Intelligent identification system based on millimeter wave cloud radar data
By using clustering algorithms, automatic threshold segmentation algorithms and deep learning technology in the meteorological radar intelligent identification system, the problems of complex meteorological conditions and lack of labeled data in the existing technology are solved, and real-time identification of non-meteorological targets and the accuracy of meteorological data processing are improved.
Patent Information
- Application Number
- CN202411730272.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-03-18
AI Technical Summary
The existing meteorological radar algorithms will reduce accuracy when dealing with complex meteorological conditions, weak echo signals or large areas of clear sky echo areas, and lack deep learning algorithms and methods for labeling data.
An intelligent identification system based on millimeter wave cloud radar data is provided, and data labeling is adopted using clustering algorithms and automatic threshold segmentation algorithms. Data features are extracted using improved feature extraction modules and multi-channel fusion technology. Combined with spatial and channel attention mechanisms, the generalization ability and interpretability of the model are improved through jump connections and optimized feature recovery modules.
Real-time identification of non-meteorological targets in complex weather environments has been achieved, the accuracy and efficiency of meteorological radar data processing has been improved, and the support capabilities of meteorological early warning and disaster prevention and control have been enhanced.
Smart Images
Figure SMS_1 
Figure SMS_2 
Figure SMS_3
Abstract
Description
Technical Field
[0001] The present invention relates to the field of radar weather forecasting, and in particular to an intelligent recognition system based on millimeter wave cloud radar data. Background Art
[0002] Meteorological radar, especially millimeter-wave cloud radar, plays a vital role in modern meteorological monitoring. Millimeter-wave cloud radar uses high-frequency electromagnetic waves to detect meteorological targets (such as precipitation, clouds, fog, etc.) with high precision, and obtains important data such as echo intensity, radial velocity, velocity spectrum width and polarization information through the backscattering effect of meteorological particle groups on electromagnetic waves. These data play a decisive role in severe convective weather warnings, extreme weather causes, and the formulation of emergency plans for disastrous weather. However, millimeter-wave cloud radars will encounter interference signals from non-meteorological targets during the detection process, mainly including ground object echoes and clear sky echoes. Ground object echoes originate from the reflected echoes generated by the radar beam irradiating the ground object. They are usually strong and do not change with time, and are often confused with precipitation echoes. Clear sky echoes originate from phenomena such as low-altitude insects, birds or atmospheric turbulence. These echoes usually appear at the near end of the radar and appear as weak echo signals, which may be confused with low cloud echoes. Nowadays, through intelligent algorithms, it is possible to automatically distinguish meteorological echoes from non-meteorological echoes from a large amount of radar data, greatly improving data processing efficiency and accuracy, thereby enhancing the performance and reliability of meteorological radars in practical applications. The implementation of deep learning technology will provide higher accuracy for the refined analysis of meteorological data, provide more reliable support for meteorological warnings, disaster prevention and control, and promote technological innovation and development in the meteorological field. However, most of the existing algorithms in China are based on traditional mathematical and physical methods, deep learning algorithms have not been widely used in the meteorological field, and there is a lack of methods for labeling data. Therefore, it is crucial to invent a deep learning algorithm and a system for labeling data for meteorological forecasting. Summary of the invention
[0003] In view of the above shortcomings, the present invention provides an intelligent recognition system based on millimeter wave cloud radar data. For meteorological data that lacks labels, the present invention provides a classification based on clustering algorithms and automatic threshold segmentation algorithms, respectively, using K-means clustering, hierarchical clustering, Otsu method, and triangulation method to process preprocessed data respectively, obtain preliminary classification labels, use open operations on preliminary classification labels, design algorithms to optimize labels according to the characteristics of the data, open operations to eliminate data noise, design threshold labeling data according to the height characteristics of the data, and finally the algorithm visualizes different label optimization results, and manually selects the label closest to the true value. In view of the problem that the accuracy of existing classification algorithms will decrease when facing complex meteorological conditions, weak echo signals, or large areas of clear sky echo areas, an improved feature extraction module and multi-channel fusion technology are used to fully extract fusion data features, so that the model can be trained using richer feature representations, and the spatial and channel attention mechanisms are used to focus attention on important areas of the data and discard irrelevant data. In view of the fact that deep learning models require a large amount of data during training, long training time and limited interpretability, we use jump connections to connect the data after feature extraction to the upsampling recovery module to retain feature details, and the attention module to enhance information flow and improve the generalization ability of the model. We use the optimized feature recovery module to further simplify feature splicing; secondly, we use the dropout module to randomly discard neurons to prevent overfitting.
[0004] The intelligent recognition system based on millimeter wave cloud radar data provided by the present invention includes a data acquisition module, a data preprocessing module, a data marking module, a data classification module, a feature extraction module, a recognition prediction module and a user interface module;
[0005] The data acquisition module is responsible for collecting millimeter-wave cloud radar data, including clear sky echo data, meteorological echo data, ground object echo data, and non-ground object echo data;
[0006] The data preprocessing module preprocesses the collected millimeter-wave cloud radar data to obtain processed meteorological data;
[0007] The data labeling module performs algorithm classification and manual optimization on the preprocessed meteorological data to obtain training label data;
[0008] The data classification module uses a Boolean mask to separate the region of interest from the processed meteorological data to obtain radar signature data;
[0009] The feature extraction module uses an optimized convolutional neural network module to perform target recognition and extract features of non-meteorological data from the classification feature data for classification;
[0010] The recognition prediction module receives new millimeter wave cloud radar data as input, uses the data classification module and the feature extraction module to perform classification and feature extraction, and outputs the classification results of meteorological targets and non-meteorological targets;
[0011] The user interface module provides a user-friendly interactive interface, allowing meteorological staff and researchers to select data, view classification visualization results and adjust parameter settings of the identification and prediction module.
[0012] Furthermore, the data labeling module uses an improved classification algorithm and manual optimization to obtain meteorological target and non-meteorological target labeling data, specifically including the following steps:
[0013] Step S1: preprocessing the raw data, taking the processed millimeter wave cloud radar data as the basis, using the clustering algorithm and the automatic threshold segmentation algorithm for classification, respectively, to obtain preliminary classification labels;
[0014] Step S2: Label optimization, use open operations on the preliminary labels, and design algorithms to optimize the labels based on the characteristics of the data;
[0015] Step S3: Manual selection, the algorithm visualizes different label optimization results, manually selects the label closest to the true value and manually optimizes it;
[0016] Step S4: label naming, using the algorithm to automatically match the name of the label and the original data to facilitate subsequent training calls;
[0017] Furthermore, in step S1, clustering algorithms and automatic threshold segmentation algorithms are used for classification. Specifically, the algorithms used are K-means clustering, staggered clustering, Otsu's method, triangulation method, etc.
[0018] Step S11: Filter out a large number of null values in the original data using a Boolean mask;
[0019] Where M is the mask matrix. When M=1, the value of X is assigned to X. valid , M=0, the value of X is ignored, and the ignored value can be marked as 2 after processing; the mask formula is as follows:
[0020] Step S12: the categories of the clustering algorithm and the automatic threshold segmentation algorithm are set to 2, the processing object is the filtered radar data, and the algorithm divides the data into 2 categories;
[0021] Furthermore, the S2 feature extraction module uses an improved convolution module to extract radar echo data, which specifically includes the following steps:
[0022] Step P1: Data input. In THI mode, the data loader provides radar echo data of five channels: Z1, V1, W1, SNR1, and LDR1. In PPI mode, the data loader provides radar echo data of four channels: Z1, V1, W1, and SNR1.
[0023] Step P2: Multi-channel feature extraction. The feature extraction module uses multi-channel fusion technology to extract data from multiple channels at the same time, and copies a copy of the extracted features to the upsampling module through a skip connection;
[0024] Step P3: Integrate the spatial channel attention mechanism into the jump connection to assist network training;
[0025] Step P4: Repeat the extraction, the input data is downsampled four times and the feature replication is passed four times;
[0026] Step P5: Feature recovery, upsampling the features of the lowest layer and combining them with the features copied by the skip connection to obtain the restored result;
[0027] Step P6: Use the improved cross entropy loss function to measure the training results;
[0028] Furthermore, the spatial channel attention mechanism in step P3 is composed of a spatial attention mechanism module and a channel attention mechanism module, wherein the channel attention module F c , spatial attention module F s The calculation formula is as follows: F c =σ(W2·ReLU(W1·f gp (X))) F s =σ(Conv2D([AvgPool(X),MaxPool(X)])) Where σ: Sigmoid activation function, used to map the output to the [0,1] range, X: input feature map, shape is (B, C, H, W), where B is the batch size, C is the number of channels, H is the height of the feature map, and W is the width of the feature map. W1W2 global pooling operation, the weight matrix of the fully connected layer, used to map features, f gp (X) is usually global average pooling or global maximum pooling, ReLU: activation function ReLU, introducing nonlinear Conv2D: convolution operation, usually 1x1 convolution, used to fuse the features after average pooling and maximum pooling.
[0029] Furthermore, the cross entropy loss function calculation formula in step P6 is as follows: Where 1 is the indicator function, which is 1 when Y(i,j) is equal to category c, otherwise it is 0, and Pc(i,j) is the probability of category c predicted by the model. The loss values of all valid pixels are averaged to obtain the final loss value.
[0030] According to claim 1, the intelligent recognition system based on millimeter wave cloud radar data is characterized in that: the recognition prediction module uses a 1×1 convolution layer to reduce the number of channels to the number of categories num_classes=3. The output shape is (batch_size,3,360,100), which represents the score of each pixel for 3 categories. The index of the category with the largest score is obtained through torch.argmax, and the value of each pixel is the index of its predicted category (0,1,2). The numbers 0, 1, and 2 represent different echo types.
[0031] The present invention can identify non-meteorological targets of millimeter-wave cloud radar in real time by analyzing and processing millimeter-wave cloud radar data. In complex weather environments, helping to analyze non-meteorological targets is of great significance to production and life such as agriculture and transportation. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 A schematic diagram of an intelligent recognition system based on millimeter-wave cloud radar data provided by the present invention;
[0033] Figure 2 An improved downsampling module of the present invention;
[0034] Figure 3 An improved upsampling module of the present invention;
[0035] Figure 4 The neural network architecture of the present invention;
[0036] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION
[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0038] For example, see Figure 1,Weather radar, especially millimeter wave cloud radar, plays a vital role in modern meteorological monitoring.,However, deep learning algorithms are not widely used in the meteorological field, and there is a lack of methods to label data.
[0039] In order to overcome the above problems, the intelligent recognition system based on millimeter wave cloud radar data provided by the present invention is characterized by comprising a data acquisition module, a data preprocessing module, a data marking module, a data classification module, a feature extraction module, an identification prediction module and a user interface module;
[0040] The data acquisition module is responsible for collecting millimeter-wave cloud radar data, including clear sky echo data, meteorological echo data, ground object echo data, and non-ground object echo data;
[0041] The data preprocessing module preprocesses the collected millimeter-wave cloud radar data to obtain processed meteorological data;
[0042] The data labeling module performs algorithm classification and manual optimization on the preprocessed meteorological data to obtain training label data;
[0043] The data classification module uses a Boolean mask to separate the region of interest from the processed meteorological data to obtain radar signature data;
[0044] The feature extraction module uses an optimized convolutional neural network module to perform target recognition, extracting non-meteorological data features and ground object echo features from radar feature data for classification;
[0045] The recognition prediction module receives new millimeter wave cloud radar data as input, uses the data classification module and the feature extraction module to perform classification and feature extraction, and outputs the classification results of meteorological targets and non-meteorological targets;
[0046] The user interface module provides a user-friendly interactive interface, allowing meteorological staff and researchers to select data, view classification visualization results and adjust parameter settings of the identification and prediction module.
[0047] Embodiment 2, based on the above embodiment, the data set acquisition module collects relevant millimeter-wave cloud radar data including five types of data in THI mode: Z1, V1, W1, SNR1, and LDR1, and four types of data in PPI mode: Z1, V1, W1, and SNR1;
[0048] Embodiment 3, based on the above embodiment, the preprocessing operation in the data preprocessing module includes data visualization, using plt.imshow to convert the two-dimensional array echo_labels into a color map, and different values are represented by colors of different intensities. The preliminary judgment of color intensity can facilitate subsequent marking.
[0049] Embodiment 4: The data labeling module performs algorithm classification and manual optimization on the pre-processed meteorological data to obtain training label data, which specifically includes the following steps:
[0050] Step S1: preprocessing the raw data. The processed millimeter-wave cloud radar data is used as the basis for classification using clustering algorithms and automatic threshold segmentation algorithms. The clustering algorithms include hierarchical clustering and Kmeans clustering. The automatic threshold segmentation methods include Otsu method and triangulation method. Then, preliminary classification labels are obtained.
[0051] Step S2: Label optimization. Open operation is used for preliminary labels to remove some noise points. Algorithms are designed to optimize labels based on the characteristics of the data. The height of clear sky echoes generally does not appear at high altitudes, and the radar velocity component of ground object echoes is generally low. A height optimization algorithm and Doppler algorithm are designed.
[0052] Step S3: Manual selection, algorithm visualization of different label optimization results, manual selection of labels closest to the true value, clustering and other algorithms cannot clearly obtain the final result. We need to remove data with poor effects and misclassification based on algorithm classification to provide efficient and accurate labels for subsequent training;
[0053] Step S4: label naming, using the algorithm to automatically match the name of the label and the original data, to facilitate subsequent training calls. After the label is manually optimized, a label will be added based on the original name, and then the label index will be added based on the original name;
[0054] Embodiment 5, this example is based on the above embodiment, and the manually selected marked data in step S3 is the result obtained through certification by professionals, and the data such as clear sky echo, weather echo, ground object echo, and non-ground object echo are marked respectively.
[0055] Embodiment 6, based on the above embodiment, the feature extraction module uses an improved convolution module to extract radar echo data, specifically including the following steps:
[0056] Step P1: Data input, the data loader in THI mode provides radar echo data of five channels: Z1, V1, W1, SNR1, LDR1; the data loader in PPI mode provides radar echo data of four channels: Z1, V1, W1, SNR1, LDR1;
[0057] Step P2: Multi-channel feature extraction: the feature extraction module uses multi-channel fusion technology to extract data from five channels at the same time, and copies a copy of the extracted features to the upsampling module through a skip connection;
[0058] Step P3: Integrate the spatial channel attention mechanism into the jump connection to assist network training;
[0059] Step P4: Repeat the extraction, refer to Figure 4 ,The input data is downsampled four times and passed through four feature replications; the number of channels varies to 5, 64, 128, 256, and 512;
[0060] Step P5: Feature recovery, reference Figure 4 , the features of the lowest layer are upsampled and combined with the features copied by the skip connection to obtain the restored result. The channels change to 512, 256, 128, 64, and 1. The final 1 channel is the classification result;
[0061] Step P6: Use the improved cross entropy loss function to measure the training results. The improved cross entropy loss can effectively quantify the difference between the predicted probability distribution and the true distribution, prompting the model to generate accurate probability outputs, thereby optimizing the classification performance;
[0062] Embodiment 7, based on the above embodiment, the spatial channel attention mechanism in step P3 is composed of a spatial attention mechanism module and a channel attention mechanism module, wherein the channel attention module F c , spatial attention module F s The calculation formula is as follows: F c =σ(W2·ReLU(W1·f gp (X))) F s =σ(Conv2D([AvgPool(X),MaxPool(X)])) Where σ: Sigmoid activation function, used to map the output to the range [0,1], X: input feature map, shape (B,C,H,W), where B is the batch size, C is the number of channels, H is the height of the feature map, and W is the width of the feature map. W1W2 global pooling operation, the weight matrix of the fully connected layer, used to map features, f gp (X) is usually global average pooling or global maximum pooling, ReLU: activation function ReLU, introducing nonlinear Conv2D: convolution operation, usually 1x1 convolution, used to fuse the features after average pooling and maximum pooling.
[0063] The calculation formula of the cross entropy loss function in step P6 is as follows: Where 1 is the indicator function, which is 1 when Y(i,j) is equal to category c, otherwise it is 0, and Pc(i,j) is the probability of category c predicted by the model. The loss values of all valid pixels are averaged to obtain the final loss value.
[0064] The present invention can identify non-meteorological targets of millimeter-wave cloud radar in real time by analyzing and processing millimeter-wave cloud radar data. In complex weather environments, helping to analyze non-meteorological targets is of great significance to production and life such as agriculture and transportation.
[0065] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0066] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
[0067] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.
Claims
1. The intelligent recognition system based on millimeter wave cloud radar data is characterized by: It includes data acquisition module, data preprocessing module, data marking module, data classification module, feature extraction module, recognition prediction module and user interface module; The data acquisition module is responsible for collecting millimeter-wave cloud radar data, including clear sky echo data, meteorological echo data, ground object echo data, and non-ground object echo data; The data preprocessing module preprocesses the collected millimeter-wave cloud radar data to obtain processed data; The data labeling module performs algorithm classification and manual optimization on the preprocessed meteorological data to obtain training label data; The data classification module uses a Boolean mask to separate the region of interest from the processed meteorological data to obtain radar signature data; The feature extraction module uses an optimized convolutional neural network module to perform target recognition, extracting non-meteorological data features and ground object echo features from radar feature data for classification; The recognition prediction module receives new millimeter wave cloud radar data as input, uses the data classification module and the feature extraction module to perform classification and feature extraction, and outputs the classification results of meteorological targets and non-meteorological targets; The user interface module provides a user-friendly interactive interface, allowing meteorological staff and researchers to select data, view classification visualization results and adjust parameter settings of the identification and prediction module.
2. The intelligent recognition system based on millimeter wave cloud radar data according to claim 1 is characterized in that: The data labeling module uses an improved unsupervised clustering algorithm and manual optimization to obtain meteorological target labels, non-meteorological target labels, ground object echo labels, and non-ground object echo labels, specifically including the following steps: Step S1: preprocessing the raw data, taking the processed millimeter wave cloud radar data as the basis, using the clustering algorithm and the automatic threshold segmentation algorithm for classification, respectively, to obtain preliminary classification labels; Step S2: Label optimization, use open operations on the preliminary labels, and design algorithms to optimize the labels based on the characteristics of the data; Step S3: Manual selection, the algorithm visualizes different label optimization results, manually selects the label closest to the true value and manually optimizes it; Step S4: Label naming, using the algorithm to automatically match the name of the label and the original data to facilitate subsequent training calls.
3. The intelligent recognition system based on millimeter wave cloud radar data according to claim 1 is characterized in that: In step S1, clustering algorithm and automatic threshold segmentation algorithm are used for classification. Specifically, the algorithms used are K-means clustering, hierarchical clustering, Otsu method, triangulation method, etc.
4. The intelligent recognition system based on millimeter wave cloud radar data according to claim 1 is characterized in that: The S2 feature extraction module uses an improved convolution module to extract radar echo data, which specifically includes the following steps: Step P1: Data input, the data loader in THI mode provides radar echo data of five channels: Z1, V1, W1, SNR1, LDR1; the data loader in PPI mode provides radar echo data of four channels: Z1, V1, W1, SNR1, LDR1; Step P2: Multi-channel feature extraction. The feature extraction module uses multi-channel fusion technology to extract data from multiple channels at the same time, and copies a copy of the extracted features to the upsampling module through a skip connection; Step P3: Integrate the spatial channel attention mechanism into the jump connection to assist network training; Step P4: Repeat the extraction, the input data is downsampled four times and the feature replication is passed four times; Step P5: Feature recovery, upsampling the features of the lowest layer and combining them with the features copied by the skip connection to obtain the restored result; Step P6: Use the improved cross entropy loss function to measure the training results.
5. The intelligent recognition system based on millimeter wave cloud radar data according to claim 4 is characterized in that: In step P3, the spatial channel attention mechanism is composed of a spatial attention mechanism module and a channel attention mechanism module, wherein the channel attention module F c , spatial attention module F s The calculation formula is as follows: F c =σ(W2·ReLU(W1·f gp (X))) (1) F s =σ(Conv2D([AvgPool(X),MaxPool(X)])) (2) Where σ: Sigmoid activation function, used to map the output to the range [0,1], X: input feature map, shape is (B,C,H,W), where B is the batch size, C is the number of channels, H is the height of the feature map, W is the width of the feature map; W1W2 global pooling operation, the weight matrix of the fully connected layer, used to map features, f gp (X) is usually global average pooling or global maximum pooling, ReLU: activation function ReLU, introducing nonlinear Conv2D: convolution operation, usually 1x1 convolution, used to fuse the features after average pooling and maximum pooling.
6. The intelligent recognition system based on millimeter wave cloud radar data according to claim 4 is characterized in that: The calculation formula of the cross entropy loss function in step P6 is as follows: Where 1 is the indicator function, which is 1 when Y(i,j) is equal to category c, otherwise it is 0, Pc(i,j) is the probability of category c predicted by the model, and the loss values of all valid pixels are averaged to obtain the final loss value.
7. The intelligent recognition system based on millimeter wave cloud radar data according to claim 1 is characterized in that: The recognition prediction module uses a 1×1 convolutional layer to reduce the number of channels to the number of categories num_classes=3. The output shape is (batch_size,3,360,100), which represents the score of each pixel for the three categories. The value of each pixel is the index of its predicted category (0,1,2). The numbers 0, 1, and 2 represent different echo types. The torch.argmax function is used to obtain the category index with the largest score and output the final prediction result.