An intelligent recognition method for millimeter-wave cloud radar echoes based on the KAN network

Abstract

The present invention discloses an intelligent recognition method for millimeter-wave cloud radar echoes based on the KAN network, comprising the following steps: (1) automatic label generation; (2) the REC-KAN-MLP network combining eigenvalue and fuzzy logic algorithm; (3) the model realizes real-time filtering after radar scanning through end-to-end deep learning; the present invention can effectively identify meteorological echoes and ground clutter in the PPI without the radar echo power spectrum.

Description

Technical Field

[0001] The present invention relates to the technical field of millimeter-wave cloud radar echo, and specifically relates to an intelligent recognition method for millimeter-wave cloud radar echo based on a KAN network. Background Art

[0002] A PPI (Plan Position Indicator) radar is a radar system widely used in meteorological observations and ground object detection. It generates a planar image through the horizontal rotation scanning of radar waves, reflecting the position and intensity of targets on the horizontal plane. The observation results of the PPI radar are presented in a planar distribution centered on the radar station, and echo data in different azimuths can be obtained in real time. These echo data contain a large amount of environmental information, such as precipitation intensity, terrain reflection, etc., but are also easily affected by ground clutter and noise interference, resulting in a decline in data quality and affecting the actual application effect. Traditional PPI radar echo processing technologies mainly rely on power spectra to distinguish meteorological targets and non-meteorological interferences. However, these methods have limitations in accurate recognition in complex environments and cannot effectively cope with non-meteorological interferences such as ground clutter and building reflections.

[0003] Currently, most mature radar echo recognition algorithms are based on the power spectrum data of radar echoes. These filtering algorithms are relatively complex, have a slow processing speed, poor migration ability, and cannot effectively filter the extended products of radar signals, greatly limiting the application effect in actual operations. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide an intelligent recognition method for millimeter-wave cloud radar echo based on a KAN network, so as to solve the problems of slow radar echo recognition speed, low accuracy, and low business value in the prior art.

[0005] Technical Solution: An intelligent recognition method for millimeter-wave cloud radar echo based on a KAN network according to the present invention includes the following steps:

[0006] (1) Using automated label generation technology to preprocess radar observation data, generating labels close to real echo classification, and constructing a label data set;

[0007] (2) Fusing eigenvalue and fuzzy logic algorithms to construct a REC-KAN-MLP network;

[0008] (3) Using the echo classification label data set to train the REC-KAN-MLP network to obtain a radar echo recognition target model;

[0009] (4) Obtaining radar observation data of the research area, using the radar echo recognition target model to perform pixel-by-pixel recognition and classification of radar echoes, and conducting evaluation and analysis of multiple scoring indicators;

[0010] (5) Conduct interpretable analysis on the model output, quantify the contributions of each feature, and construct membership functions.

[0011] Furthermore, step (1) includes the following steps:

[0012] (11) Set thresholds in combination with the characteristics of ground clutter echoes;

[0013] (12) Based on the feature engineering technology of radar echo characteristics, use the initial detection data of the radar to extract secondary data that further characterizes the semantic type of radar echoes for model input;

[0014] (13) Perform morphological processing of erosion and dilation on the labels.

[0015] Furthermore, step (11) is specifically as follows:

[0016] (111) Preliminary classification: When the radar reflectivity factor is a NaN value, it is marked as the background field; when the average radial velocity value is less than 0.2 m / s, it is marked as ground clutter echoes, otherwise it is marked as meteorological echoes;

[0017] (112) Refine the labels according to more conditions:

[0018] a When the average radial velocity value is less than 1, the radar reflectivity factor is less than -5 dBZ, and the velocity spectrum width is greater than 0.4 m / s, it is marked as meteorological echoes, otherwise it is marked as ground clutter echoes;

[0019] b When the average radial velocity value is greater than 0.2 m / s and the radar reflectivity factor is greater than 18 dBZ, it is marked as meteorological echoes, otherwise it is marked as ground clutter echoes; when the average radial velocity value is greater than 0.2 m / s and the signal-to-noise ratio is greater than 18, it is marked as meteorological echoes, otherwise it is marked as ground clutter echoes;

[0020] c When the average radial velocity value is less than 0.2 m / s, when the number of NaN values in the 5×5 kernel window centered on the judgment point does not exceed 2, the maximum value of the average radial velocity values in two directions exceeds 0.1 m / s and the difference in the average radial velocity within the kernel exceeds 0.4 m / s, it is marked as meteorological echoes, otherwise it is marked as ground clutter echoes;

[0021] Furthermore, step (12) is specifically as follows: For PPI-type radar echo data, 11 features are extracted for further identification of ground clutter echoes and meteorological echoes, including 4 original radar data: radar reflectivity factor, average radial velocity V, velocity spectrum width W, signal-to-noise ratio SNR of radar echoes, the polar coordinate radius calculated according to the data array, and 6 features extracted by combining the commonly used fuzzy logic algorithms for radar echo classification: the texture of echo intensity and the sign change along the radial direction , Degree of variation between libraries along the radial direction , Regional average value of radial velocity , Variance , Regional average value of velocity spectrum width .

[0022] Furthermore, in step (13), the morphological processing of erosion and dilation is to remove noise and fill holes.

[0023] Furthermore, in step (2), the overall REC-KAN-MLP network structure is a REC structure. The membership function of the KAN network is introduced for fitting, and the MLP module is used for weight allocation; in the REC-KAN-MLP network structure, the hidden layer neurons are KAN neurons , the output feature of the output layer is 1. The MLP module receives the output probability of KAN, and after linear transformation, it is activated by the sigmoid function, and the output is the probability of the echo type .

[0024] Furthermore, step (3) is specifically as follows: First, each factor is used as an input alone to train its exclusive classification model; then, the purification technique is used to screen out relatively important neurons and remove other neurons; finally, the remaining neurons are used for training again to obtain the final model. In this model, the activation function of the first-layer network is a function of the membership function, and the membership function of the second-layer network is an external function. If multiple activation functions are retained in the same layer network, it is a relationship of equal-weight addition, and the output range after passing through the two-layer network is the probability in [0,1].

[0025] Furthermore, in step (3), during the training process, 1 to 3 factors in 10% of the data are randomly masked each time and not input into the network for training.

[0026] Furthermore, the evaluation and analysis of multiple scoring metrics in step (4) are specifically as follows:

[0027] Recognition rate Accuracy: ;

[0028] Precision: ;

[0029] Recall: ;

[0030] F1 Score: ;

[0031] False alarm rate FAR: ;

[0032] Miss rate miss: ;

[0033] Among them, the calculation method of the recognition rate of ground clutter echoes in the PPI dataset: TP represents the number of targets that are actually ground clutter echoes and are recognized as ground clutter echoes; FP represents the number of targets that are actually non-ground clutter echoes and are recognized as ground clutter echoes; TN represents the number of targets that are actually non-ground clutter echoes and are recognized as non-ground clutter echoes; FN represents the number of targets that are actually ground clutter echoes and are recognized as non-ground clutter echoes.

[0034] Furthermore, step (5) is specifically as follows: Use the KAN network to fit the membership function to explain the inference process of the model; Use SHAP to rank the contribution degrees of the input factors of the model. Description of the Drawings

[0035] Figure 1 is the flowchart of the present invention;

[0036] Figure 2 is the technical flowchart of the PPI automatic label generation of the present invention;

[0037] Figure 3 is the network structure diagram of the REC-KAN-MLP of the present invention;

[0038] Figure 4 is the schematic diagram of the overall filtering score effect of the radar echo of the present invention;

[0039] Figure 5 is the schematic diagram of the ranking of the importance of each factor (reflectivity, radial velocity, extracted feature values, etc.) of the model by the SHAP method of the present invention. Detailed Embodiments

[0040] As Figure 1 shown, the embodiment of the present invention provides a millimeter-wave cloud radar echo intelligent recognition method based on a KAN network, including the following steps:

[0041] Step S100, use the automatic label generation technology to preprocess the radar observation data, generate labels close to the real echo classification, and thus construct a training dataset. The specific preprocessing process: perform standardization processing on the radar data; According to the detection needs, classify the radar echo data according to the PPI type for identifying meteorological echoes and ground clutter. As Figure 2 shown, the specific steps are as follows:

[0042] S101 Set a threshold in combination with the characteristics of ground clutter echoes:

[0043] (S1) Preliminary classification: When the radar reflectivity factor is a NaN value, it is marked as the background field; When the average radial velocity value is less than 0.2 m / s, it is marked as a ground clutter echo, otherwise it is marked as a meteorological echo;

[0044] (S2)Refine the labels according to more conditions: When the average radial velocity value is less than 1 m / s, the radar reflectivity factor is less than -5 dBZ, and the velocity spectrum width is greater than 0.4 m / s, it is marked as meteorological echo; otherwise, it is marked as ground clutter echo. When the average radial velocity value is greater than 0.2 m / s and the radar reflectivity factor is greater than 18 dBZ, it is marked as meteorological echo; otherwise, it is marked as ground clutter echo. When the average radial velocity value is greater than 0.2 m / s and the signal-to-noise ratio is greater than 18, it is marked as meteorological echo; otherwise, it is marked as ground clutter echo. When the average radial velocity value is less than 0.2 m / s, if the number of NaN values in the 5×5 kernel window centered on the judgment point does not exceed 2, the maximum value of the average radial velocity values in two directions exceeds 0.1 m / s, and the difference in the average radial velocity within the kernel exceeds 0.4 m / s, it is marked as meteorological echo; otherwise, it is marked as ground clutter echo.

[0045] (S3)Morphological processing of erosion and dilation: Erosion operation means eroding the edges of the image. That is, in an image represented by 0 / 1, a rectangular kernel with a value of 1 is set. For each pixel, placed at the center of the kernel, if all the parts of the image covered by the kernel are 1, then this pixel is 1; otherwise, it is 0. Contrary to the erosion operation, the dilation operation means dilating the contour of the image. That is, in an image represented by 0 / 1, a rectangular kernel with a value of 1 is set. For each pixel, placed at the center of the kernel, if all the parts of the image covered by the kernel are 0, then this pixel is 0; otherwise, it is 1.

[0046] Since the filtering effect on single-point ground clutter echo is still weak after the above operations, erosion and dilation operations are mainly used to remove ground clutter echo noise. The meteorological echo array marked as 0 is subjected to 2 erosion operations and 1 dilation operation through a 3×3 structuring element.

[0047] Step S200, Feature engineering technology based on the characteristics of radar echo. Using the initial detection data of the radar, extract secondary data that can further characterize the semantic type of radar echo for model input. The specific features extracted by feature engineering and the calculation methods are as follows:

[0048] PPI feature extraction: For PPI-type radar echo data, a total of 11 features are extracted for better identification of ground clutter echo and meteorological echo, including 4 features in the original radar data: radar reflectivity factor Z, average radial velocity V, velocity spectrum width W, signal-to-noise ratio SNR of the radar echo, the polar coordinate radius calculated according to the data array, and 6 features extracted by combining the fuzzy logic algorithm commonly used in radar echo classification. The specific introduction is as follows:

[0049] Polar coordinate radius: In the polar coordinate system, the radius not only represents the positional relationship relative to the radar site but also reflects the density of coordinate points. There are significant differences in echo judgment corresponding to different densities.

[0050] Texture of echo intensity : A feature extracted from the echo intensity, used to represent the variance between adjacent range gates of the reflectivity factor, mainly reflecting the local variation of the reflectivity factor within a certain range. The definition formula is:

[0051] ;

[0052] Sign change along the radial direction , Degree of change between bins along the radial direction : A feature extracted from the echo intensity, indicating the consistency of the echo intensity variation along the radial direction, indicating the sign change of the echo intensity variation along the radial direction, reflecting the variation of the reflectivity factor along the radial direction. The definition formula is:

[0053] ;

[0054] ;

[0055] Regional average value of radial velocity , Variance : A feature extracted from the radial velocity, representing the radial velocity values after median filtering within a certain range, representing the variance of the radial velocity. The definition formula is:

[0056] ;

[0057] ;

[0058] Regional average value of velocity spectrum width : A feature extracted from the velocity spectrum width, representing the spectrum width values after median filtering within a certain range. The definition formula is:

[0059] ;

[0060] Among them, represents the range of the number of calculation points defined in the range direction, represents the range of the number of calculation points defined in the azimuth direction.

[0061] Step S300, use the echo classification label dataset to train the KAN network deep learning model to obtain a radar echo recognition target model; as Figure 3As shown, the overall structure of the "REC-KAN-MLP" network is the REC framework. The membership function is fitted by the KAN network, and the weight assignment is fitted by the MLP module. The structure of the KAN network is as follows: the number of input layers is equal to the number of features; the number of hidden layers and the number of hidden layer features are adjustable parameters, and the hidden layer neurons are KAN neurons , and the output feature of the output layer is 1. The MLP module receives the output probability of the KAN, and after linear transformation, it is activated by the sigmoid function, and the output is the probability of the echo type .

[0062] In step S400, each factor is first used as an input to train its exclusive classification model; then, the purification technology is used to screen out the more important neurons and remove the other neurons; finally, the remaining neurons are used to train again to obtain the final model. In this model, the activation function of the first-layer network is the function of the membership function, and the membership function of the second-layer network is the external function. If multiple activation functions are retained in the same layer network, they are in an equal-weight addition relationship. The output range after passing through the two-layer network is in [0,1]. During the training process, 1 to 3 factors in 10% of the data are randomly masked each time and not input into the network for training

[0063] In step S500, the radar observation data of the research area is obtained, and the KAN network deep learning model is used to perform pixel-by-pixel recognition and classification of the radar echoes, and multiple scoring indicators are evaluated and analyzed. Specifically as follows: the processed data is standardized to obtain a data set, which is divided into a training set, a validation set, and a test set according to the ratio of 8:1:1. The KAN network is trained to obtain a target model; the evaluation criteria are as follows

[0064] Recognition rate Accuracy: ;

[0065] Precision: ;

[0066] Recall: ;

[0067] F1 Score: ;

[0068] False alarm rate FAR: ;

[0069] Miss rate miss: ;

[0070] Among them, the calculation method of the recognition rate of ground clutter echoes in the PPI dataset: TP represents the number of targets that are actually ground clutter echoes and are recognized as ground clutter echoes; FP represents the number of targets that are actually non-ground clutter echoes and are recognized as ground clutter echoes; TN represents the number of targets that are actually non-ground clutter echoes and are recognized as non-ground clutter echoes; FN represents the number of targets that are actually ground clutter echoes and are recognized as non-ground clutter echoes.

[0071] The recognition ability of the intelligent radar echo recognition technology is evaluated using multiple scoring metrics, and networks with different parameter magnitudes are compared. The parameter quantity relationship is big model > small model > mini. For example, KAN_big is the model with the largest parameter quantity of "REC-KAN-MLP", and MLP_mini is the model with the smallest parameter quantity under the structure of "REC-MLP-MLP". The overall evaluation of the models is as Figure 4 shown.

[0072] Step S600, use the KAN network to fit the membership function to explain the inference process of the model. The specific steps are as follows: First, train the KAN network according to the data. When the training is completed, remove the weight functions of the edges with too low weight influence; then, analyze the remaining activation functions obtained from the training and compare them with the known function library. Finally, replace the activation functions obtained from the training with specific known functions to achieve the interpretability of the network. This process can be summarized as "sparsification - visualization - pruning - symbolization".

[0073] Step S700, use SHAP to rank the contribution degrees of the input factors of the model to explain the importance of the model factors. In the ground clutter echo recognition network, the SHAP values of each factor are as Figure 5 shown. It can be seen that the average radial velocity V plays an absolute dominant role. Although its average SHAP value is near 0, from the relationship between the SHAP value and the magnitude of V, it can be seen that when the average radial velocity is very small, the SHAP value is also very small, which is exactly the characteristic of ground clutter echoes; while when V is larger, the corresponding SHAP value is also larger, that is, a large Doppler velocity generally corresponds to meteorological echoes.

Claims

1. An intelligent recognition method for millimeter-wave cloud radar echo based on KAN network, characterized in that It includes the following steps: (1) Preprocess the radar observation data using automated label generation technology to generate labels close to real echo classification, and construct a label dataset; it includes the following steps: (11) Set thresholds in combination with the characteristics of ground clutter echoes; specifically as follows: (111) Preliminary classification: When the radar reflectivity factor is a NaN value, it is marked as the background field; when the average radial velocity value is less than 0.2 m / s, it is marked as ground clutter echo, otherwise it is marked as meteorological echo; (112) Refine the labels according to more conditions: a When the average radial velocity value is less than 1, the radar reflectivity factor is less than -5 dBZ, and the velocity spectrum width is greater than 0.4 m / s, it is marked as meteorological echo, otherwise it is marked as ground clutter echo; b When the average radial velocity value is greater than 0.2 m / s and the radar reflectivity factor is greater than 18 dBZ, it is marked as meteorological echo, otherwise it is marked as ground clutter echo; when the average radial velocity value is greater than 0.2 m / s and the signal-to-noise ratio is greater than 18, it is marked as meteorological echo, otherwise it is marked as ground clutter echo; c When the average radial velocity value is less than 0.2 m / s, when the number of NaN values in the 5×5 kernel window centered on the judgment point does not exceed 2, the maximum value of the average radial velocity values in two directions exceeds 0.1 m / s and the difference in the average radial velocity within the kernel exceeds 0.4 m / s, it is marked as meteorological echo, otherwise it is marked as ground clutter echo; (12)Feature engineering technology based on the characteristics of radar echoes, which uses the initial detection data of the radar to extract secondary data that further characterizes the semantic types of radar echoes for input into the REC-KAN-MLP network model; specifically as follows: for PPI-type radar echo data, 11 features are extracted for further identification of ground echoes and meteorological echoes, including 4 original radar data: radar reflectivity factor, average radial velocity V, velocity spectrum width W, signal-to-noise ratio SNR of the radar echo, polar coordinate radius calculated from the data array, and 6 features extracted using the fuzzy logic algorithm commonly used in radar echo classification: polar coordinate radius, texture of echo intensity , sign change along the radial direction , degree of variation between bins along the radial direction , regional average of radial velocity , variance , regional average of velocity spectrum width ; (13) Perform morphological processing of erosion and dilation on the labels; (2)Integrate the eigenvalue and fuzzy logic algorithm to construct the REC-KAN-MLP network; the overall structure of the REC-KAN-MLP network is the REC structure, introduce the membership function of the KAN network for fitting, and use the MLP module for weight allocation; in the REC-KAN-MLP network structure, the hidden layer neurons are KAN neurons , the output feature of the output layer is 1, the MLP module receives the output probability of KAN, and after linear transformation, it is activated by the sigmoid function, and the output is the probability of the echo type ; (3) Use the echo classification label dataset to train the REC-KAN-MLP network to obtain a radar echo recognition target model; specifically as follows: First, each factor is used as an input separately to train its exclusive KAN classification model; then, the purification technology is used to screen out relatively important neurons and remove other neurons; finally, the remaining neurons are used for training again to obtain the final radar echo recognition target model. In this model, the activation function of the first-layer network is a function of the membership function, and the membership function of the second-layer network is an external function. If multiple activation functions are retained in the same layer network, it is a relationship of equal-weight addition. The output range after passing through the two-layer network is the probability within [0,1]; (4) Obtain the radar observation data of the study area, use the radar echo recognition target model to perform pixel-by-pixel recognition and classification of radar echoes, and conduct evaluation and analysis of multiple scoring metrics; the evaluation and analysis of multiple scoring metrics are specifically as follows: Recognition rate Accuracy: ; Precision: ; Recall rate: ; F1 Score: ; False Alarm Rate FAR: ; False negative rate miss: ; Among them, the calculation method of the recognition rate of ground clutter echoes in the PPI dataset: TP represents the number of targets that are actually ground clutter echoes and are recognized as ground clutter echoes; FP represents the number of targets that are actually non-ground clutter echoes and are recognized as ground clutter echoes; TN represents the number of targets that are actually non-ground clutter echoes and are recognized as non-ground clutter echoes; FN represents the number of targets that are actually ground clutter echoes and are recognized as non-ground clutter echoes; (5) Conduct interpretable analysis on the output of the radar echo recognition target model, quantify the contributions of each feature and construct a membership function.

2. The intelligent recognition method for millimeter-wave cloud radar echo based on the KAN network according to claim 1, wherein, In step (13), the morphological processing of erosion and dilation is to remove noise and fill holes.

3. An intelligent recognition method for millimeter-wave cloud radar echo based on KAN network according to claim 1, characterized in that, In step (3), during the training process, 1 to 3 factors in 10% of the data are randomly masked each time and not input into the network for training.

4. An intelligent recognition method for millimeter-wave cloud radar echo based on KAN network according to claim 1, characterized in that, Step (5) is specifically as follows: Use the KAN network to fit the membership function to explain the inference process of the model; Use SHAP to rank the contribution degrees of the input factors of the model.

Images