Millimeter wave cloud radar echo intelligent identification method based on KAN network

Abstract

The invention discloses a millimeter wave cloud radar echo intelligent identification method based on a KAN network. The method comprises the following steps: (1) automatic tag generation; (2) an REC-KAN-MLP network is established in combination with characteristic values and a fuzzy logic algorithm; (3) the model realizes real-time filtering after radar scanning through end-to-end deep learning; according to the method, meteorological echoes and ground clutters in the PPI can be effectively identified without a radar echo power spectrum.

Description

Technical Field

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

[0002] PPI (Plan Position Indicator) radar is a radar system widely used in meteorological observation and ground object detection. It generates a plane image through horizontal rotation scanning of radar waves, reflecting the position and intensity of the target on the horizontal plane. The observation results of PPI radar are distributed in a plane with the radar station as the center, and echo data in different directions can be obtained in real time. These echo data contain a lot of environmental information, such as precipitation intensity, terrain reflection, etc., but they are also easily interfered by ground clutter and noise, resulting in a decrease in data quality and affecting the actual application effect. Traditional PPI radar echo processing technology mainly relies on power spectrum to distinguish meteorological targets from non-meteorological interference. However, these methods have limitations in accurate identification in complex environments and cannot effectively deal with non-meteorological interference such as ground clutter and building reflection.

[0003] Currently, the more mature radar echo recognition algorithms are mostly based on the power spectrum data of radar echoes. These filtering algorithms are relatively complex, slow in processing speed, poor in mobility, and cannot effectively filter the extended products of radar signals, which greatly limits the application effect in actual business. Summary of the invention

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

[0005] Technical solution: The method for intelligently identifying millimeter-wave cloud radar echoes based on a KAN network described in the present invention comprises the following steps: (1) Using automatic label generation technology to preprocess radar observation data, generate labels close to the actual echo classification, and construct a label dataset; (2) Integrate eigenvalue and fuzzy logic algorithms to construct the REC-KAN-MLP network; (3) The REC-KAN-MLP network is trained using the echo classification label dataset to obtain a radar echo target recognition model; (4) Obtain radar observation data in the study area, use the radar echo target recognition model to identify and classify the radar echo pixel by pixel, and evaluate and analyze multiple scoring indicators; (5) Perform interpretable analysis on the model output, quantify the contribution of each feature and construct membership functions.

[0006] Furthermore, step (1) includes the following steps: (11) Set the threshold based on the characteristics of the ground object echo; (12) Feature engineering technology based on radar echo characteristics uses the initial detection data of the radar to extract secondary data that further characterizes the semantics of the radar echo type for model input; (13) Perform morphological processing of corrosion and expansion on the labels.

[0007] Furthermore, step (11) is specifically as follows: (111) Preliminary classification: When the radar reflectivity factor is NaN, it is marked as background field; when the average radial velocity value is less than 0.2m / s, it is marked as ground echo, otherwise it is marked as meteorological echo; (112) Refine the label based on more conditions: a When the average radial velocity value is less than 1, the radar reflectivity factor is less than -5dBZ, and the velocity spectrum width is greater than 0.4m / s, it is marked as a weather echo, otherwise it is marked as a ground object echo; b. When the average radial velocity value is greater than 0.2m / s and the radar reflectivity factor is greater than 18dBZ, it is marked as a weather echo, otherwise it is marked as a ground object echo; when the average radial velocity value is greater than 0.2m / s and the signal-to-noise ratio is greater than 18, it is marked as a weather echo, otherwise it is marked as a ground object echo; c. When the average radial velocity value is less than 0.2 m / s, when the NaN value in the 5×5 kernel window centered on the judgment point does not exceed 2, the maximum value of the average radial velocity in the two directions exceeds 0.1 m / s and the difference of the average radial velocity in the kernel exceeds 0.4 m / s, it is marked as a meteorological echo, otherwise it is marked as a ground object echo; Furthermore, step (12) is as follows: for PPI type radar echo data, 11 features are extracted for further identification of ground object echo and weather echo, including 4 original radar data: radar reflectivity factor, average radial velocity V, velocity spectrum width W, signal-to-noise ratio SNR of radar echo, polar coordinate radius calculated from data array, and 6 features extracted by combining fuzzy logic algorithm commonly used in radar echo classification: texture of echo intensity , along the radial direction , the degree of variation between reservoirs along the radial direction , the regional average of the radial velocity ,variance , the regional average of the velocity spectrum width .

[0008] Furthermore, in step (13), the morphological processing of corrosion and dilation is performed to remove noise and fill holes.

[0009] Furthermore, in step (2), the overall REC-KAN-MLP network structure is a REC structure, the KAN network membership function is introduced for fitting, and the MLP module is used for weight distribution; 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 KAN output probability, and after linear change, it is activated by the sigmoid function, and the output is the probability of the echo type .

[0010] Furthermore, step (3) is as follows: first, each factor is used as input to train its own classification model; then, purification technology is used to screen the more important neurons and remove other neurons; finally, the retained neurons are used for training again to obtain the final model. In this model, the activation function of the first layer of the network is a function of the membership function, and the membership function of the second layer of the network is an external function. If multiple layers of activation functions are retained in the same layer of the network, they are in an equal-weighted addition relationship, and the probability that the output after passing through the two layers of the network is in the range of [0,1].

[0011] 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.

[0012] Furthermore, the evaluation and analysis of multiple scoring indicators in step (4) are as follows: Recognition rate Accuracy: ; Precision: ; Recall: ; F1 Score: ; False alarm rate FAR: ; Miss rate: ; Among them, the calculation method of the ground object echo recognition rate in the PPI dataset is as follows: TP represents the number of targets that are actually ground object echoes and are recognized as ground object echoes; FP represents the number of targets that are actually non-ground object echoes and are recognized as ground object echoes; TN represents the number of targets that are actually non-ground object echoes and are recognized as non-ground object echoes; FN represents the number of targets that are actually ground object echoes and are recognized as non-ground object echoes.

[0013] Furthermore, step (5) is as follows: fitting the membership function using the KAN network to explain the reasoning process of the model; and ranking the contribution of the input factors of the model using SHAP. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flow chart of the present invention; Figure 2 It is a flow chart of the PPI automated label generation technology of the present invention; Figure 3 It is a REC-KAN-MLP network structure diagram of the present invention; Figure 4 It is a schematic diagram of the overall scoring effect of filtering of radar echoes of the present invention; Figure 5 It is a schematic diagram of ranking the importance of each model factor (reflectivity, radial velocity, extracted eigenvalues, etc.) using the SHAP method of the present invention. DETAILED DESCRIPTION

[0015] like Figure 1 As shown, an embodiment of the present invention provides a millimeter wave cloud radar echo intelligent recognition method based on a KAN network, comprising the following steps: Step S100, using automatic label generation technology to pre-process radar observation data, generate labels close to the real echo classification, and thus construct a training data set. Specific pre-processing process: standardize radar data; according to detection needs, radar echo data is classified by PPI type to identify meteorological echoes and ground clutter. Figure 2 As shown, the specific steps are as follows: S101 sets the threshold based on the characteristics of ground object echo: (S1) Preliminary classification: When the radar reflectivity factor is NaN, it is marked as background field; when the average radial velocity value is less than 0.2m / s, it is marked as ground echo, otherwise it is marked as weather echo; (S2) Refine the label according to more conditions: when the average radial velocity value is less than 1m / s, the radar reflectivity factor is less than -5dBZ, and the velocity spectrum width is greater than 0.4m / s, it is marked as a weather echo, otherwise it is marked as a ground object echo; when the average radial velocity value is greater than 0.2m / s and the radar reflectivity factor is greater than 18dBZ, it is marked as a weather echo, otherwise it is marked as a ground object echo; when the average radial velocity value is greater than 0.2m / s and the signal-to-noise ratio is greater than 18, it is marked as a weather echo, otherwise it is marked as a ground object echo; when the average radial velocity value is less than 0.2m / s, when the NaN value in the 5×5 kernel window centered on the judgment point does not exceed 2, the maximum value of the average radial velocity in two directions exceeds 0.1m / s and the average radial velocity difference in the kernel exceeds 0.4m / s, it is marked as a weather echo, otherwise it is marked as a ground object echo; (S3) Morphological processing of erosion and dilation: The erosion operation refers to the erosion of the edge of the image, that is, in an image represented by 0 / 1, a rectangular kernel with a value of 1 is set, and for each pixel, it is placed at the center of the kernel. If the kernel and the covered image part are all 1, then the pixel point is 1, otherwise it is 0. In contrast to the erosion operation, the dilation operation refers to the dilation of the outline of the image, that is, in an image represented by 0 / 1, a rectangular kernel with a value of 1 is set, and for each pixel, it is placed at the center of the kernel. If the kernel and the covered image part are all 0, then the pixel point is 0, otherwise it is 1.

[0016] Since the filtering effect of single-point ground object echo is still weak after the above operations, corrosion and expansion operations are mainly used to remove ground object echo noise. The meteorological echo array marked as 0 is subjected to two corrosion operations and one expansion operation through the 3×3 structured element.

[0017] Step S200, based on the feature engineering technology of radar echo characteristics, the initial detection data of the radar is used to extract secondary data that can further characterize the semantics of the radar echo type for model input; the specific features extracted by the feature engineering and the calculation method are as follows: PPI feature extraction: For PPI type radar echo data, a total of 11 features are extracted to better identify ground object echoes and weather echoes, including radar reflectivity factor Z, average radial velocity V, velocity spectrum width W, radar echo signal-to-noise ratio SNR in 4 original radar data, polar coordinate radius calculated from data array, and 6 features extracted by fuzzy logic algorithm commonly used in radar echo classification. The specific introduction is as follows: Polar coordinate radius: The radius in the polar coordinate system not only expresses the positional relationship relative to the radar site, but also reflects the density of the coordinate points. There are significant differences in the echo judgments corresponding to different densities.

[0018] Echo intensity texture : Features extracted from echo intensity, used to represent the variance between adjacent range gates of reflectivity factor, mainly reflecting the local change of reflectivity factor within a certain range, and the definition is: ; Sign change along the radial direction , the degree of variation between reservoirs along the radial direction : Features extracted from echo intensity, Indicates the consistency of echo intensity change along the radial direction, The sign change of the echo intensity along the radial direction reflects the change of the reflectivity factor along the radial direction. The definition is: ; ; Regional average of radial velocity ,variance : Features extracted from radial velocity, Represents the radial velocity value processed by median filtering within a certain range. represents the variance of radial velocity and is defined as: ; ; Regional average of velocity spectrum width : The feature extracted from the velocity spectrum width represents the spectrum width value processed by median filtering within a certain range. The definition is: ; in, Indicates the range of calculation points defined in the distance direction, Indicates the range of calculation points defined in the azimuth direction.

[0019] Step S300, using the echo classification label data set to train the KAN network deep learning model to obtain a radar echo target recognition model; Figure 3 As shown in the figure, the overall structure of the "REC-KAN-MLP" network is a REC framework, the membership function is fitted by the KAN network, and the weight distribution is fitted by the MLP module. The KAN network structure is: 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 , the output feature of the output layer is 1, the MLP module receives the KAN output probability, and after linear change, it is activated by the sigmoid function, and the output is the probability of the echo type .

[0020] Step S400, first use each factor as input to train its own classification model; then use purification technology to screen more important neurons and remove other neurons; finally use the retained neurons to train again to obtain the final model. In this model, the activation function of the first layer of the network is a function of the membership function, and the membership function of the second layer of the network is an external function. If multiple layers of activation functions are retained in the same layer of the network, they are equally weighted. The output range after two layers of networks is [0,1]. During the training process, 1 to 3 factors in 10% of the data are randomly masked each time and are not input into the network for training.

[0021] Step S500, obtain radar observation data of the study area, use the KAN network deep learning model to identify and classify radar echoes pixel by pixel, and evaluate and analyze multiple scoring indicators. Specifically, the processed data is standardized to obtain a data set, which is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The KAN network is trained to obtain a target model; the evaluation criteria are as follows: Recognition rate Accuracy: ; Precision: ; Recall: ; F1 Score: ; False alarm rate FAR: ; Miss rate: ; Among them, the calculation method of the ground object echo recognition rate in the PPI dataset is as follows: TP represents the number of targets that are actually ground object echoes and are recognized as ground object echoes; FP represents the number of targets that are actually non-ground object echoes and are recognized as ground object echoes; TN represents the number of targets that are actually non-ground object echoes and are recognized as non-ground object echoes; FN represents the number of targets that are actually ground object echoes and are recognized as non-ground object echoes.

[0022] The recognition capability of intelligent radar echo recognition technology is evaluated using multiple scoring indicators. Networks with different parameter magnitudes are used for comparison. The parameter magnitude of big model>small model>mini. For example, KAN_big is the model with the largest parameter magnitude of "REC-KAN-MLP", and MLP_mini is the model with the smallest parameter magnitude under the "REC-MLP-MLP" structure. The overall evaluation of the model is as follows: Figure 4 shown.

[0023] Step S600, use the KAN network to fit the membership function, so as to explain the reasoning process of the model. The specific steps are: first, train the KAN network according to the data, and when the training is completed, remove the weight function of the edge with too low weight influence; then, analyze the remaining activation function obtained by training and compare it with the known function library, and finally replace the activation function obtained by training with a specific known function, so as to achieve the interpretability of the network. The process can be summarized as "sparseness-visualization-pruning-symbolization".

[0024] Step S700, SHAP is used to rank the contribution of the input factors of the model, thereby explaining the importance of the model factors. In the ground feature echo recognition network, the SHAP value of each factor is as follows: Figure 5As shown, it can be seen that the average radial velocity V plays an absolutely dominant role. Although its average SHAP value is near 0, it can be seen from the relationship between the SHAP value and the size of V that when the average radial velocity is very small, the SHAP value is also small, which is exactly the characteristic of the ground object echo; and when V is large, the corresponding SHAP value is also large, that is, a large Doppler velocity generally corresponds to a meteorological echo.

Claims

1. A millimeter wave cloud radar echo intelligent recognition method based on KAN network, characterized in that: The following steps are involved: (1) Using automatic label generation technology to preprocess radar observation data, generate labels close to the actual echo classification, and construct a label dataset; (2) Integrate eigenvalue and fuzzy logic algorithms to construct the REC-KAN-MLP network; (3) The REC-KAN-MLP network is trained using the echo classification label dataset to obtain a radar echo target recognition model; (4) Obtain radar observation data in the study area, use the radar echo target recognition model to identify and classify the radar echo pixel by pixel, and evaluate and analyze multiple scoring indicators; (5) Perform interpretable analysis on the model output, quantify the contribution of each feature and construct membership functions.

2. According to the method for intelligent identification of millimeter wave cloud radar echo based on KAN network in claim 1, it is characterized in that: Step (1) includes the following steps: (11) Set the threshold based on the characteristics of the ground object echo; (12) Feature engineering technology based on radar echo characteristics uses the initial detection data of the radar to extract secondary data that further characterizes the semantics of the radar echo type for model input; (13) Perform morphological processing of corrosion and expansion on the labels.

3. According to claim 2, a method for intelligent identification of millimeter wave cloud radar echoes based on a KAN network is characterized in that: Step (11) is as follows: (111) Preliminary classification: When the radar reflectivity factor is NaN, it is marked as background field; when the average radial velocity value is less than 0.2m / s, it is marked as ground echo, otherwise it is marked as meteorological echo; (112) Refine the label based on more conditions: a When the average radial velocity value is less than 1, the radar reflectivity factor is less than -5dBZ, and the velocity spectrum width is greater than 0.4m / s, it is marked as a weather echo, otherwise it is marked as a ground object echo; b. When the average radial velocity value is greater than 0.2m / s and the radar reflectivity factor is greater than 18dBZ, it is marked as a weather echo, otherwise it is marked as a ground object echo; when the average radial velocity value is greater than 0.2m / s and the signal-to-noise ratio is greater than 18, it is marked as a weather echo, otherwise it is marked as a ground object echo; c When the average radial velocity value is less than 0.2m / s, when the NaN value in the 5×5 kernel window centered on the judgment point does not exceed 2, the maximum value of the average radial velocity in the two directions exceeds 0.1m / s and the average radial velocity difference in the kernel exceeds 0.4m / s, it is marked as a meteorological echo, otherwise it is marked as a ground object echo.

4. According to the method for intelligent identification of millimeter wave cloud radar echo based on KAN network in claim 2, it is characterized in that: Step (12) is as follows: For PPI type radar echo data, 11 features are extracted for further identification of ground object echoes and weather echoes, including 4 original radar data: radar reflectivity factor, average radial velocity V, velocity spectrum width W, radar echo signal-to-noise ratio SNR, polar coordinate radius calculated from the data array, and 6 features extracted by combining the fuzzy logic algorithm commonly used in radar echo classification: polar coordinate radius, echo intensity texture , along the radial direction , the degree of variation between reservoirs along the radial direction , the regional average of the radial velocity ,variance , the regional average of the velocity spectrum width .

5. The method for intelligent identification of millimeter wave cloud radar echoes based on KAN network according to claim 2 is characterized in that: In step (13), the morphological processing of corrosion and dilation is to remove noise and fill holes.

6. The method for intelligently identifying millimeter wave cloud radar echoes based on a KAN network according to claim 1, characterized in that: In step (2), the overall structure of the REC-KAN-MLP network is the REC structure, the KAN network membership function is introduced for fitting, and the MLP module is used for weight distribution; 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 KAN output probability, and after linear change, it is activated by the sigmoid function, and the output is the probability of the echo type .

7. The method for intelligently identifying millimeter wave cloud radar echoes based on a KAN network according to claim 1, characterized in that: Step (3) is as follows: first, each factor is used as input to train its own classification model; then, purification technology is used to screen the more important neurons and remove other neurons; finally, the retained neurons are used for training again to obtain the final model. In this model, the activation function of the first layer of the network is a function of the membership function, and the membership function of the second layer of the network is an external function. If multiple layers of activation functions are retained in the same layer of the network, they are in an equal-weighted addition relationship, and the probability that the output after passing through the two layers of the network is in the range of [0,1].

8. The method for intelligently identifying millimeter wave cloud radar echoes based on a KAN network according to claim 1, characterized in that: In step (3), during the training process, 10% of the data is randomly masked with 1 to 3 factors each time and is not input into the network for training.

9. The method for intelligent identification of millimeter wave cloud radar echoes based on KAN network according to claim 1, characterized in that: Step (4) The evaluation and analysis of multiple scoring indicators are as follows: Recognition rate Accuracy: ; Precision: ; Recall: ; F1 Score: ; False alarm rate FAR: ; Miss rate: ; Among them, the method for calculating the recognition rate of ground object echoes in the PPI dataset is as follows: TP represents the number of targets that are actually ground object echoes and are recognized as ground object echoes; FP indicates the number of targets that are actually non-ground object echoes but identified as ground object echoes; TN indicates the number of targets that are actually non-ground object echoes but identified as non-ground object echoes; and FN indicates the number of targets that are actually ground object echoes but identified as non-ground object echoes.

10. The method for intelligently identifying millimeter wave cloud radar echoes based on a KAN network according to claim 1, characterized in that: Step (5) is as follows: Use the KAN network to fit the membership function to explain the reasoning process of the model; use SHAP to rank the contribution of the input factors of the model.

Images