

[0024] Among them, gamma and bata are distribution moderating features.

[0025] Furthermore, the step of obtaining a shared identity token based on the initial identity token using the associated information bridge module includes:

[0026] The associated information bridge module includes an associated encoder, which includes two graph convolutions.

[0027] Based on the prior associations between tasks, a dark knowledge graph is predefined. The dark knowledge graph is then input into the association encoder to extract k weighted weights.

[0028] The initial identity token is fitted with k single Gaussian distributions, and the k single Gaussian distributions are weighted with k weighted weights to obtain a Gaussian mixture distribution model. The distribution characteristics of the weighted Gaussian mixture distribution model are used as the shared identity token.

[0029] Furthermore, the step of predefining a dark knowledge graph based on prior associations between tasks includes:

[0030] The types of associated data are used as the x-axis of the dark knowledge graph, and the tasks related to the associated data are used as the y-axis. The dark knowledge graph is predefined as an N×N matrix.

[0031] If the type of associated data is related to the task, the corresponding value in the matrix is ​​1; if the type of associated data is not related to the task, the corresponding value in the matrix is ​​0.

[0032] Furthermore, the step of inputting the spliced ​​features of different tasks into the identity-guided estimation module to obtain estimated wind speed, pressure, core wind radius, and outer core wind radius includes:

[0033] The identity-guided estimation module includes four identity attention estimation heads, each of which includes a self-attention mechanism, three fully connected layers, and two activation functions.

[0034] The spliced ​​features from different tasks are input into four identity attention estimation heads to obtain estimated wind speed, pressure, core wind circle radius, and outer core wind circle radius.

[0035] Furthermore, identity constraint loss is used to supervise shared identity tokens and special identity tokens for different tasks. The identity constraint loss is the difference between the sample similarity matrix of the identity token in each batch of data and the sample similarity matrix of the true label value in each batch of data.

[0036] Compared with the prior art, the significant advantages of this invention are:

[0037] 1. Task-Dependent Flow Learning: Utilizing the Holland prior wind field model of typhoons, this approach decouples and models the dependencies between multiple tasks. First, it learns special identity tokens representing typhoon intensity—wind speed and pressure—by developing encoded features of the data. Then, it uses a proposed prior approximator to model the Holland model, converting these special identity tokens into special identity tokens for the kernel and outer kernel wind circle estimation tasks, respectively. Task-dependent flow learning makes the model's multi-task modeling and learning process more explicit, avoiding the conceptual bias problem that arises from learning multiple tasks simultaneously.

[0038] 2. Linked Information Bridge: Based on the associated data of typhoons and the task associations they imply, a predefined dark knowledge graph is used. A Gaussian mixture model is applied to the weighted initial identity tokens obtained from the dark knowledge graph to obtain shared identity tokens for multiple tasks. This shared identity token acts as an information bridge, linking the model's multimodal inputs and multi-task outputs through the physical semantic information learned from the latent association graph.

[0039] 3. Identity-oriented multi-task estimation: Existing technologies rely solely on data-driven black-box feature learning to achieve simultaneous estimation of multiple tasks. The method of this invention learns prior-driven shared and specific identity tokens for tasks through task-dependent flow learning and associated information bridges, and incorporates identity information from different tasks through global attention, guiding the model to estimate multiple typhoon attributes more accurately and effectively. Attached Figure Description

[0040] Figure 1 This is a flowchart of the typhoon multi-task estimation method based on prior identity guidance and physical invariance of the present invention;

[0041] Figure 2 This is a schematic diagram of the network estimation model of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] like Figure 1 As shown, this invention provides a typhoon multi-task estimation method based on prior identity guidance and physical invariance, including:

[0044] Step S1, Identity Token Initialization: Multimodal features (initial features) are extracted from temporal multi-channel infrared data and associated data using an encoding network. The mean and variance of these multimodal features are then calculated, assuming a normal distribution. Distribution features of a specified dimension are sampled from these features as the initial identity token. The associated data includes estimated historical (12 hours prior to the estimation time) typhoon fullness, concentration index, energy index, and width index. Referring to the calculation method for typhoon fullness, the concentration index = maximum wind speed / core wind circle, the energy index = pressure / maximum wind speed, and the width index = pressure / outer core wind circle.

[0045] This embodiment is based on a priori identity-guided and physically invariant typhoon multi-task estimation method. It involves training and constructing an estimation network model, and then using the trained estimation network model to process real-time typhoon data to estimate the wind speed, pressure, and inner and outer core wind circles at the current moment.

[0046] The estimation network model constructed in this embodiment is as follows: Figure 2 As shown, it includes an identity token initialization module, a task-dependent flow learning module, an association information bridge module, and an identity-guided estimation module. The identity token initialization module consists of two parts: an encoding network and normal distribution sampling; the task-dependent flow learning module consists of two parts: an encoder and two prior approximators; the association information bridge module consists of two parts: an association encoder and a Gaussian mixture model; and the identity-guided estimation module consists of multiple identity attention estimation heads.

[0047] Specifically, the encoding network consists of two deep convolutional networks (VGG), an associative encoder, and a self-attention long short-term memory network (LSTM). Multi-channel satellite brightness temperature data is input into the encoding network to extract initial features related to the typhoon. In one specific embodiment, the deep convolutional network has 13 layers, the self-attention LSTM network has 512 input channels, a temporal step size of 2, and the initial identity token dimension after normal distribution sampling is set to 16.

[0048] Step S2, Task-dependent Stream Learning: Encode the development data (the intensity level and duration of the typhoon at the previous moment) using the development encoder, transform the initial identity tokens with the obtained development features to obtain special identity tokens for the two tasks of pressure and wind speed; Model the prior model (Holland) of the typhoon through the prior approximator to realize the transformation of identity tokens, thereby obtaining special identity tokens for the two tasks of inner and outer core wind circles.

[0049] Specifically, the development encoder consists of two identical multilayer perceptrons (MLPs). The typhoon's previous intensity level and duration data are input into the development encoder to extract the typhoon's development characteristics. In one specific embodiment, the two MLPs have identical structures but different parameters. Each MLP includes two fully connected layers. The typhoon's previous intensity level and duration data are input into the two identical MLPs respectively to obtain special identity tokens for the two estimation tasks of typhoon wind speed and pressure. The two fully connected layers are used to learn two outputs from the input data. Either output is used as the distribution adjustment feature gamma, and the other as the distribution adjustment feature bata. gamma and bata are used as coefficients and biases, respectively, to transform the initial identity token, i.e., gamma × initial identity token + bata, thereby obtaining the special identity tokens for the two estimation tasks of typhoon wind speed and pressure (i.e.,...). Figure 2 The tokens are v and p. The two multilayer perceptrons have the same structure, with each of the two fully connected layers containing 2 neurons. The weights and biases of the two fully connected layers are defined as floating-point variables. It should be noted that all the fully connected layers are followed by a ReLU activation function.

[0050] The prior approximator consists of two graph convolutions and a self-attention long short-term memory network (LSTM). The wind speed's unique identifier is input into the prior approximator to extract the unique identifier of the core wind circle (i.e.,...). Figure 2 The token ri in the middle), the special identity token of pressure is input into the prior approximator to extract the special identity token of the outer core wind circle (i.e. Figure 2 The token ro in the LSTM. In a specific embodiment, the prior approximator consists of two graph convolutions, a self-attention long short-term memory network, and learnable parameters for fitting unknown values ​​in the prior model. The two graph convolutional layers are used for encoding the input and its associated graph in the forget gate and output gate of the LSTM, respectively, and the number of graph nodes is set to 4. The original temporal cyclic fusion of the self-attention long short-term memory network is used for the exponential operation in fitting the prior model, and the step size is predefined as 20. If the interpolation result of the iterative operation is lower than the predefined threshold of 0.001, the fitting loop process is interrupted.

[0051] It should be noted that there is a prior model between the intensity and the wind circle radius estimated in this invention. Therefore, by using a prior approximator to transform the special identity tokens corresponding to different tasks, the dependency decoupling and modeling of multi-task learning are realized, which can avoid the concept offset problem that exists in multi-task simultaneous learning, that is, the problem that multiple association mapping functions are difficult to learn effectively at the same time.

[0052] Step S3, Association Information Bridge: Based on the prior association between tasks, a dark knowledge graph is predefined, and the association encoder is used to learn the graph. The resulting weighted weights are used to weight the mixture Gaussian distribution model of the initial identity token, and finally the shared identity token is obtained.

[0053] Specifically, the dark knowledge graph is initially predefined as an N×N matrix A, which can be obtained based on the definition of associated data. Its horizontal axis index represents the associated data, and the vertical axis represents the task related to that data. For example, typhoon fullness (index-x=0) is related to the inner and outer core wind circles of a typhoon (index-y=2 / 3). Therefore, A... 02 and A 03 The value at irrelevant indices is 0, and the value at irrelevant indices is 1. The association encoder consists of two graph convolutions. A predefined dark knowledge graph is input into the association encoder to extract the weighting weights used for the weighted Gaussian mixture model, namely gamma_1 to gamma_k and bata_1 to bata_k, where the former is the weighting coefficient and the latter is the bias. The initial identity token is fitted with k single Gaussian distributions, and then the k weights output by the association encoder are weighted and summed to obtain the Gaussian mixture model. The distribution features of the weighted Gaussian mixture model are used as the shared identity token. The Gaussian mixture model consists of multiple sets of learnable parameters conforming to a Gaussian distribution. In a specific embodiment, the number of components k in the Gaussian mixture model is set to 4, the number of graph nodes in the graph convolution operation is 4, and all parameters in the model are defined as floating-point variables.

[0054] It should be noted that this embodiment utilizes latent association graph learning to guide the constraint and modeling of shared feature distributions, which can maximize the semantic information representing the input and output ends, thereby alleviating the covariance and label offset problems existing in the input and output data. In addition, the task-specific and shared identity tokens obtained in steps S2 and S3 are supervised using identity constraint loss. Specifically, the identity constraint loss is calculated by the difference between the sample similarity matrix of the identity token in each batch of data and the sample similarity matrix of the true label value in each batch of data. The sample similarity matrix of the identity token and the true label value in each batch of data is calculated using matrix multiplication, with the following formula:

[0055]

[0056] in, This represents the sample similarity matrix of the identity token or tag's true value within each batch of data, with a shape of (batch_size, batch_size). This represents the data for each batch of identity tokens or authenticity tags, with a shape of (batch_size, n). This indicates transpose.

[0057] Step S4, Identity-oriented Typhoon Estimation: The shared identity token, special identity tokens specific to different tasks, and the initial multimodal features are concatenated and input into the four identity attention estimation heads of the identity-oriented estimation module to obtain multiple estimation results for wind speed, pressure, and inner and outer core wind circles.

[0058] Specifically, the identity tokens output by the task-dependent flow learning module and the associated information bridge module, both based on a self-attention mechanism, are used to fuse the initial features of the network and feed them into the corresponding estimation heads to simultaneously estimate the typhoon's wind speed, pressure, and the radii of its inner and outer core wind circles. This method, guided by a physically invariant identity distribution, yields robust multi-attribute estimations of typhoons, which is extremely important for disaster prevention applications related to typhoons and the natural disasters they cause.

[0059] The identity attention estimation head consists of a self-attention mechanism, three fully connected layers, and two activation functions. It is used to output four regression values ​​of typhoon wind speed, pressure, and inner and outer core wind circle radii. The number of neurons in the three fully connected layers is set to 128, 64, and 1, respectively. The number of attention heads is set to 4, and the feature dimension is set to 16.

[0060] The training process for the estimation network model in this invention is as follows:

[0061] First, the dataset was preprocessed. The dataset was divided into three parts: a training set, a validation set, and a test set. The training set contained data on 235 typhoons that occurred between 2015 and 2021, within the latitude and longitude range of 60°S-60°N and 80°E-160°W. The validation and test sets contained data on 36 typhoons that occurred in 2022 and 32 typhoons that occurred in 2023, respectively. The one-dimensional data (correlation and development data) and two-dimensional data (time-series multi-channel infrared data) in the dataset were normalized to make the data more conducive to model training. The training set data was then rotated clockwise by 45°, 90°, 135°, 180°, 225°, and 270° for data augmentation.

[0062] During training:

[0063] The parameters of the convolutional and fully connected layers in the model were initialized with random weights, and a total of 200 complete training iterations were performed. The learning rate of the model optimizer Adam was set to 0.0001, and the data size of each batch was 16.

[0064] Input some one-dimensional data and two-dimensional time-series multi-channel satellite data of the typhoon into the coding network; input the duration of existence and the intensity level 12 hours before the estimated time into the task-dependent flow learning module; input the typhoon fullness, concentration index, energy index and width index into the association information bridge module.

[0065] The encoding network, task-dependent flow learning module, and associated information bridge module respectively extract features and model the distribution constraints of the input time-series multi-channel infrared data and associated data. Finally, the shared representation, shared identity token, and specific identity token applicable to different estimation tasks are fused and sent to the estimation module, and the estimated values ​​of typhoon wind speed, pressure, and inner and outer core wind circles are output.

[0066] The error between the estimated typhoon and the actual typhoon value is calculated to obtain four L1 losses. These losses are then summed, and the total error is fed back into the network. The network parameters are adjusted by minimizing the loss function using the Adam algorithm to optimize the estimation performance.

[0067] Train the model until it converges, and obtain the best-performing model for typhoon estimation.

[0068] The prior knowledge of this invention comprises two parts: first, the prior model Holland, which is approximated by a prior approximator; and second, a dark knowledge graph derived from associated data (learning prior associations between different tasks). This invention utilizes the learning of these two types of prior knowledge to obtain task-shared and task-specific prior identities (tokens, representing learned physical invariances), which are then used to guide the estimation of different tasks.

[0069] This invention also provides experimental data, comparing the technical solution of this invention with some traditional technical solutions (ADT, MTCSWA) and other deep learning methods (STIA, NS, TC-MTLNet and DeepTCNet). The performance metrics used for comparison are mean absolute error (MAE) and root mean square error (RMSE). The comparison results are shown in Table 1.

[0070] Table 1

[0071]

[0072] As can be seen from the comparison results in Table 1, the technical solution of the present invention outperforms traditional technical solutions and other deep learning methods.

[0073] Among them, ADT (Advanced Dvorak Technique) is an objective intensity estimation technique based on infrared satellite imagery, which inherits and improves upon the traditional Dvorak technique. It establishes a regression equation through cloud feature parameters (such as cloud symmetry and the size of the central dense cloud area), transforms cloud image features into intensity indices, and then into sea level pressure or maximum wind speed.

[0074] MTCSWA (Multi-platform Tropical Cyclone Surface Wind Analysis): NOAA (National Oceanic and Atmospheric Administration) operational multi-platform satellite wind field analysis technology that integrates multi-source satellite data (such as scatterometers and microwave radiometers) to estimate surface wind fields.

[0075] STIA (Spatial-Temporal Image Analysis Model): Based on Himawari-8 satellite imagery, this model estimates the intensity of tropical cyclones in the Northwest Pacific using a spatiotemporal analysis approach. It combines the temporal and spatial characteristics of infrared cloud images and extracts intensity indicators through deep learning or image processing techniques.

[0076] NS (Novel Size Estimation Model): A size estimation model based on convolutional neural networks (CNN) that directly learns cyclone-scale features (such as wind circle radius) using geostationary satellite imagery.

[0077] TC-MTLNet (Physics-Augmented Deep Learning): A physics-augmented multi-task deep learning network that simultaneously estimates strength and size. It introduces physical constraints (such as wind pressure relationships) on top of CNNs.

[0078] DeepTCNet (Lightweight Multi-Task Learning): A lightweight multi-task learning model that simultaneously optimizes strength and size estimation through an adaptive loss balancing mechanism. It uses a 2D CNN to process the spatial features of infrared images and transfer learning to adapt to different sea areas.

[0079] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A typhoon multitask estimation method based on prior identity orientation and physical invariance, characterized in that, The typhoon multi-task estimation method based on prior identity orientation and physical invariance comprises: acquiring time-series multi-channel infrared data and associated data; constructing an estimation network model, including an identity token initialization module, a task dependency flow learning module, an associated information bridge module, and an identity orientation estimation module; extracting multi-modal features from the time-series multi-channel infrared data and the associated data by using the identity token initialization module, and obtaining an initial identity token based on the multi-modal features; obtaining special identity tokens of four tasks of pressure, wind speed, inner core wind circle, and outer core wind circle based on development data by using the task dependency flow learning module, wherein the development data is intensity level of a typhoon at a previous time and typhoon existence duration data; obtaining a shared identity token based on the initial identity token by using the associated information bridge module; splicing the shared identity token, the special identity tokens of different tasks, and the multi-modal features to obtain spliced features of different tasks; inputting the spliced features of different tasks into the identity orientation estimation module to obtain estimated wind speed, pressure, inner core wind circle radius, and outer core wind circle radius.

2. The typhoon multitask estimation method based on prior identity orientation and physical invariance according to claim 1, characterized in that, The identity token initialization module comprises an encoding network and a normal distribution sampling part, wherein the encoding network comprises two deep convolutional networks, an associated encoder, and a self-attention long short-term memory network. The encoding network is used to extract multi-modal features from the time-series multi-channel infrared data and the associated data. The normal distribution sampling is used to statistically obtain mean and variance of the multi-modal features, define a normal distribution with the mean and the variance, and sample distribution features of a specified dimension from the normal distribution as the initial identity token. The task dependency flow learning module comprises a development encoder and two prior approximators, the development encoder comprises two identical multi-layer perceptrons, and the prior approximator comprises two graph convolutions and a self-attention long short-term memory network.

3. The typhoon multitask estimation method based on prior identity orientation and physical invariance according to claim 1, characterized in that, The intensity level of the typhoon at the previous time and the typhoon existence duration data are input into the two multi-layer perceptrons respectively, each multi-layer perceptron obtains two distribution adjustment features, the two distribution adjustment features are used as a coefficient and a bias respectively, the initial identity token is transformed, and the two multi-layer perceptrons obtain special identity tokens of wind speed and pressure tasks respectively. The special identity tokens of the wind speed and pressure tasks are input into the prior approximator to extract special identity tokens of inner core wind circle and outer core wind circle tasks. The transformation of the initial identity token is as follows: Special identity token of inner core wind circle or outer core wind circle task = gamma x initial identity token + bata; 4. The typhoon multitask estimation method based on prior identity orientation and physical invariance according to claim 3, characterized in that, wherein gamma and bata are distribution adjustment features. The associated information bridge module comprises an associated encoder, and the associated encoder comprises two graph convolutions. ​ 5. The typhoon multitask estimation method based on prior identity orientation and physical invariance according to claim 1, characterized in that, ​ ​ The dark knowledge graph is predefined according to prior associations between tasks, the dark knowledge graph is input into an association encoder, and k weighted weights are extracted; The initial identity token is fitted by using k single Gaussian distributions, the k single Gaussian distributions are weighted by using the k weighted weights, a mixed Gaussian distribution model is obtained, and distribution features of the weighted mixed Gaussian distribution model are used as the shared identity token.

6. The typhoon multitask estimation method based on prior identity orientation and physical invariance according to claim 5, characterized in that, The predefined dark knowledge graph according to the prior associations between tasks comprises: The category of the association data is used as the horizontal coordinate of the dark knowledge graph, the task related to the association data is used as the vertical coordinate, and the predefined dark knowledge graph is an N*N matrix. If the category of the association data is related to the task, the corresponding value in the matrix is 1, and if the category of the association data is not related to the task, the corresponding value in the matrix is 0.

7. The typhoon multitask estimation method based on prior identity orientation and physical invariance according to claim 1, characterized in that, The spliced features of different tasks are input into an identity orientation estimation module to obtain estimated wind speed, pressure, core wind circle radius and outer core wind circle radius, and the identity orientation estimation module comprises: The identity orientation estimation module comprises four identity attention estimation heads, the identity attention estimation head comprises a self-attention mechanism, three fully connected layers and two activation functions. The spliced features of different tasks are input into the four identity attention estimation heads respectively to obtain estimated wind speed, pressure, core wind circle radius and outer core wind circle radius.

8. The typhoon multitask estimation method based on prior identity orientation and physical invariance according to claim 1, characterized in that, Identity constraint loss is used to supervise the shared identity token and the special identity token of different tasks, and the identity constraint loss is the difference between a sample similarity matrix of the identity token in each batch of data and a sample similarity matrix of a label real value in each batch of data.

Images