An intelligent recognition method for non-meteorological targets based on deep learning

Abstract

The present invention designs a method for intelligent recognition of non-meteorological targets based on deep learning, and identifies non-meteorological targets existing in millimeter-wave cloud radar data; specifically, by designing a millimeter-wave cloud radar data annotation algorithm and a data iteration algorithm, the corresponding labels are quickly obtained; a data expansion method for expanding millimeter-wave cloud radar data is designed to obtain more diverse training data, and feature extraction and feature restoration are designed to improve the recognition performance of non-meteorological targets; and an excellent recognition model is designed using the encoder-decoder concept. Through the above set of processes, the accuracy of intelligent recognition of non-meteorological targets of millimeter-wave cloud radar is improved. Furthermore, the present invention enhances prompts for specific labels by constructing a prompt enhancement module, thereby improving the sensitivity of the model to specific labels, reducing the computational complexity of the intelligent recognition model, and improving the generalization ability.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recognition of non-meteorological targets by millimeter-wave cloud radar, and in particular to an intelligent recognition method of non-meteorological targets based on deep learning. Background Art

[0002] Millimeter-wave cloud radar is a new type of meteorological radar that uses high-frequency electromagnetic waves to detect meteorological targets with high precision. It can be used for refined detection of meteorological targets such as precipitation, clouds, and fog. It uses the backscattering effect of meteorological particle groups on electromagnetic waves to obtain the echo intensity, radial velocity, velocity spectrum width, and polarization information of meteorological targets, and further generates secondary data products. Meteorological targets have the characteristics of weak echo signal intensity, slow movement speed, and obvious distributed characteristics. These characteristics cause millimeter-wave cloud radar to inevitably receive a large number of echo signals from non-meteorological targets at the same time during detection. These clutter sources exist in large numbers within the detection range of the radar, so the original echo data of the millimeter-wave cloud radar is the superposition of meteorological target echoes and non-meteorological target echoes, which brings difficulties to meteorological researchers when judging weather changes such as precipitation areas, precipitation characteristics, and cloud base height. Therefore, the design of intelligent recognition of non-meteorological targets of millimeter-wave cloud radar can effectively solve the key problems of non-meteorological target recognition technology.

[0003] At present, there are many non-meteorological target intelligent recognition algorithms, such as: non-meteorological target intelligent recognition based on fuzzy logic, intelligent recognition based on neural network, and recognition algorithm based on machine learning. These methods all need to expand more feature data, calculate the correlation of features, and provide labeled data as training data, which makes the intelligence level insufficient, the recognition accuracy is not high, and the generalization ability is poor. Therefore, in response to the above problems, this paper uses the millimeter wave cloud radar data labeling algorithm, the millimeter wave cloud radar data data expansion method, the millimeter wave cloud radar data detail information and global information extraction module, and the millimeter wave cloud radar data feature information fusion module to build an efficient and reliable millimeter wave cloud radar non-meteorological target intelligent model based on deep learning.

[0004] Specifically, due to the differences in detection environments and conditions, the data obtained from millimeter-wave cloud radar are of varying quality. How to use easy-to-classify samples in existing data to label difficult-to-classify samples and thus improve the utilization of detection data is also a technical problem that needs to be solved in this field.

[0005] Furthermore, the U-Net architecture is a classic convolutional neural network, which is mainly used for image segmentation tasks, especially in the field of medical image segmentation. However, when used to process massive data, the processing speed is slow and the generalization effect is not good. Summary of the invention

[0006] In view of the deficiencies of the prior art, the present invention provides an intelligent recognition method for non-meteorological targets based on deep learning, the method comprising:

[0007] Step S1: labeling the easy-to-classify samples in the millimeter-wave cloud radar data, where the millimeter-wave cloud radar data includes vertical upward scanning mode data and plane position scanning mode data, thereby constructing a labeled data set T1 of easy-to-classify samples in the vertical upward scanning mode data and a labeled data set P1 of easy-to-classify samples in the plane position scanning mode data; in the labeling process, selecting several specific labels as enhanced prompts;

[0008] Step S2: Use the data set of easy-to-classify samples in step S1 as positive samples for pseudo-label learning, so as to annotate the difficult-to-classify samples in the millimeter-wave cloud radar data, thereby constructing a data set of difficult-to-classify samples, and merge the data set of difficult-to-classify samples with the data set of easy-to-classify samples in step S1 to obtain a data set T3 of vertical upward scanning mode data and a data set P3 of plane position scanning mode data;

[0009] Step S3: expanding the data set T3 and the data set P3 in step S2 to obtain a total training data set T of vertical upward scanning mode data and a total training data set P of plane position scanning mode data;

[0010] Step S4: construct an intelligent recognition model based on the U-Net structure, which consists of an encoder, a prompt enhancement module, a feature fusion layer, a decoder, and a classification head; import the millimeter-wave cloud radar data of the total training data set T of the vertical upward scanning mode data or the total training data set P of the plane position scanning mode data into the encoder to obtain the encoding features;

[0011] The prompt enhancement module is designed in parallel with the encoder, and the enhanced prompt in step S1 is input into the prompt enhancement module to obtain enhanced prompt features consistent with the decoder resolution;

[0012] Step S5: importing the encoding features and enhanced prompt features into the decoder to obtain initial decoding features;

[0013] Step S6: Import the encoding features of step S4 and the initial decoding features of step S5 into the feature fusion layer for feature fusion, obtain the fused features, and upsample the fused features to the decoder. After the decoder restores the features again, obtain the final decoding features;

[0014] Step S7: The classification head processes the final decoded features of step S6 to obtain the predicted category of the millimeter wave cloud radar data.

[0015] Furthermore, step S1 is specifically as follows:

[0016] Step S11: dividing the vertical upward scanning mode data by altitude, coordinate information and reflectivity factor to obtain easy-to-classify samples and difficult-to-classify samples, where the easy-to-classify samples include clear sky echoes, weather echoes and background waves; generating a labeled data set T1 of easy-to-classify samples in the vertical upward scanning mode data;

[0017] Step S12: Divide the plane position scanning mode data through window filtering to obtain easy-to-classify samples and difficult-to-classify samples. The easy-to-classify samples include ground object echoes, clear sky echoes and background echoes; generate a labeled data set P1 of easy-to-classify samples in the plane position scanning mode data.

[0018] Furthermore, step S2 is specifically as follows:

[0019] Step S21: taking the data set T1 and the data set P1 obtained in step S1 as positive samples, and taking the plane position scanning mode data and the vertical upward scanning mode data in step S1 as training set D;

[0020] Step S22: constructing a preliminary classification model, and training the preliminary classification model by performing pseudo-label learning on the training set D based on positive samples. The preliminary classification model obtained after training annotates the difficult-to-classify samples in the vertical upward scanning mode data and the plane position scanning mode data, and obtains a data set T2 of difficult-to-classify samples in the vertical upward scanning mode data and a data set P2 of difficult-to-classify samples in the plane position scanning mode data;

[0021] Step S23: fine-tune the data sets T2 and P2 so that the reflectivity factors in the data sets T2 and P2 are both higher than the preset threshold value, merge the fine-tuned data set T2 and data set T1 to obtain a data set T3 of the vertical upward scanning mode data, and merge the fine-tune data set P2 and data set P1 to obtain a data set P3 of the plane position scanning mode data.

[0022] Furthermore, the data expansion of step S3 is specifically as follows:

[0023] Step S31: for the vertical upward scanning mode data set T3, randomly translate the samples and corresponding labels of the data set T3 by several units along the X-axis direction by random translation, and expand it into the vertical upward scanning mode data set T4; randomly flip the samples and corresponding labels of the data set T3 with the y-axis as the center axis by random flipping, and expand it into the vertical upward scanning mode data set T5; merge the data sets T3, T4, and T5 to obtain the total training data set T of the vertical upward scanning mode data;

[0024] Step S32: For the data set P3 of the plane position scanning pattern data, randomly translate the samples and corresponding labels of the data set P3 by several units along the X-axis direction through random translation, and expand it into the data set P4 of the plane position scanning pattern data; flip the samples and corresponding labels of the data set P3 with the y-axis as the center axis through random flipping, and expand it into the data set P5 of the plane position scanning pattern data; cover the samples and corresponding labels of the data set P3 to the data gaps through random replication, and expand it into the data set P6 of the plane position scanning pattern data; merge the data sets P3, P4, P5, and P6 to obtain the total training data set P of the plane position scanning pattern data.

[0025] Further, the encoder in step S4 is composed of three encoding layers with the same structure and two maximum pooling layers with the same structure, specifically a first encoding layer, a first maximum pooling layer, a second encoding layer, a second maximum pooling layer and a third encoding layer connected in series in sequence; the first, second and third encoding layers perform feature extraction on the input data to obtain a first encoding feature, a second encoding feature and a third encoding feature, respectively, and the maximum pooling layer performs maximum pooling processing on the input data;

[0026] The coding layers in step S4 are composed of three layers of depth residual blocks connected in series. The feature extraction process of the coding layer is expressed as:

[0027] ;

[0028] ;

[0029] in, is the output of a single-layer depth residual block, is the input of a single-layer depth residual block, is a convolution operation with a convolution kernel size of 3×3. is the batch normalization operation, To activate the operation, is the output of the encoding layer.

[0030] Further, the decoder is composed of three decoding layers with the same structure and two transposed upsampling layers with the same structure, specifically, a first decoding layer, a first transposed upsampling layer, a second decoding layer, a second transposed upsampling layer and a third decoding layer connected in series in sequence; the first, second and third decoding layers perform feature restoration on the input data to obtain a first initial decoding feature, a second initial decoding feature and a third initial decoding feature respectively; the transposed upsampling layer performs transposed upsampling after splicing the output of the decoding layer and the enhanced prompt feature in step S4;

[0031] The decoding layer performs convolution, batch normalization and rectified linear unit operations on the input data in different scales in turn to complete feature restoration. The processing process of the decoding layer is expressed as:

[0032] ;

[0033] in, is the output of the decoding layer, is the input of the decoding layer, The convolution kernel is a 3×3 convolution. is the batch normalization operation, is a nonlinear activation function used to correct linear units.

[0034] Further, the feature fusion layer includes three fusion layers with the same structure, specifically, a first fusion layer, a second fusion layer and a third fusion layer; the first coding feature and the third initial decoding feature are feature fused through the first fusion layer to obtain a first fusion feature; the second coding feature and the second initial decoding feature are feature fused through the second fusion layer to obtain a second fusion feature; the third coding feature and the first initial decoding feature are feature fused through the third fusion layer to obtain a third fusion feature; the third fusion feature is input into the first decoding layer, the output of the first decoding layer is imported into the first transposed upsampling layer, the output of the first transposed upsampling layer is spliced ​​with the second fusion feature and then imported into the second decoding layer, the output of the second decoding layer is input into the second transposed upsampling layer, the output of the second transposed upsampling layer is spliced ​​with the first fusion feature and then imported into the third decoding layer, and the third decoding layer outputs the final decoding feature;

[0035] The feature fusion layer in step S6 reduces the number of channels of the input data through 1×1 convolution and then performs pixel-by-pixel addition for feature fusion. Then, an attention weight is generated through a 1×1 convolution layer and a Sigmoid activation function, and the attention weight is applied to the output of the encoding layer. Finally, the fused feature is obtained, which is expressed as:

[0036] ;

[0037] in, represents the attention weight, To fusion features, is the output of the encoding layer, is the output of the decoding layer, is a 1×1 convolution operation, is the batch normalization operation, is a nonlinear activation operation, Sigmoid activation operation.

[0038] Furthermore, the classification head adjusts the number of channels of the final decoded feature to 3 through a 3×3 convolutional layer, and then uses the Softmax activation operation to obtain the probability value of each pixel of the final decoded feature belonging to each category, and predicts the category of the millimeter-wave cloud radar data by statistical probability values.

[0039] Furthermore, in step S22, the preliminary classification model is constructed by performing pseudo-label learning on the training set D based on positive samples, thereby training the preliminary classification model, specifically:

[0040] Step S221: Preliminary classification model for labeled positive samples And unlabeled difficult-to-classify samples Performing random change operations for preprocessing, the random change operations include rotation, inversion, cropping and grayscale transformation;

[0041] The preprocessed labeled positive samples And difficult to classify samples Input into the preliminary classification model for feature extraction operation, and obtain the corresponding labeled extracted features and unlabeled feature extraction ;

[0042] Using density clustering algorithm to extract features from labeled and unlabeled feature extraction Perform unsupervised clustering; specifically:

[0043] Step S222: For each data point , calculate the value of this data point Neighborhood, expressed as:

[0044] ;

[0045] in, express of Neighborhood, represents the Euclidean distance, Represents data points The adjacent data points of

[0046] The data points As the core point, The points in the neighborhood are grouped into the same cluster and continue to Neighborhood expansion: adjacent data points ,if It can also be used as a core point for further expansion Neighborhood; the constraints are expressed as:

[0047] ;

[0048] ;

[0049] in, is the minimum number of cluster core points, express Neighborhood;

[0050] Repeat the expansion process until the constraints cannot be satisfied;

[0051] Marks that are not core points and do not belong to any core points The points in the neighborhood are noise points, and no processing is done on them;

[0052] Step S223: After completing the clustering operation, obtain the clustering prototype , expressed as:

[0053] ;

[0054] in, is assigned to The feature set of each cluster; It is The number of features contained in each cluster; is assigned to data points in clusters;

[0055] Step S224: Cluster prototype As pseudo labels, a self-supervised loss function is constructed based on the pseudo labels, and the self-supervised loss function is minimized to further optimize the preliminary classification model;

[0056] Among them, the self-supervised loss function includes the contrast loss function and the mean square error consistency loss function. The contrast loss function It is expressed as:

[0057] ;

[0058] in, represents the pseudo labels corresponding to samples i and j, represents the unlabeled extracted features corresponding to sample i and sample j, is the indicator function, represents the Euclidean distance;

[0059] Among them, the mean square error consistency loss function It is expressed as:

[0060] ;

[0061] in, Indicates the total number of samples;

[0062] Self-supervised loss function It is expressed as:

[0063] .

[0064] Furthermore, in step S4, the enhanced prompt is input into the feature enhancement module to obtain the enhanced prompt feature consistent with the decoder resolution; specifically:

[0065] Input the enhanced prompt, use 3×3 convolution to downsample the enhanced prompt, and obtain the enhanced prompt feature consistent with the decoder resolution, expressed as:

[0066] ;

[0067] in, To enhance the prompt features, For a specific preset label, To preset another specific tag, is a 3×3 convolution operation, This is the batch normalization operation.

[0068] The beneficial effects of the present invention are:

[0069] The present invention provides a method for intelligently identifying non-meteorological targets based on deep learning. The method first performs pseudo-label learning on easy-to-classify samples in existing millimeter-wave cloud radar data, thereby annotating difficult-to-classify samples in the millimeter-wave cloud radar data to complete data expansion, thereby improving the utilization rate of the millimeter-wave cloud radar data. Furthermore, the present invention improves the feature classification capability of millimeter-wave cloud radar data by constructing an intelligent recognition model based on the U-Net architecture through an encoder-decoder structure. Furthermore, the present invention enhances prompts for specific labels by constructing a prompt enhancement module, thereby improving the sensitivity of the model to specific labels, reducing the computational complexity of the intelligent recognition model, and improving the generalization capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0071] Figure 1 A flowchart of the steps of an intelligent identification method for non-meteorological targets based on deep learning provided by the present invention. DETAILED DESCRIPTION

[0072] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0074] Reference Figure 1 , an intelligent recognition method for non-meteorological targets based on deep learning, the method comprising:

[0075] Step S1: labeling the easy-to-classify samples in the millimeter-wave cloud radar data, where the millimeter-wave cloud radar data includes vertical upward scanning mode data and plane position scanning mode data, thereby constructing a labeled data set T1 of easy-to-classify samples in the vertical upward scanning mode data and a labeled data set P1 of easy-to-classify samples in the plane position scanning mode data; in the labeling process, selecting several specific labels as enhanced prompts;

[0076] Step S2: Use the data set of easy-to-classify samples in step S1 as positive samples for pseudo-label learning, so as to annotate the difficult-to-classify samples in the millimeter-wave cloud radar data, thereby constructing a data set of difficult-to-classify samples, and merge the data set of difficult-to-classify samples with the data set of easy-to-classify samples in step S1 to obtain a data set T3 of vertical upward scanning mode data and a data set P3 of plane position scanning mode data;

[0077] Step S3: expanding the data set T3 and the data set P3 in step S2 to obtain a total training data set T of vertical upward scanning mode data and a total training data set P of plane position scanning mode data;

[0078] Step S4: construct an intelligent recognition model based on the U-Net structure, which consists of an encoder, a prompt enhancement module, a feature fusion layer, a decoder, and a classification head; import the millimeter-wave cloud radar data of the total training data set T of the vertical upward scanning mode data or the total training data set P of the plane position scanning mode data into the encoder to obtain the encoding features;

[0079] The prompt enhancement module is designed in parallel with the encoder, and the enhanced prompt in step S1 is input into the prompt enhancement module to obtain enhanced prompt features consistent with the decoder resolution;

[0080] Step S5: importing the encoding features and enhanced prompt features into the decoder to obtain initial decoding features;

[0081] Step S6: Import the encoding features of step S4 and the initial decoding features of step S5 into the feature fusion layer for feature fusion, obtain the fused features, and upsample the fused features to the decoder. After the decoder restores the features again, obtain the final decoding features;

[0082] Step S7: The classification head processes the final decoded features of step S6 to obtain the predicted category of the millimeter wave cloud radar data.

[0083] Furthermore, step S1 is specifically as follows:

[0084] Step S11: dividing the vertical upward scanning mode data by altitude, coordinate information and reflectivity factor to obtain easy-to-classify samples and difficult-to-classify samples, where the easy-to-classify samples include clear sky echoes, weather echoes and background waves; generating a labeled data set T1 of easy-to-classify samples in the vertical upward scanning mode data;

[0085] Step S12: Divide the plane position scanning mode data through window filtering to obtain easy-to-classify samples and difficult-to-classify samples. The easy-to-classify samples include ground object echoes, clear sky echoes and background echoes; generate a labeled data set P1 of easy-to-classify samples in the plane position scanning mode data.

[0086] Furthermore, step S2 is specifically as follows:

[0087] Step S21: taking the data set T1 and the data set P1 obtained in step S1 as positive samples, and taking the plane position scanning mode data and the vertical upward scanning mode data in step S1 as training set D;

[0088] Step S22: constructing a preliminary classification model, and training the preliminary classification model by performing pseudo-label learning on the training set D based on positive samples. The preliminary classification model obtained after training annotates the difficult-to-classify samples in the vertical upward scanning mode data and the plane position scanning mode data, and obtains a data set T2 of difficult-to-classify samples in the vertical upward scanning mode data and a data set P2 of difficult-to-classify samples in the plane position scanning mode data;

[0089] Step S23: fine-tune the data sets T2 and P2 so that the reflectivity factors in the data sets T2 and P2 are both higher than the preset threshold value, merge the fine-tuned data set T2 and data set T1 to obtain a data set T3 of the vertical upward scanning mode data, and merge the fine-tune data set P2 and data set P1 to obtain a data set P3 of the plane position scanning mode data.

[0090] Furthermore, the data expansion of step S3 is specifically as follows:

[0091] Step S31: for the vertical upward scanning mode data set T3, randomly translate the samples and corresponding labels of the data set T3 by several units along the X-axis direction by random translation, and expand it into the vertical upward scanning mode data set T4; randomly flip the samples and corresponding labels of the data set T3 with the y-axis as the center axis by random flipping, and expand it into the vertical upward scanning mode data set T5; merge the data sets T3, T4, and T5 to obtain the total training data set T of the vertical upward scanning mode data;

[0092] Step S32: For the data set P3 of the plane position scanning pattern data, randomly translate the samples and corresponding labels of the data set P3 by several units along the X-axis direction through random translation, and expand it into the data set P4 of the plane position scanning pattern data; flip the samples and corresponding labels of the data set P3 with the y-axis as the center axis through random flipping, and expand it into the data set P5 of the plane position scanning pattern data; cover the samples and corresponding labels of the data set P3 to the data gaps through random replication, and expand it into the data set P6 of the plane position scanning pattern data; merge the data sets P3, P4, P5, and P6 to obtain the total training data set P of the plane position scanning pattern data.

[0093] Further, the encoder in step S4 is composed of three encoding layers with the same structure and two maximum pooling layers with the same structure, specifically a first encoding layer, a first maximum pooling layer, a second encoding layer, a second maximum pooling layer and a third encoding layer connected in series in sequence; the first, second and third encoding layers perform feature extraction on the input data to obtain a first encoding feature, a second encoding feature and a third encoding feature, respectively, and the maximum pooling layer performs maximum pooling processing on the input data;

[0094] The coding layers in step S4 are composed of three layers of depth residual blocks connected in series. The feature extraction process of the coding layer is expressed as:

[0095] ;

[0096] ;

[0097] in, is the output of a single-layer depth residual block, is the input of a single-layer depth residual block, is a convolution operation with a convolution kernel size of 3×3. is the batch normalization operation, To activate the operation, is the output of the encoding layer.

[0098] Further, the decoder is composed of three decoding layers with the same structure and two transposed upsampling layers with the same structure, specifically, a first decoding layer, a first transposed upsampling layer, a second decoding layer, a second transposed upsampling layer and a third decoding layer connected in series in sequence; the first, second and third decoding layers perform feature restoration on the input data to obtain a first initial decoding feature, a second initial decoding feature and a third initial decoding feature respectively; the transposed upsampling layer performs transposed upsampling after splicing the output of the decoding layer and the enhanced prompt feature in step S4;

[0099] The decoding layer performs convolution, batch normalization and rectified linear unit operations on the input data in different scales in turn to complete feature restoration. The processing process of the decoding layer is expressed as:

[0100] ;

[0101] in, is the output of the decoding layer, is the input of the decoding layer, The convolution kernel is a 3×3 convolution. is the batch normalization operation, is a nonlinear activation function used to correct linear units.

[0102] Further, the feature fusion layer includes three fusion layers with the same structure, specifically, a first fusion layer, a second fusion layer and a third fusion layer; the first coding feature and the third initial decoding feature are feature fused through the first fusion layer to obtain a first fusion feature; the second coding feature and the second initial decoding feature are feature fused through the second fusion layer to obtain a second fusion feature; the third coding feature and the first initial decoding feature are feature fused through the third fusion layer to obtain a third fusion feature; the third fusion feature is input into the first decoding layer, the output of the first decoding layer is imported into the first transposed upsampling layer, the output of the first transposed upsampling layer is spliced ​​with the second fusion feature and then imported into the second decoding layer, the output of the second decoding layer is input into the second transposed upsampling layer, the output of the second transposed upsampling layer is spliced ​​with the first fusion feature and then imported into the third decoding layer, and the third decoding layer outputs the final decoding feature;

[0103] The feature fusion layer in step S6 reduces the number of channels of the input data through 1×1 convolution and then performs pixel-by-pixel addition for feature fusion. Then, an attention weight is generated through a 1×1 convolution layer and a Sigmoid activation function, and the attention weight is applied to the output of the encoding layer. Finally, the fused feature is obtained, which is expressed as:

[0104] ;

[0105] in, represents the attention weight, To fusion features, is the output of the encoding layer, is the output of the decoding layer, is a 1×1 convolution operation, is the batch normalization operation, is a nonlinear activation operation, Sigmoid activation operation.

[0106] Furthermore, the classification head adjusts the number of channels of the final decoded feature to 3 through a 3×3 convolutional layer, and then uses the Softmax activation operation to obtain the probability value of each pixel of the final decoded feature belonging to each category, and predicts the category of the millimeter-wave cloud radar data by statistical probability values.

[0107] Furthermore, in step S22, the preliminary classification model is constructed by performing pseudo-label learning on the training set D based on positive samples, thereby training the preliminary classification model, specifically:

[0108] Step S221: Preliminary classification model for labeled positive samples And unlabeled difficult-to-classify samples Performing random change operations for preprocessing, the random change operations include rotation, inversion, cropping and grayscale transformation;

[0109] The preprocessed labeled positive samples And difficult to classify samples Input into the preliminary classification model for feature extraction operation, and obtain the corresponding labeled extracted features and unlabeled feature extraction ;

[0110] Using density clustering algorithm to extract features from labeled and unlabeled feature extraction Perform unsupervised clustering; specifically:

[0111] Step S222: For each data point , calculate the value of this data point Neighborhood, expressed as:

[0112] ;

[0113] in, express of Neighborhood, represents the Euclidean distance, Represents data points The adjacent data points of

[0114] The data points As the core point, The points in the neighborhood are grouped into the same cluster and continue to Neighborhood expansion: adjacent data points ,if It can also be used as a core point for further expansion Neighborhood; the constraints are expressed as:

[0115] ;

[0116] ;

[0117] in, is the minimum number of cluster core points, express Neighborhood;

[0118] Repeat the expansion process until the constraints cannot be satisfied;

[0119] Marks that are not core points and do not belong to any core points The points in the neighborhood are noise points, and no processing is done on them;

[0120] Step S223: After completing the clustering operation, obtain the clustering prototype , expressed as:

[0121] ;

[0122] in, is assigned to The feature set of each cluster; It is The number of features contained in each cluster; is assigned to data points in clusters;

[0123] Step S224: Cluster prototype As pseudo labels, a self-supervised loss function is constructed based on the pseudo labels, and the self-supervised loss function is minimized to further optimize the preliminary classification model;

[0124] Among them, the self-supervised loss function includes the contrast loss function and the mean square error consistency loss function. The contrast loss function It is expressed as:

[0125] ;

[0126] in, represents the pseudo labels corresponding to samples i and j, represents the unlabeled extracted features corresponding to sample i and sample j, is the indicator function, represents the Euclidean distance;

[0127] Among them, the mean square error consistency loss function It is expressed as:

[0128] ;

[0129] in, Indicates the total number of samples;

[0130] Self-supervised loss function It is expressed as:

[0131] .

[0132] Furthermore, in step S4, the enhanced prompt is input into the feature enhancement module to obtain the enhanced prompt feature consistent with the decoder resolution; specifically:

[0133] Input the enhanced prompt, use 3×3 convolution to downsample the enhanced prompt, and obtain the enhanced prompt feature consistent with the decoder resolution, expressed as:

[0134] ;

[0135] in, To enhance the prompt features, For a specific preset label, To preset another specific tag, is a 3×3 convolution operation, This is the batch normalization operation.

[0136] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. An intelligent recognition method for non-meteorological targets based on deep learning, characterized in that: The method comprises: Step S1: labeling the easy-to-classify samples in the millimeter-wave cloud radar data, where the millimeter-wave cloud radar data includes vertical upward scanning mode data and plane position scanning mode data, thereby constructing a labeled data set T1 of easy-to-classify samples in the vertical upward scanning mode data and a labeled data set P1 of easy-to-classify samples in the plane position scanning mode data; in the labeling process, selecting several specific labels as enhanced prompts; Step S2: Use the data set of easy-to-classify samples in step S1 as positive samples for pseudo-label learning, so as to annotate the difficult-to-classify samples in the millimeter-wave cloud radar data, thereby constructing a data set of difficult-to-classify samples, and merge the data set of difficult-to-classify samples with the data set of easy-to-classify samples in step S1 to obtain a data set T3 of vertical upward scanning mode data and a data set P3 of plane position scanning mode data; Step S3: expanding the data set T3 and the data set P3 in step S2 to obtain a total training data set T of vertical upward scanning mode data and a total training data set P of plane position scanning mode data; Step S4: construct an intelligent recognition model based on the U-Net structure, which consists of an encoder, a prompt enhancement module, a feature fusion layer, a decoder, and a classification head; import the millimeter-wave cloud radar data of the total training data set T of the vertical upward scanning mode data or the total training data set P of the plane position scanning mode data into the encoder to obtain the encoding features; The prompt enhancement module is designed in parallel with the encoder, and the enhanced prompt in step S1 is input into the prompt enhancement module to obtain enhanced prompt features consistent with the decoder resolution; Step S5: importing the encoding features and enhanced prompt features into the decoder to obtain initial decoding features; Step S6: Import the encoding features of step S4 and the initial decoding features of step S5 into the feature fusion layer for feature fusion, obtain the fused features, and upsample the fused features to the decoder. After the decoder restores the features again, obtain the final decoding features; Step S7: The classification head processes the final decoded features of step S6 to obtain the predicted category of the millimeter wave cloud radar data.

2. According to claim 1, a method for intelligent identification of non-meteorological targets based on deep learning is characterized in that: Step S1 is specifically as follows: Step S11: dividing the vertical upward scanning mode data by altitude, coordinate information and reflectivity factor to obtain easy-to-classify samples and difficult-to-classify samples, wherein the easy-to-classify samples include clear sky echoes, weather echoes and background echoes; Generate a labeled dataset T1 of easily classified samples in vertical upward scanning mode data; Step S12: Divide the plane position scanning mode data through window filtering to obtain easy-to-classify samples and difficult-to-classify samples. The easy-to-classify samples include ground object echoes, clear sky echoes and background echoes; generate a labeled data set P1 of easy-to-classify samples in the plane position scanning mode data.

3. The method for intelligently identifying non-meteorological targets based on deep learning according to claim 2, characterized in that: Step S2 is specifically as follows: Step S21: taking the data set T1 and the data set P1 obtained in step S1 as positive samples, and taking the plane position scanning mode data and the vertical upward scanning mode data in step S1 as training set D; Step S22: constructing a preliminary classification model, and training the preliminary classification model by performing pseudo-label learning on the training set D based on positive samples. The preliminary classification model obtained after training annotates the difficult-to-classify samples in the vertical upward scanning mode data and the plane position scanning mode data, and obtains a data set T2 of difficult-to-classify samples in the vertical upward scanning mode data and a data set P2 of difficult-to-classify samples in the plane position scanning mode data; Step S23: fine-tune the data sets T2 and P2 so that the reflectivity factors in the data sets T2 and P2 are both higher than the preset threshold value, merge the fine-tuned data set T2 and data set T1 to obtain a data set T3 of the vertical upward scanning mode data, and merge the fine-tune data set P2 and data set P1 to obtain a data set P3 of the plane position scanning mode data.

4. The method for intelligently identifying non-meteorological targets based on deep learning according to claim 3 is characterized in that: The data expansion of step S3 is specifically as follows: Step S31: for the vertical upward scanning mode data set T3, randomly translate the samples and corresponding labels of the data set T3 by several units along the X-axis direction by random translation, and expand it into the vertical upward scanning mode data set T4; randomly flip the samples and corresponding labels of the data set T3 with the y-axis as the center axis by random flipping, and expand it into the vertical upward scanning mode data set T5; merge the data sets T3, T4, and T5 to obtain the total training data set T of the vertical upward scanning mode data; Step S32: For the data set P3 of the plane position scanning pattern data, randomly translate the samples and corresponding labels of the data set P3 by several units along the X-axis direction through random translation, and expand it into the data set P4 of the plane position scanning pattern data; flip the samples and corresponding labels of the data set P3 with the y-axis as the center axis through random flipping, and expand it into the data set P5 of the plane position scanning pattern data; cover the samples and corresponding labels of the data set P3 to the data gaps through random replication, and expand it into the data set P6 of the plane position scanning pattern data; merge the data sets P3, P4, P5, and P6 to obtain the total training data set P of the plane position scanning pattern data.

5. The method for intelligently identifying non-meteorological targets based on deep learning according to claim 1, characterized in that: The encoder in step S4 is composed of three encoding layers with the same structure and two maximum pooling layers with the same structure, specifically a first encoding layer, a first maximum pooling layer, a second encoding layer, a second maximum pooling layer and a third encoding layer connected in series in sequence; the first, second and third encoding layers perform feature extraction on the input data to obtain the first encoding feature, the second encoding feature and the third encoding feature respectively, and the maximum pooling layer performs maximum pooling processing on the input data; The coding layers in step S4 are composed of three layers of depth residual blocks connected in series. The feature extraction process of the coding layer is expressed as: ； ； in, is the output of a single-layer depth residual block, is the input of a single-layer depth residual block, is a convolution operation with a convolution kernel size of 3×3. is the batch normalization operation, To activate the operation, is the output of the encoding layer.

6. The method for intelligently identifying non-meteorological targets based on deep learning according to claim 1, characterized in that: The decoder is composed of three decoding layers with the same structure and two transposed upsampling layers with the same structure, specifically, a first decoding layer, a first transposed upsampling layer, a second decoding layer, a second transposed upsampling layer and a third decoding layer connected in series in sequence; the first, second and third decoding layers restore the features of the input data to obtain the first initial decoding feature, the second initial decoding feature and the third initial decoding feature respectively; The transpose upsampling layer concatenates the output of the decoding layer and the enhanced hint feature in step S4 and then performs transpose upsampling; The decoding layer performs convolution, batch normalization and rectified linear unit operations on the input data in different scales in turn to complete feature restoration. The processing process of the decoding layer is expressed as: ； in, is the output of the decoding layer, is the input of the decoding layer, The convolution kernel is a 3×3 convolution. is the batch normalization operation, is a nonlinear activation function used to correct linear units.

7. The method for intelligently identifying non-meteorological targets based on deep learning according to claim 6, characterized in that: The feature fusion layer includes three fusion layers with the same structure, specifically a first fusion layer, a second fusion layer and a third fusion layer; the first coding feature and the third initial decoding feature are fused through the first fusion layer to obtain a first fusion feature; The second encoding feature and the second initial decoding feature are fused through a second fusion layer to obtain a second fused feature; The third encoding feature and the first initial decoding feature are fused through the third fusion layer to obtain the third fusion feature; the third fusion feature is input to the first decoding layer, the output of the first decoding layer is imported into the first transposed upsampling layer, the output of the first transposed upsampling layer is spliced ​​with the second fusion feature and then imported into the second decoding layer, the output of the second decoding layer is input to the second transposed upsampling layer, the output of the second transposed upsampling layer is spliced ​​with the first fusion feature and then imported into the third decoding layer, and the third decoding layer outputs the final decoding feature; The feature fusion layer in step S6 reduces the number of channels of the input data through 1×1 convolution and then performs pixel-by-pixel addition for feature fusion. Then, an attention weight is generated through a 1×1 convolution layer and a Sigmoid activation function, and the attention weight is applied to the output of the encoding layer. Finally, the fused feature is obtained, which is expressed as: ； in, represents the attention weight, To fusion features, is the output of the encoding layer, is the output of the decoding layer, is a 1×1 convolution operation, is the batch normalization operation, is a nonlinear activation operation, Sigmoid activation operation.

8. The method for intelligently identifying non-meteorological targets based on deep learning according to claim 7, characterized in that: The classification head adjusts the number of channels of the final decoded feature to 3 through a 3×3 convolutional layer, and then uses the Softmax activation operation to obtain the probability value of each pixel of the final decoded feature belonging to each category, and predicts the category of the millimeter-wave cloud radar data by statistical probability values.

9. The method for intelligently identifying non-meteorological targets based on deep learning according to claim 3, characterized in that: The construction of the preliminary classification model in step S22 is performed by performing pseudo-label learning on the training set D based on positive samples, thereby training the preliminary classification model, specifically: Step S221: Preliminary classification model for labeled positive samples And unlabeled difficult-to-classify samples Performing random change operations for preprocessing, the random change operations include rotation, inversion, cropping and grayscale transformation; The preprocessed labeled positive samples And difficult to classify samples Input into the preliminary classification model for feature extraction operation, and obtain the corresponding labeled extracted features and unlabeled feature extraction ; Using density clustering algorithm to extract features from labeled and unlabeled feature extraction Perform unsupervised clustering; specifically: Step S222: For each data point , calculate the value of this data point Neighborhood, expressed as: ； in, express of Neighborhood, represents the Euclidean distance, Represents data points The adjacent data points of The data points As the core point, The points in the neighborhood are grouped into the same cluster and continue to Neighborhood expansion: adjacent data points ,if It can also be used as a core point for further expansion Neighborhood; the constraints are expressed as: ； ； in, is the minimum number of cluster core points, express Neighborhood; Repeat the expansion process until the constraints cannot be satisfied; Marks that are not core points and do not belong to any core points The points in the neighborhood are noise points, and no processing is done on them; Step S223: After completing the clustering operation, obtain the clustering prototype , expressed as: ； in, is assigned to The feature set of each cluster; It is The number of features contained in each cluster; is assigned to data points in clusters; Step S224: Cluster prototype As pseudo labels, a self-supervised loss function is constructed based on the pseudo labels, and the self-supervised loss function is minimized to further optimize the preliminary classification model; Among them, the self-supervised loss function includes the contrast loss function and the mean square error consistency loss function. The contrast loss function It is expressed as: ； in, represents the pseudo labels corresponding to samples i and j, represents the unlabeled extracted features corresponding to sample i and sample j, is the indicator function, represents the Euclidean distance; Among them, the mean square error consistency loss function It is expressed as: ； in, Indicates the total number of samples; Self-supervised loss function It is expressed as: 。 10. The method for intelligently identifying non-meteorological targets based on deep learning according to claim 1, characterized in that: In step S4, the enhanced prompt is input into the feature enhancement module to obtain the enhanced prompt feature consistent with the decoder resolution; Specifically: Input the enhanced prompt, use 3×3 convolution to downsample the enhanced prompt, and obtain the enhanced prompt feature consistent with the decoder resolution, which is expressed as: ； in, To enhance the prompt features, For a specific preset label, To preset another specific tag, is a 3×3 convolution operation, This is the batch normalization operation.

Images