A Quality Control Method and System for Multi-Radar Reflectivity Mosaic Based on Artificial Intelligence

Abstract

This invention provides a quality control method and system for multi-radar reflectivity mosaic based on artificial intelligence, comprising: acquiring an original multi-dimensional radar dataset; modeling the original multi-dimensional radar dataset based on category theory to construct a radar morphological network; performing propagation error analysis on the radar morphological network based on a preset data error analysis model to generate a propagation error feature set; performing risk probability modeling based on the radar morphological network and the propagation error feature set to obtain a multi-radar reflectivity fusion probability field; and employing deep reinforcement learning to simultaneously combine the multi-radar reflectivity fusion probability field to perform mosaicking, thereby generating a multi-radar reflectivity mosaic and improving the quality and reliability of the multi-radar reflectivity mosaic.

Description

Technical Field

[0001] This invention relates to the field of digital processing technology, and more specifically, to a method and system for quality control of multi-radar reflectivity mosaic based on artificial intelligence. Background Technology

[0002] Multi-radar reflectivity mosaicking technology, by fusing observation data from multiple radars, can generate a three-dimensional atmospheric reflectivity field with wider coverage and more uniform detection capabilities, making it an important tool for modern meteorological monitoring, short-term forecasting, and hydrological early warning. However, traditional methods often have certain limitations when constructing high-quality, highly consistent mosaicking products.

[0003] First, traditional methods often employ fixed processing pipelines, making it difficult to adaptively optimize data preprocessing strategies for specific radars or weather conditions, resulting in inconsistent data quality. Second, traditional methods often rely on empirical corrections based on climate statistics or simple models, frequently lacking modeling of the impact of atmospheric propagation. Third, when observations of the same area from multiple radars differ or even conflict, traditional methods mostly obtain a single optimal value through mathematical means, often lacking analysis of the uncertainties inherent in the observations, thus leading to defects in the accuracy of the final generated three-dimensional atmospheric reflectivity field. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a multi-radar reflectivity mosaic quality control method based on artificial intelligence, the method comprising: Obtain the original multidimensional radar dataset, model the original multidimensional radar dataset based on category theory, and construct the radar state-radial network; Based on a preset data error analysis model, the propagation error analysis is performed on the radar state-to-ground network to generate a propagation error feature set. Risk probability modeling is performed based on radar reflectivity network and propagation error feature set to obtain multi-radar reflectivity fusion probability field; Deep reinforcement learning is used to simultaneously combine the multi-radar reflectivity fusion probability field to generate a multi-radar reflectivity mosaic.

[0005] As a further aspect of the present invention, the original multidimensional radar dataset is obtained, and a radar state-to-radius network is constructed based on category theory, including: A preset meta-learning controller is used to perform data feature analysis on the original multidimensional radar dataset to obtain feature analysis results; Based on the feature analysis results, the corresponding preprocessing operator sequence is selected for the original multidimensional radar dataset to perform data preprocessing and obtain the initial multidimensional radar dataset. Graph neural networks and category theory are used to model the initial multidimensional radar dataset to generate a radar state-to-beam network. The nodes of the radar state-to-ground network are represented by the data source and processing module corresponding to each radar, while the edges represent the data flow and the corresponding transformation relationship. Data quality is evaluated for nodes within the radar dynamic network based on a deep neural network, and corresponding data quality scores are generated.

[0006] As a further aspect of the present invention, propagation error analysis is performed on the radar state-to-ground network based on a preset data error analysis model to generate a propagation error feature set, including: Obtain a radar propagation path environment dataset, which is used to describe the environmental information on the radar's corresponding propagation path and includes at least temperature, air pressure, humidity, and wind vector data. Based on the radar propagation path environment dataset, the environment of each radar ray path is modeled to generate a path state feature vector. The data error analysis model predicts propagation error based on path state feature vectors, generating an error correction field and an error sensitivity map. Based on the error correction field, the initial multidimensional radar dataset contained in the radar state-to-ground network is corrected for errors, and the error correction performance results are obtained. The error correction field, error correction effectiveness results, and error sensitivity map are encapsulated into a propagation error feature set for output.

[0007] As a further aspect of the present invention, the method further includes: A data error analysis model is constructed based on a physical information neural network. This data error analysis model is used to predict and correct propagation errors. A supervised learning strategy is adopted, and a pre-set loss function is simultaneously combined to pre-train the data error analysis model; The error correction field is generated by fusing the phase error correction field and the attenuation error correction field, and includes at least the corresponding error correction amount and uncertainty index. The error sensitivity map is generated by backpropagation to calculate the gradient of the propagation error on each dimension of the path state feature vector, and is used to quantify the contribution of environmental information on each propagation path in different regions to the propagation error.

[0008] As a further aspect of the present invention, risk probability modeling is performed based on radar reflectivity network and propagation error feature set to obtain a multi-radar reflectivity fusion probability field, including: Obtain the target grid point set of the multi-radar reflectivity mosaic, and construct a global probabilistic inference graph based on the spatial proximity relationship between each target grid point, using all target grid points in the target grid point set as nodes. The nodes in the global probability inference graph are used to represent the reflectivity variables of the current target grid point, and each node corresponds to an observation likelihood function. The inner edge connection of the global probability inference graph corresponds to a potential function, which is used to encode the preset spatial continuity prior. Based on the global probability inference graph, a belief propagation algorithm is used to perform approximate inference to generate a radar reflectivity fusion probability field. The radar reflectivity fusion probability field is used to describe the probability distribution of all possible values ​​of nodes within the global probabilistic inference graph.

[0009] As a further aspect of the present invention, deep reinforcement learning is employed to simultaneously combine the multi-radar reflectivity fusion probability field for mosaicking, generating a multi-radar reflectivity mosaic, including: An adversarial reinforcement learning environment, comprising generative and discriminative agents, is constructed based on cellular automata and trained using a curriculum learning strategy. The adversarial reinforcement learning environment is initialized based on the multi-radar reflectivity fusion probability field, and the generated agent and the discriminative agent engage in multiple rounds of adversarial game based on the current environment state to generate the initial multi-radar reflectivity fusion field. Perform regional feature analysis on the initial multi-radar reflectivity fusion field to obtain the regional feature analysis results; Based on the regional feature analysis results, the initial multi-radar reflectivity fusion field is locally enhanced to generate the optimal multi-radar reflectivity fusion field, which is the multi-radar reflectivity mosaic.

[0010] As a further aspect of the present invention, an adversarial reinforcement learning environment comprising a generative agent and a discriminative agent is constructed based on cellular automata, and a course learning strategy is employed for training, including: The cellular automaton is used to predict the state of a cell at the next moment based on the current state of the cell itself and its neighboring cells. The state of the cell includes adjustable variables and background variables. The state space of an adversarial reinforcement learning environment consists of a radar reflectivity fusion probability field and the corresponding environmental feature vector. The action space of the generated agent is a set of parameters randomly sampled from the radar reflectivity fusion probability field, which serves as the initial reflectivity field of the cell. Cellular automata then perform predictions and deductions based on the current state space to obtain the predicted reflectivity field, and simultaneously combine this with the first reward function to obtain the first reward. The discriminative agent then combines the second reward function to obtain the second reward, and weightedly merges the first and second rewards to generate the total reward; The training employs a course-based learning strategy, aiming to maximize total reward, until a preset training termination condition is met, at which point the training ends.

[0011] Furthermore, embodiments of the present invention also provide an artificial intelligence-based multi-radar reflectivity mosaic quality control system, including: The data processing module is used to acquire the original multidimensional radar dataset and model the original multidimensional radar dataset based on category theory to construct a radar state-radial network. An error analysis module is used to perform propagation error analysis on the radar state-to-ground network according to a data error analysis model, and generate a propagation error feature set. The probability analysis module is used to perform risk probability modeling based on the radar state-to-surface network and propagation error feature set, and to obtain a multi-radar reflectivity fusion probability field. The decision puzzle module is used to combine deep reinforcement learning with the multi-radar reflectivity fusion probability field to generate a multi-radar reflectivity puzzle.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The original multidimensional radar dataset is obtained, and modeled based on category theory to construct a radar state-radial network. This step provides a unified mathematical description framework for the original multidimensional radar data by using category theory modeling. At the same time, graph neural networks and meta-learning controllers are combined to improve the system's adaptability to different weather conditions and radar states, providing a high-quality data foundation for subsequent processes. Based on the preset data error analysis model, the propagation error analysis is performed on the radar state-to-ground network to generate a propagation error feature set. This step uses a physical information neural network to predict and correct atmospheric propagation errors, thereby improving the model's generalization ability and the physical rationality of the prediction results in complex or rare weather scenarios. Risk probability modeling is performed based on radar dynamic network and propagation error feature set to obtain multi-radar reflectivity fusion probability field. This step, through Bayesian inference and spatial consistency constraints, can obtain a globally coordinated probability solution while taking into account the credibility of each radar observation, providing a richer and more reliable data foundation for subsequent data fusion. Deep reinforcement learning is employed to simultaneously combine the multi-radar reflectivity fusion probability field to generate a multi-radar reflectivity mosaic. This step uses a reinforcement learning framework to encode user needs and physical laws into decision constraints, making the generation process of the multi-radar reflectivity mosaic clear in its target orientation and physically interpretable, thereby improving the credibility of the multi-radar reflectivity mosaic. Attached Figure Description

[0013] Figure 1 This is a flowchart of the steps of the multi-radar reflectivity mosaic quality control method based on artificial intelligence of the present invention; Figure 2 This is a flowchart of the steps in generating a radar reflectivity fusion probability field in the multi-radar reflectivity mosaic quality control method based on artificial intelligence of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating the steps of the multi-radar reflectivity mosaic quality control method based on artificial intelligence, as described in this invention. Figure 2 This is a flowchart of the steps in generating a radar reflectivity fusion probability field in the AI-based multi-radar reflectivity mosaic quality control method of the present invention. The following is a detailed introduction to the AI-based multi-radar reflectivity mosaic quality control method.

[0015] Step S1: Obtain the original multidimensional radar dataset, model the original multidimensional radar dataset based on category theory, and construct the radar state-radial network.

[0016] Specifically, a preset meta-learning controller is used to perform data feature analysis on the original multidimensional radar dataset to obtain feature analysis results. Based on the feature analysis results, a corresponding preprocessing operator sequence is selected for data preprocessing to obtain the initial multidimensional radar dataset.

[0017] Furthermore, graph neural networks and category theory are used to model the initial multidimensional radar dataset to generate a radar morphological network. The nodes of the radar morphological network represent the data source and processing module corresponding to each radar, while the edges represent the data flow and the corresponding transformation relationship. Based on a deep neural network, the data quality of the nodes in the radar morphological network is evaluated to generate corresponding data quality scores.

[0018] In one possible implementation, the radar observation range is defined. its object collection A set of morphisms containing all possible data state types. It includes all possible data transformation operations. Specifically, for each radar... Define its local category This category contains at least the original base data objects. This includes raw radar observations such as I / Q signals and pulse sequences; and preprocessed data objects. Represented as base data after processing such as pulse compression and clutter suppression; physical quantity data object Includes meteorological physical quantities such as reflectivity factor, radial velocity, and spectral width; quality control data objects. Includes data processed with quality control markings; coordinate system transformation data objects. Includes data transformed to a unified geographic coordinate system; where each object Defined by a triple, the triple can be represented as ,in For data value tensors, Metadata (including timestamps, spatial ranges, processing history, etc.) This is a vector of quality indicators.

[0019] The set of states must contain at least: pulse processing states Used to achieve the conversion from raw signal to base data; physical quantity inversion state projection It is used to calculate physical quantities such as reflectivity and velocity; mass control state emission. Used to characterize quality control algorithms; coordinate transformation state projection , used to transform to a unified coordinate system.

[0020] For example, for an S-band Doppler radar Its original data object If it contains a continuous I / Q sampling sequence, then the pulse processing state is determined according to the application of matched filtering and pulse pair processing. The output is a preprocessed object containing the power spectral density matrix and autocorrelation function. Subsequently, the state reversal was performed by combining physical quantities estimated through logarithmic transformation and velocity estimation. Generate reflectivity factor matrix and radial velocity matrix ,based on and Constituent objects ; In each state, metadata is updated while performing data processing. The processing history is analyzed, and the corresponding quality indicators are calculated. Such as signal-to-noise ratio, data consistency metrics, etc.

[0021] It should be noted that, in order to establish a mapping from multi-radar data to a unified mosaic, it is necessary to construct a framework that extends from the local scope of each radar to the global mosaic scope. letter Specifically, only the data after physical quantity inversion is mapped to the global mesh object, while the original data and intermediate processed data exist only in the local scope, i.e., only contain the following object mappings: ,in It is radar The observation covers a three-dimensional grid, and the grid point values ​​are physical quantities in the radar coordinate system. ,in It is quality-controlled radar grid data; ,in It is grid data after being transformed to a unified geographic coordinate system; while morphological mapping can be represented as Quality control operations are mapped to states in the global scope. ; The coordinate transformation mapping is to a morphism in the global domain. ,in The gridding used to implement the quality control algorithm includes at least noise filtering, ground clutter suppression, and distance attenuation correction. Corresponding coordinate transformation and resampling algorithms are used to convert radar polar coordinate data into a unified geographic grid. Furthermore, in order to coordinate the system differences between different radars, system bias correction is performed based on calibration parameters cross-validated between radars, and spatiotemporal alignment is performed to conduct observation geometry compensation and time synchronization adjustment. For example, the differences in radar beamwidth and elevation angle are standardized, and the scanning time of different radars is aligned.

[0022] After the category theory-based modeling process is completed, a graph neural network is used to define a standard directed acyclic graph for the entire multi-radar network. , where nodes Used to represent the data processing status, corresponding to objects within the category; edge Used to represent data processing operations, corresponding to morphisms in the category; weights This is used to represent the quality and efficiency indicators of data processing. Specifically, a multi-layer message-passing neural network is designed, with each layer containing at least the following update rules: For each node... In the Layer representation First, its aggregation of messages from neighbors can be represented as ,in The function typically takes the mean, maximum, or attention-weighted sum; secondly, its update node can be represented as... The function is implemented using a learnable neural network; for example, in radar data processing diagrams, a node representing reflectivity data quality control... The function learns how to dynamically adjust the state of a node based on current data quality characteristics and consistency information from neighboring radar nodes, in order to reflect the credibility of the fused data and the corrected data values.

[0023] For example, consider a network containing three radars. A graph neural network receives the following input features: node features: data quality metrics, timestamps, spatial coverage, system status flags; edge features: processing algorithm type, computational complexity, historical success rate; global features: network topology, communication latency, computational resource status. After three layers of graph convolution, each node can predict the optimal next processing node, processing algorithm parameters such as filter cutoff frequencies, and the expected output quality. For instance, when the radar... When a clutter suppression node detects a high-anomaly ground object echo, the graph neural network may suggest skipping the standard filter and using a more aggressive wavelet filter, and pass this decision to downstream nodes.

[0024] Receive real-time data streams from multiple radars, wherein the real-time data streams include at least raw base data, radar status parameters, and environmental information, combined with the directed acyclic graph. A radar state-to-state network is constructed. Specifically, for each radar, a corresponding sub-category node is instantiated based on its unique identifier. This node contains multiple sub-state nodes, namely, raw data nodes, preprocessing nodes, physical quantity inversion nodes, quality control nodes, and coordinate transformation nodes. Directed edge connections are established between the above sub-state nodes to form the data processing state-to-state chain within the radar. Based on the spatial topology between radars, beam coverage overlap areas, and data transmission links, coordination edges are established between the corresponding state nodes of different radars to form a cross-radar state-to-state network. The coordination edges are used to characterize the interaction relationships between nodes, such as data fusion, consistency constraints, and error propagation. All nodes and edges constitute the initial radar state-to-state network, where each node is assigned an initial feature vector, including static attributes such as radar hardware parameters, data timestamps, geographical location, and scanning mode, as well as dynamic features extracted from real-time data, such as signal-to-noise ratio and data coverage.

[0025] For example, suppose the system connects to three radars: R1 (S-band, located in an urban plain); R2 (C-band, located in a coastal area); and R3 (X-band, located in a mountainous area). After the real-time data from these three radars arrives, the system instantiates five corresponding sub-state nodes for each radar and obtains the radar deployment information: R1 and R2 are 120 kilometers apart, and their beam coverage partially overlaps; R3 is farther from R1 and R2 but at a higher altitude, and its beam covers the upper atmosphere. Since there is an overlapping area between R1 and R2 that needs to be fused, a coordination edge is established between R1 and R2 at the physical quantity inversion node and the quality control node. R3 mainly establishes connections with other radars through the coordinate transformation node. At the same time, the sub-state nodes within each radar are connected sequentially through directed edges to form their respective data processing chains.

[0026] After the radar dynamic network is constructed, the system uses a preset meta-learning controller to perform feature analysis on the original data nodes of each radar in the graph and adaptively selects the optimal preprocessing operator sequence. Specifically, firstly, feature extraction is performed on each original data node to generate a multi-dimensional feature vector. The multi-dimensional feature vector includes at least: signal-level features, such as the power spectral density and phase noise of I / Q signals; data-level features, such as the proportion of missing values ​​and outlier statistics; environmental-level features, such as current weather conditions and terrain obstruction coefficient; and graph structure features, such as the adjacency relationship and initial features of the node in the radar dynamic network. Subsequently, the meta-learning controller outputs a preprocessing operator sequence and corresponding parameters for the current data of the radar based on the multi-dimensional feature vector. Finally, based on the output operator sequence, preprocessing is performed on the original data to generate a standardized initial multi-dimensional radar dataset and update the state of the corresponding preprocessed nodes.

[0027] For example, for the raw data nodes of the mountain radar R3, feature analysis shows that the signal-to-noise ratio is low due to terrain clutter, and the data loss rate is high due to beam obstruction. In addition, the current weather is heavy precipitation. After receiving these features, the meta-learning controller refers to its meta-knowledge in similar scenarios and outputs a preprocessing operator sequence: first, an adaptive clutter suppression operator based on the terrain database is applied; then, a spatiotemporal interpolation operator is used to fill in the data loss; and finally, a precipitation attenuation correction operator is used. The system executes these operator sequences to generate preprocessed data and update the status of the corresponding nodes.

[0028] It should be noted that the meta-learning controller is represented as a neural network trained using model-independent meta-learning. Its core objective is to enable the controller to quickly adapt based on a small number of new scene samples, thereby allocating appropriate preprocessing operator sequences for different radar and meteorological conditions. Specifically, the training data consists of historical radar observation scenes, with each scene as an independent task. Each scene includes at least the original radar data, corresponding ideal preprocessing operator sequences generated by expert annotation or simulation, and a preprocessed performance evaluation. Each task is divided into a support set and a query set. The training process involves randomly sampling a batch of tasks in each iteration. For each task, the controller uses its current parameters as the initial point and performs a... One or more steps of gradient descent are used to obtain task-specific parameters. Then, the loss is calculated on the query set, and the initial parameters of the controller are meta-updated based on the loss, so that the initial parameters of the controller are adjusted in the direction of adapting to the new task. Its loss function includes sequence matching loss and data fidelity loss. The sequence matching loss is expressed as the difference between the operator sequence output by the controller and the expert-annotated sequence measured by cross-entropy. The data fidelity loss is expressed as the mean square error between the preprocessed data and the high-quality reference data. The training is carried out with the goal of minimizing the expected loss of all tasks on the query set until the average loss of the controller on unseen tasks decreases by less than a threshold or reaches the preset maximum number of training cycles in a series of consecutive training cycles.

[0029] After preprocessing, the system uses a pre-trained deep neural network to evaluate the data quality of each state node in the radar state-to-noise network, generating a quantified quality score. Specifically, for each state node in the radar state-to-noise network, such as the physical quantity inversion node, the following two types of features are extracted: first, the node's own features, including statistical features calculated from the data such as mean, variance, and signal-to-noise ratio, and physical rationality features such as reflectivity range and velocity range; second, graph context features, i.e., the state information of neighboring nodes obtained through the message passing mechanism of the graph neural network, such as the consistency of adjacent radar data and the quality indicators of upstream processing nodes. These features are input into a multi-task deep neural network. The network simultaneously outputs multiple quality dimension scores, such as signal-to-noise ratio (SNR) score, spatial consistency score, temporal continuity score, and physical rationality score. Each dimension score is normalized to the [0,1] interval using the sigmoid function. Finally, the system integrates these dimension scores according to preset weights to obtain the overall data quality score of the node. The preset weights are set by analyzing historical data. For example, the weight ratio of SNR score, spatial consistency score, temporal continuity score, and physical rationality score is 1:1:1:1, and the corresponding values ​​after normalization by the sigmoid function are 0.1, 0.3, 0.3, and 0.3, respectively. The final overall data quality score is 1.0.

[0030] It should be noted that the pre-training data of the deep neural network consists of historical multi-radar observation data and their corresponding manually labeled or automatically labeled quality tags with high confidence. These tags at least cover signal-to-noise ratio, spatial consistency, temporal continuity, physical plausibility, and processing confidence. The training process uses multi-task learning, which outputs predictions for multiple quality dimensions in parallel by sharing the underlying feature extraction network. The loss function is the sum of the weighted mean square error between the predicted value and the true label of each dimension, with the goal of minimizing the overall loss on the independent validation set. Training ends when the loss on the validation set no longer decreases significantly over multiple consecutive training cycles, such as when the fluctuation range does not exceed 0.02, and the correlation coefficient between the predicted value and the true label of each dimension reaches a preset threshold.

[0031] Step S2: Perform propagation error analysis on the radar state-to-ground network based on a preset data error analysis model to generate a propagation error feature set.

[0032] In this embodiment, step S2 includes: Step S2-1: Construct and train the data error analysis model.

[0033] Specifically, a data error analysis model is constructed based on a physical information neural network. This model is used to predict and correct propagation errors. A supervised learning strategy is adopted, and the data error analysis model is pre-trained simultaneously with a preset loss function.

[0034] In one possible embodiment, firstly, a model training dataset is constructed based on historical multi-radar synchronous observation data. Specifically, for each historical radar ray sample, starting from the radar, the actual propagation path of the ray in three-dimensional space is calculated according to the corresponding radar equation and the atmospheric ray model considering refraction. A series of key atmospheric parameters are extracted along this path as features, including at least the temperature, air pressure, specific humidity, and liquid water content at each point. Based on these features, derivative physical quantities such as the corrected refractive index N, refractive index gradient dN / dh, and specific attenuation coefficient K are further calculated to generate a high-dimensional feature vector describing the ray propagation environment. Each sample is then labeled. The labels include at least phase error and attenuation error. For phase error, inversion calculation is performed using dual radar or geometric cross-validation method of radar and GNSS. For attenuation error, it is estimated by comparing the observation differences of radars of different frequencies on the same path, such as comparing the difference between S-band and X-band, or by using polarization quantities that are not sensitive to attenuation, such as differential phase Kdp. In addition, under the condition of known fine atmospheric profile, a large amount of simulated "ground truth" error data is generated in combination with the corresponding electromagnetic simulation model to expand the training set, thereby covering extreme and rare weather conditions, such as using an electromagnetic simulation model based on vector radiative transfer theory to simulate extreme environments.

[0035] For example, to train a model to predict the attenuation error of S-band radar in warm cloud precipitation, a stratiform cloud precipitation event from historical data is selected. First, data from two nearby radars with different wavelengths simultaneously observing the same volume are extracted. For a specific ray, the temperature, humidity, and liquid water profiles along its path are extracted from the NWP field. By comparing the reflectivity differences between the two radars along this path, and simultaneously combining a dual-frequency attenuation correction algorithm, the total attenuation error of the S-band radar along this path is calculated. This forms a training sample with atmospheric feature vectors along the path as input and the calculated total attenuation error as the output label. At the same time, using the temperature, humidity, and liquid water profiles extracted from the NWP field along its path, the corresponding electromagnetic simulation model is run to generate another set of simulation samples with subtle perturbations, thereby enhancing the robustness of the model to parameter changes.

[0036] Next, a data error analysis model is constructed using a physical information neural network. Specifically, the physical information neural network adopts a deep residual structure, which contains eight fully connected layers, each containing 256 neurons, and uses the Swish activation function. Its input layer receives path state feature vectors, and the output layer is divided into two branches: a main branch and an uncertainty branch. The main branch is used to output the phase error. and attenuation error The predicted value is then output, and the uncertainty branch is used to output the corresponding prediction uncertainty. and .

[0037] For the main branch, the pre-training process is divided into two stages using a phased, course-based approach. Specifically, in the first stage, mean squared error is used as the loss function, aiming to minimize the loss between the network's predicted value and the supervision label. Pre-training is performed simultaneously using stochastic gradient descent and a learning rate decay strategy until the model's prediction error on the independent validation set no longer significantly decreases. If the fluctuation of the prediction error does not exceed a preset threshold, the model will have initially grasped the mapping relationship between atmospheric environmental characteristics and propagation error. In the second stage, a composite loss function is used. ,in, Huber loss is used to quantify the difference between the prediction error and the label; This can be represented as using automatic differentiation techniques to embed the partial differential equations controlling electromagnetic wave propagation as soft constraints into the training process. These constraints require that the output predictions and their derivatives with respect to input features approximately satisfy the differential equations describing electromagnetic wave propagation. For example, for attenuation error prediction... It will force the model to require that the total path attenuation predicted by the model and the liquid water content of the path integral satisfy a positive correlation that conforms to Mie scattering theory or empirical approximation; To control the strength of physical constraints, this stage forces the model to learn mapping relationships that conform to the underlying physical laws of the current domain by minimizing this composite loss and performing backpropagation, thereby ensuring the physical rationality of the prediction results.

[0038] For the uncertain branch, a Dropout operation is introduced in the last three layers of the fully connected layer, with a pre-set random dropout rate, such as maintaining a neuron random dropout rate of 0.2 during training. The output layer of this branch is designed with two independent channels to predict the phase error separately. and attenuation error Standard deviation and The total loss function used in its training is expressed as an addition of an uncertainty calibration loss to the loss function used in the second stage of training in the main branch. This uncertainty calibration loss is in the form of negative log-likelihood, and can be expressed as follows: ,in, Represented as the predicted true value, The mean of the predicted values, This is represented by the corresponding variance.

[0039] Finally, an early stopping method is adopted, which stops training when the total loss on the validation set decreases below a preset threshold over several consecutive training cycles to avoid overfitting.

[0040] Step S2-2: Perform error analysis based on the data error analysis model.

[0041] Specifically, a radar propagation path environment dataset is obtained. This dataset describes the environmental information along the radar's corresponding propagation path and includes at least temperature, air pressure, humidity, and wind vector data. Based on this dataset, the environment of each radar ray path is modeled to generate a path state feature vector.

[0042] The data error analysis model predicts propagation error based on path state feature vectors, generates an error correction field and an error sensitivity map, and performs error correction on the initial multidimensional radar dataset contained in the radar state-to-ground network based on the error correction field to obtain the error correction performance results.

[0043] The error correction field, error correction effectiveness results, and error sensitivity map are encapsulated into a propagation error feature set for output.

[0044] Understandably, the error correction field is generated by fusing the phase error correction field and the attenuation error correction field, and includes at least the corresponding error correction amount and uncertainty index; the error sensitivity map is generated by backpropagation to calculate the gradient of the propagation error on each dimension of the path state feature vector, and is used to quantify the contribution of environmental information on each propagation path in different regions to the propagation error.

[0045] In one possible implementation, three-dimensional ray tracing is performed for each volume scan, each radial ray, and each observation point corresponding to each range library for each radar at the current moment. For example, using the real-time analysis field or short-term forecast field of numerical weather prediction as the background atmosphere, considering the continuous changes in the curvature of the Earth and atmospheric refraction, the actual curved ray path from the radar antenna to the target point is calculated. Along the calculated real ray path, the atmospheric feature vector sequence that is completely consistent with the training phase is extracted from the NWP field at high resolution to generate a path state feature vector. For example, for an S-band radar located on a plain, when detecting a strong convective cell, the system first performs ray tracing based on the radar's current elevation angle, azimuth angle, and target distance, and simultaneously combines the atmospheric vertical profile provided by the real-time NWP. The analysis reveals that due to the strong superrefractive stratification in the lower layer, the ray path is significantly curved downwards compared to a straight line, resulting in the actual detection height of the radar beam being about 300 meters lower than the geometric height. At this time, features such as temperature, air pressure, humidity, and liquid water content are extracted along this curved path and fused to generate a multi-dimensional feature vector.

[0046] The path state feature vectors corresponding to all rays are input into the data error analysis model in batches. The main branch of the model outputs the phase error and attenuation error of each ray. The phase error represents the measurement deviation of the target in slant range and height caused by the change in the length and direction of the wave propagation path due to atmospheric refraction. The attenuation error represents the attenuation of radar echo power caused by absorption and scattering of atmospheric gases and cloud and rain particles, that is, the reflectivity is underestimated.

[0047] The uncertainty branch, during the forward inference phase of the model, performs Monte Carlo dropout sampling 10 times in parallel on the path state feature vector of each ray path. Specifically, in each forward propagation, a portion of neurons are randomly dropped, such as 20% of neurons, resulting in 10 sets of predictions. Due to the randomness of dropout, each propagation will generate a slightly different set of predictions. and the corresponding standard deviation estimate The mean of the predicted value set is used as the final predicted value, and the standard deviation is used as the prediction uncertainty. Furthermore, through numerical weather prediction ensemble dispersion analysis, the error covariance matrix of atmospheric parameters such as temperature and humidity is calculated, and the corresponding total uncertainty is obtained through the error propagation formula. Ultimately, the total uncertainty of each ray .

[0048] For example, after data from a certain band is processed by a model, the output mean is... , The parameter uncertainty caused by the discreteness of the NWP set is calculated through propagation. Then the total uncertainty The final output will be:

[0049] All obtained ray error prediction values , and its uncertainty , Mapped to three-dimensional space, each ray point corresponds to a spatial location including longitude, latitude, and altitude, along with a corresponding error value. A regular three-dimensional mesh field is generated using a physically constrained co-kriging interpolation method. Specifically, for each mesh point to be interpolated, all ray points within its neighborhood are found, such as within a 50-kilometer radius, and the following co-kriging equations are established: in, It is the covariance matrix between error values, and this matrix is ​​a The symmetric matrix, the The number of known ray points. Represented as the first The sample point and the first The covariance between error values ​​at each sample point; and These are the cross-covariance matrices between the error values ​​and the covariates; these two matrices describe the error... with covariates Spatial correlation between them It is an n×p matrix. yes Transpose, into a Matrix, where It is the number of covariates. The Middle Line 1 Column elements Represented as the first The error value at the nth sample point and the nth sample point The covariance of the vectors of each covariate at all ray points; It is the covariance matrix between covariates, and this matrix is A symmetric matrix, in which the first... Line 1 The element C_mm(k,l) in the column represents the... The first covariate and the first The covariance between covariates describes the spatial correlation structure between different atmospheric parameter fields, such as the spatial correlation structure between the refractive index gradient field and the liquid water content field. It is a weight vector. It is a Lagrange multiplier. and This is represented as the covariance vector between the point to be interpolated and the known point.

[0050] It should be noted that the covariates are the atmospheric parameters corresponding to the path state feature vector, such as the refractive index gradient field, liquid water content field, and convective available potential energy field from the NWP; the covariance function adopts an exponential model, i.e. ,in It is spatial distance. The relevant lengths, such as 30 kilometers horizontally and 1 kilometer vertically, are determined by solving the co-kriging equations to obtain the weights, and then the grid points... The error estimate is The estimated variance is ,in, and Represented as and The transposed row vector, for example, assuming there are 2 sample points and 1 covariate, yields the weight vector after solving the system of equations. Lagrange multipliers And the error measurement value corresponding to the sample point The covariate value at the point to be estimated is Then the error estimate of the point to be estimated is

[0051] Finally, by solving the phase error and attenuation error The corresponding co-kriging equations map the error and variance estimates of each grid point in the three-dimensional grid field, thereby constructing the corresponding phase error correction field and attenuation error correction field. The phase error correction field and attenuation error correction field together constitute the error correction field.

[0052] Each grid point in the error correction field corresponds to a radar observation point. Error correction is performed based on the values ​​at corresponding positions in the phase error correction field and the attenuation error correction field, and the corrected value is used as the result of the error correction effectiveness. Specifically, based on... Adjust the slope distance of the observation point and recalculate its geographic coordinates; the new slope distance... ,in This is the original slant range; similarly, the reflectivity is corrected based on the attenuation error correction field, and the corrected reflectivity is... ,in The original reflectance, and the total uncertainty of the corrected reflectance. ,in The overall data quality score obtained in step S1.

[0053] For example, if an observation point has an initial slant distance of 3.0 km, an initial reflectance of 45 dBZ, and an overall data quality score of 0.8, its value in the error correction field... =8.2dB, =0.6dB, =120 meters =5 meters, then the error correction performance results include the corrected slant distance of 2.88 km, reflectivity of 53.2 dBZ, and total uncertainty of 1.0.

[0054] It should be noted that since the data error analysis model is based on a physical information neural network, and physical information neural networks are differentiable, the gradient of the prediction error with respect to the given-dimensional environmental features within the input feature vector can be calculated through differentiation. , This involves normalizing the environmental features contained in the path state feature vector, such as specific humidity and liquid water content, to obtain the contribution percentage of each feature, thereby generating the corresponding error sensitivity map.

[0055] Step S3: Based on the radar state-to-surface network and propagation error feature set, perform risk probability modeling to obtain the multi-radar reflectivity fusion probability field.

[0056] Specifically, a set of target grid points from a multi-radar reflectivity mosaic is obtained, and a global probabilistic inference graph is constructed based on the spatial proximity relationships between all target grid points, with all target grid points in the set as nodes.

[0057] It should be noted that the nodes in the global probability inference graph are used to represent the reflectivity variable of the current target grid point, and each node corresponds to an observation likelihood function; the edge connection in the global probability inference graph corresponds to a potential function, which is used to encode the preset spatial continuity prior.

[0058] Furthermore, based on the global probabilistic inference graph, a belief propagation algorithm is used for approximate inference to generate a radar reflectivity fusion probability field, which is used to describe the probability distribution of all possible values ​​of nodes within the global probabilistic inference graph.

[0059] In one possible implementation, for each 3D grid point within the fusion target area, all radar observation data covering that point are collected. For each grid point... Suppose there is It was observed by radar from the 1st The radar observation provides a corrected reflectivity value. and the corresponding total uncertainty Based on this information, a local probability model is constructed for the grid point to describe the true reflectance value. In the absence of known circumstances, the probability of observing the current dataset, where each radar observation is treated as a Gaussian distribution component with mean ______. Standard deviation This directly reflects the uncertainty of the observation. The mixing weight w_k of each component integrates the reliability of the radar observation and is related to... Proportional; grid points The observation likelihood function can be expressed as

[0060] For example, suppose a grid point is observed by three radars R1, R2, and R3. After correction, R1 reports the reflectivity. Uncertainty dB; R2 reports reflectance dBZ, dB; R3 report reflectance dBZ, dB, then its initial weight is Therefore, the local probability model at this point is a mixture of the following three Gaussian distributions: The weights are 0.50, 0.28, and 0.22, respectively.

[0061] Construct a Markov random field to encode the spatial constraints between adjacent points, where each grid point It is a node whose state is the true reflectance value to be inferred. Nodes are connected by edges, such as connecting directly adjacent grid points in space, i.e., grid points in the vertical, horizontal, front, and back neighborhoods. A potential function is defined on the edges connecting the nodes; for example, a potential function that encourages spatial smoothness can be expressed as... ,in and They are adjacent nodes. It is a parameter that controls the smoothing intensity; this function is in and When the difference is small, the value is close to 1; when the difference is large, the value is close to 0. The corresponding potential function is defined according to the relationship between nodes, such as the potential function of meteorological prior knowledge like encoding vertical descent rate and zero-degree layer bright band characteristics.

[0062] For the set of all adjacent pairs of points The entire reflectivity field The joint probability distribution is proportional to The objective of the Markov random field is to find the joint posterior probability that maximizes this probability. The probability distribution.

[0063] For example, taking the grid point and its neighboring point to the east in the above case as an example, suppose the neighboring point is observed by another set of radars, and its local probability model tends to be 40.0 dBZ. Without spatial constraints, the two may independently take their respective highest probability values, say 35.0 dBZ and 40.0 dBZ, resulting in a sharp boundary of 5 dBZ. According to the potential function of the edge connecting the two points, this large difference will be penalized. Therefore, in the subsequent inference process, in order to obtain a high joint probability, the estimated values ​​of the two points will reach a consensus and eventually tend towards a smoother transition, such as taking values ​​of 37.0 dBZ and 38.0 dBZ respectively, so that the generated boundary is more in line with the characteristics of spatial continuity.

[0064] After constructing the Markov random field, the belief propagation algorithm is used for each node. The posterior edit probability distribution is calculated, and the belief propagation algorithm performs iterative message passing between nodes in the graph. Specifically, from node... Send to neighboring nodes News The nodes are summarized. From its own observation and All information obtained from other neighbors, and regarding The possible values ​​of the message are suggested. At the beginning of the iteration, all messages are initialized to a uniform distribution. Subsequently, in each iteration, each node updates its own belief (the currently estimated posterior distribution) based on the incoming messages from all neighbors and its own local likelihood, and calculates new outgoing messages to send to all neighbors. After multiple iterations, the messages gradually converge. After convergence, each node's belief... It is its approximate posterior marginal probability distribution. By combining the posterior distributions of all grid points, a complete three-dimensional probability field is formed. This three-dimensional probability field is the multi-radar reflectivity fusion probability field.

[0065] It should be noted that the formula for message updates can be expressed as: ; in, Represented as nodes remove All neighbors except those mentioned above; It can be represented as: .

[0066] Step S4: Deep reinforcement learning is used to simultaneously combine the multi-radar reflectivity fusion probability field to create a mosaic, generating a multi-radar reflectivity mosaic.

[0067] In this embodiment, step S4 includes: Step S4-1: Build and train an adversarial reinforcement learning environment.

[0068] Specifically, an adversarial reinforcement learning environment is constructed based on cellular automata, which includes generative agents and discriminative agents. The cellular automata are used to predict the state of a cell at the next moment based on the current state of the cell itself and its neighboring cells. The state of the cell includes adjustable variables and background variables.

[0069] It should be noted that the state space of the adversarial reinforcement learning environment is composed of the radar reflectivity fusion probability field and the corresponding environmental feature vector; the action space of the generated agent is a set of parameters randomly sampled from the radar reflectivity fusion probability field as the initial reflectivity field of the cell.

[0070] The cellular automaton performs prediction and deduction based on the current state space to obtain the predicted reflectivity field. It then combines the first reward function to obtain the first reward. The discriminative agent combines the second reward function to obtain the second reward. The first and second rewards are weighted and fused to generate the total reward. The weight coefficients of the first and second rewards are preset weights, and the initial values ​​of these weights are obtained after training with historical data.

[0071] Furthermore, a course-based learning strategy is employed to train with the goal of maximizing total reward until the preset training termination condition is met, at which point the training ends.

[0072] In one possible embodiment, a radar reflectivity evolution simulation environment based on preset rules is constructed as a training and decision-making scenario for reinforcement learning agents. The simulation environment can discretize the three-dimensional space into a regular cellular grid. The state of each cell includes reflectivity value and physical quantities such as liquid water content, water vapor mixing ratio, vertical velocity, and temperature perturbation. The preset rules include at least: mass conservation rules, such as maintaining the balance of water vapor and liquid water flux between adjacent cells to ensure that the total mass remains unchanged; convection development rules, such as triggering convection development when conditional instability exceeds a preset threshold and vertical velocity reaches a preset critical value, with reflectivity increasing according to empirical relationships extracted from historical data; particle interaction rules, such as simulating the growth and falling of raindrops based on gravitational settling velocity and collision efficiency parameters; and cell transport rules, such as transporting water substances from one cell to a downwind cell according to the environmental wind field. The simulation environment can perform multi-step iterations based on a given initial reflectivity field and the preset rules to simulate the short-term evolution process of the reflectivity field.

[0073] The environment-generated agent is constructed using a deep reinforcement learning algorithm based on an actor-critic framework. Its observation space includes: the multi-radar reflectivity fusion probability field of the current state (i.e., the mean and variance of the reflectivity probability distribution for each grid point), the environmental physical state (such as temperature, humidity, and wind field data obtained from the NWP field), and the task objective; the action space is represented as the action for grid point i. ,in This indicates the position selected from the quantiles of the probability distribution. This represents the fine-tuning amount based on the selected value; the agent's neural network adopts an encoder-decoder structure, where the encoder part is a three-dimensional convolutional neural network used to extract the spatial features of the entire probability field and physical field; the decoder part is divided into two branches: the actor network and the critic network. The actor network branch outputs the action distribution of each grid point, i.e., the mean and variance, while the critic network branch outputs the value estimate of the current state.

[0074] For example, in a strong convection monitoring task, the agent's observations may include the three-dimensional probability field of the entire region as well as physical fields such as the available convective potential energy field and the vertical wind shear field. The encoder extracts multiple feature maps based on the above data. Suppose that some feature maps highlight areas with large available convective potential energy values ​​and high probability field variance. Then, the actions output by the actor network in these areas may tend to select high quantiles in the probability distribution and give positive fine-tuning to enhance the convective signal; while in stable regions, values ​​close to the median are selected and negative fine-tuning is performed to suppress false echoes.

[0075] The reward function for generating the intelligent agent can be expressed as follows: ,in, Represented as a physical consistency reward, it is used to measure the consistency of the initial field chosen by the agent after the simulation environment evolves. For example, it can be used to calculate the relative change in the total amount of liquid water in the entire field. The smaller the value, the higher the reward. Or it can be used to calculate the MS-SSIM between the initial reflectivity field and the evolved reflectivity field. The larger the value, the higher the reward. This is represented as a data probability matching reward item. For example, for each grid point, the log-likelihood of the selected value in its corresponding posterior probability distribution within the multi-radar reflectivity fusion probability field is calculated, and the negative mean is taken. That is, the closer the selected value is to the peak of the probability distribution, the smaller the penalty and the higher the reward. and The weighting coefficient for the corresponding reward item is assigned based on the analysis of historical data, and the value of this reward function is the first reward.

[0076] The observation space of the discriminative agent includes: the reflectivity field generated by the agent, the multi-radar reflectivity fusion probability field, and the corresponding environmental physical state; the action space consists of the output physical rationality score and the task suitability score; its reward function is the weighted sum of the physical rationality score and the task suitability score, thereby obtaining a second reward.

[0077] It should be noted that the physical rationality score is obtained based on a deep binary classification network. The deep binary classification network uses real reflection field data from historical data as positive samples and predictive reflection field data inferred by the agent based on the data corresponding to the positive samples as negative samples. By using contrastive learning technology, the deep binary classification network is trained to learn to distinguish the differences between positive and negative samples in terms of spatial structure, texture, and statistical features. Iterative training is performed simultaneously with the corresponding loss function. After training, the deep binary classification network can output a normalized scalar value, which is the physical rationality score.

[0078] The task suitability score is obtained through inference from multiple parallel lightweight convolutional sub-networks within the discriminative agent. Each sub-network is specifically trained for a particular task, such as strong convection identification, stratiform cloud precipitation estimation, and aviation turbulence warning. Specifically, each sub-network is a regression or classification model responsible for extracting feature patterns highly relevant to the specific task from the input field and outputting a suitability score. These lightweight convolutional sub-networks utilize high-quality labeled data from a large amount of task data, such as expert-annotated convective cell regions and rain gauge-calibrated precipitation fields. Supervised training is performed to learn task-related discriminative features. For categorical tasks such as identifying strong convective regions, the cross-entropy loss function is used. For regression tasks such as quantitative precipitation estimation, Huber loss is used as the loss function. The training aims to minimize the loss on the validation set. After training, the lightweight convolutional subnetwork can extract feature patterns and normalize the fit value with the current task. Users set the corresponding business tasks and their weights according to the current monitoring task requirements, and the weighted fit value of the corresponding subnetwork is used as the task fit score.

[0079] For example, if a user only focuses on the goal of strong convection identification, the weight of the strong convection identification subnetwork can be set to 1, and the rest to 0. In this case, the task suitability score is equal to the output score of the strong convection identification subnetwork. If the user focuses on both strong convection and aviation turbulence, the weight ratios can be set to 0.6, 0.0, and 0.4, corresponding to strong convection, stratiform cloud precipitation, and aviation turbulence identification, respectively. In this case, the task suitability score is equal to the weighted sum of the output scores of the strong convection and aviation turbulence identification subnetworks.

[0080] Step S4-2: Generate a multi-radar reflectivity mosaic based on the adversarial reinforcement learning environment.

[0081] Specifically, the adversarial reinforcement learning environment is initialized based on the multi-radar reflectivity fusion probability field. The generated agent and the discriminative agent engage in multiple rounds of adversarial game based on the current environment state to generate an initial multi-radar reflectivity fusion field. Regional feature analysis is performed on the initial multi-radar reflectivity fusion field to obtain the regional feature analysis results.

[0082] Furthermore, based on the regional feature analysis results, the initial multi-radar reflectivity fusion field is locally enhanced to generate an optimal multi-radar reflectivity fusion field, which is the multi-radar reflectivity mosaic.

[0083] In one possible embodiment, two state information items are acquired: a multi-radar reflectivity fusion probability field and user task instructions. Based on the state information, the corresponding parameters of the adversarial reinforcement learning environment are initialized. An agent is generated to output the action distribution of each grid point according to the state space of the current environment. The action is a two-dimensional vector: the first dimension is the quantile selection parameter. The first dimension indicates which quantile of the probability distribution corresponding to that point is used to select the base value; the second dimension is the fine-tuning amount. This indicates the adjustment range based on the base value; for each grid point, the system adjusts according to... The initial reflectivity field is generated by finding the corresponding quantile value from its probability distribution function as the base reflectivity and then superimposing the corresponding fine-tuning amount. The initial reflectivity field is generated by evaluating the initial reflectivity field and outputting the corresponding physical rationality score and task suitability score. The two are then weighted and fused into a second reward. The generated initial reflectivity field is input into the current environment for evolution and combined with the first reward function to obtain the corresponding first reward. The first and second rewards are weighted and fused to generate a total reward. If the total reward is lower than a preset threshold, it indicates that the current strategy is not good. The generating agent will update its actor network parameters based on the current advantage function through the policy gradient. The critic network will evaluate the value of the current state and guide the update direction. Starting from the current state, the cycle of decision-making, generation, simulation, evaluation, and updating is repeated for several iterations until the total reward reaches the preset threshold or the maximum number of iterations is reached. The final initial reflectivity field is used as the initial multi-radar reflectivity fusion field.

[0084] First, the initial multi-radar reflectivity fusion field, physical simulation evolution results, and evaluation output of the discriminative agent are analyzed. A lightweight regional importance assessment network is used to calculate the enhancement priority score of each local region in real time. This score is obtained by weighting task relevance, physical inconsistency, and probabilistic conflict. Task relevance represents the importance of a region in the user-specified task, such as urban areas versus suburban areas; physical inconsistency represents the degree of change in the physical quantities of the region before and after simulation evolution; and probabilistic conflict represents the variance of the probability distribution of multi-radar observations within the region. High-priority regions are selected based on this enhancement priority score.

[0085] Subsequently, a local optimization loop is initiated for high-priority regions. Specifically, the generating agent maintains its global network parameters unchanged, but activates an additional, higher-capacity local policy head for a specified region. This local policy head represents a lightweight neural network of the generating agent, used to receive global features and high-resolution contextual information of the local region, and output more refined and aggressive action adjustments for that region, such as allowing a larger range of fine-tuning or more flexible quantile selection parameters. The simulation environment performs local evolution on the corresponding region within the initial multi-radar reflectivity fusion field based on the action adjustments, and simultaneously obtains the total reward value corresponding to that region. Only optimizations that reach a preset optimization threshold are accepted. The preset optimization threshold is obtained based on the analysis of historical data, until all high-priority regions have completed local optimization.

[0086] It should be noted that the local optimization loop only performs a limited number of loop operations for each high-priority region. After reaching the preset number of loops, if the local optimization loop operation does not improve the reflectivity of the current region, the value of the region corresponding to the initial multi-radar reflectivity fusion field is directly adopted; if the reflectivity of the current region is improved, the reflectivity value that makes the region have the highest total reward value during the loop process is adopted as the optimized reflectivity value of the current region.

[0087] Finally, the initial multi-radar reflectivity fusion field after local optimization loop is taken as the optimal multi-radar reflectivity fusion field, which is the multi-radar reflectivity mosaic.

[0088] Specifically, the AI-based multi-radar reflectivity mosaic quality control system includes: The data processing module is used to acquire the original multidimensional radar dataset and model the original multidimensional radar dataset based on category theory to construct a radar state-radial network.

[0089] An error analysis module is used to perform propagation error analysis on the radar state-to-ground network based on a data error analysis model, and generate a propagation error feature set.

[0090] The probability analysis module is used to perform risk probability modeling based on the radar state-to-surface network and propagation error feature set, and to obtain a multi-radar reflectivity fusion probability field.

[0091] The decision puzzle module is used to combine deep reinforcement learning with the multi-radar reflectivity fusion probability field to generate a multi-radar reflectivity puzzle.

[0092] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A multi-radar reflectivity mosaic quality control method based on artificial intelligence, characterized in that, It includes the following steps: Obtain the original multidimensional radar dataset, model the original multidimensional radar dataset based on category theory, and construct the radar state-radial network; Based on a preset data error analysis model, the propagation error analysis is performed on the radar state-to-ground network to generate a propagation error feature set. Risk probability modeling is performed based on radar reflectivity network and propagation error feature set to obtain multi-radar reflectivity fusion probability field; Deep reinforcement learning is used to simultaneously combine the multi-radar reflectivity fusion probability field to generate a multi-radar reflectivity mosaic.

2. The multi-radar reflectivity mosaic quality control method based on artificial intelligence according to claim 1, characterized in that, Obtain the original multidimensional radar dataset, model the original multidimensional radar dataset based on category theory, and construct a radar state-radius network, including: A preset meta-learning controller is used to perform data feature analysis on the original multidimensional radar dataset to obtain feature analysis results; Based on the feature analysis results, the corresponding preprocessing operator sequence is selected for the original multidimensional radar dataset to perform data preprocessing and obtain the initial multidimensional radar dataset. Graph neural networks and category theory are used to model the initial multidimensional radar dataset to generate a radar state-to-beam network. The nodes of the radar state-to-ground network are represented by the data source and processing module corresponding to each radar, while the edges represent the data flow and the corresponding transformation relationship. Data quality is evaluated for nodes within the radar dynamic network based on a deep neural network, and corresponding data quality scores are generated.

3. The multi-radar reflectivity mosaic quality control method based on artificial intelligence according to claim 1, characterized in that, Based on a preset data error analysis model, propagation error analysis is performed on the radar dynamic network to generate a propagation error feature set, including: Obtain a radar propagation path environment dataset, which is used to describe the environmental information on the radar's corresponding propagation path and includes at least temperature, air pressure, humidity, and wind vector data. Based on the radar propagation path environment dataset, the environment of each radar ray path is modeled to generate a path state feature vector. The data error analysis model predicts propagation error based on path state feature vectors, generating an error correction field and an error sensitivity map. Based on the error correction field, the initial multidimensional radar dataset contained in the radar state-to-ground network is corrected for errors, and the error correction performance results are obtained. The error correction field, error correction effectiveness results, and error sensitivity map are encapsulated into a propagation error feature set for output.

4. The multi-radar reflectivity mosaic quality control method based on artificial intelligence according to claim 3, characterized in that, The method further includes: A data error analysis model is constructed based on a physical information neural network. This data error analysis model is used to predict and correct propagation errors. A supervised learning strategy is adopted, and a pre-set loss function is simultaneously combined to pre-train the data error analysis model; The error correction field is generated by fusing the phase error correction field and the attenuation error correction field, and includes at least the corresponding error correction amount and uncertainty index. The error sensitivity map is generated by backpropagation to calculate the gradient of the propagation error on each dimension of the path state feature vector, and is used to quantify the contribution of environmental information on each propagation path in different regions to the propagation error.

5. The multi-radar reflectivity mosaic quality control method based on artificial intelligence according to claim 1, characterized in that, Risk probability modeling is performed based on radar reflectivity networks and propagation error feature sets to obtain a multi-radar reflectivity fusion probability field, including: Obtain the target grid point set of the multi-radar reflectivity mosaic, and construct a global probabilistic inference graph based on the spatial proximity relationship between each target grid point, using all target grid points in the target grid point set as nodes. The nodes in the global probability inference graph are used to represent the reflectivity variables of the current target grid point, and each node corresponds to an observation likelihood function. The inner edge connection of the global probability inference graph corresponds to a potential function, which is used to encode the preset spatial continuity prior. Based on the global probability inference graph, a belief propagation algorithm is used to perform approximate inference to generate a radar reflectivity fusion probability field. The radar reflectivity fusion probability field is used to describe the probability distribution of all possible values ​​of nodes within the global probabilistic inference graph.

6. The multi-radar reflectivity mosaic quality control method based on artificial intelligence according to claim 1, characterized in that, Deep reinforcement learning is used to simultaneously combine the multi-radar reflectivity fusion probability field to generate a multi-radar reflectivity mosaic, including: An adversarial reinforcement learning environment, comprising generative and discriminative agents, is constructed based on cellular automata and trained using a curriculum learning strategy. The adversarial reinforcement learning environment is initialized based on the multi-radar reflectivity fusion probability field, and the generated agent and the discriminative agent engage in multiple rounds of adversarial game based on the current environment state to generate the initial multi-radar reflectivity fusion field. Perform regional feature analysis on the initial multi-radar reflectivity fusion field to obtain the regional feature analysis results; Based on the regional feature analysis results, the initial multi-radar reflectivity fusion field is locally enhanced to generate the optimal multi-radar reflectivity fusion field, which is the multi-radar reflectivity mosaic.

7. The multi-radar reflectivity mosaic quality control method based on artificial intelligence according to claim 6, characterized in that, An adversarial reinforcement learning environment based on cellular automata, comprising generative and discriminative agents, is constructed and trained using a curriculum learning strategy, including: The cellular automaton is used to predict the state of a cell at the next moment based on the current state of the cell itself and its neighboring cells. The state of the cell includes adjustable variables and background variables. The state space of an adversarial reinforcement learning environment consists of a radar reflectivity fusion probability field and the corresponding environmental feature vector. The action space of the generated agent is a set of parameters randomly sampled from the radar reflectivity fusion probability field, which serves as the initial reflectivity field of the cell. Cellular automata then perform predictions and deductions based on the current state space to obtain the predicted reflectivity field, and simultaneously combine this with the first reward function to obtain the first reward. The discriminative agent then combines the second reward function to obtain the second reward, and weightedly merges the first and second rewards to generate the total reward; The training employs a course-based learning strategy, aiming to maximize total reward, until a preset training termination condition is met, at which point the training ends.

8. An AI-based multi-radar reflectivity mosaic quality control system, used to implement the method described in any one of claims 1 to 7, characterized in that, include: The data processing module is used to acquire the original multidimensional radar dataset and model the original multidimensional radar dataset based on category theory to construct a radar state-radial network. An error analysis module is used to perform propagation error analysis on the radar state-to-ground network according to a data error analysis model, and generate a propagation error feature set. The probability analysis module is used to perform risk probability modeling based on the radar state-to-surface network and propagation error feature set, and to obtain a multi-radar reflectivity fusion probability field. The decision puzzle module is used to combine deep reinforcement learning with the multi-radar reflectivity fusion probability field to generate a multi-radar reflectivity puzzle.

Images