An intelligent identification method and system for cloud seeding conditions of a cold cloud aircraft

Abstract

The application belongs to the field of electric digital data processing, and particularly relates to a cold cloud aircraft cloud seeding condition intelligent identification method and system, which comprises the following steps: obtaining detection data and performing preprocessing; embedding position coding and season coding on the basis of the preprocessed data, obtaining embedded feature data through an activation layer and a discard layer; establishing an adjacent matrix of a graph structure according to the dimension of the embedded feature data, taking a parameter as a node of the graph structure, traversing the embedded feature data, updating an edge weight and aggregating a node vector at each time, and obtaining spatial feature data after traversing the sample; traversing the embedded feature data, jointly inputting the embedded feature data and the spatial feature data at each time into a long short-term memory unit to obtain a time feature representation at the time, and outputting time feature data after traversing the complete sample; aggregating the time feature data and the spatial feature data through a feature aggregation unit to obtain spatiotemporal feature data; and classifying the spatiotemporal feature data based on a classifier to obtain a discrimination result.

Description

Technical Field

[0001] This application belongs to the field of electronic digital data processing technology, and specifically relates to a method and system for intelligent identification of cold cloud aircraft cloud seeding conditions. Background Technology

[0002] Weather modification, as an important means of disaster prevention and mitigation, plays a positive role in ensuring agricultural production and water resource management. Global research on artificial rain enhancement technology shows that this technology can increase precipitation by an average of about 15%. Cold clouds are currently the main targets for artificial rain enhancement operations. The principle is based on the Bergeron effect, which states that in sub-zero cloud layers, when ice crystals and supercooled water droplets coexist, the difference in supersaturated vapor pressure between the ice and water surfaces causes water droplet evaporation and ice crystal sublimation, promoting precipitation. Therefore, seeding supercooled layers with silver iodide as artificial ice nuclei is an effective method to promote precipitation, and its scientific basis has been confirmed by numerous indoor experiments, numerical simulations, and field tests. The latest assessment of global rain enhancement activities by the World Meteorological Organization's Specialist Group on Weather Modification indicates that aircraft cloud seeding technology is generally more effective than other methods such as ground generators, customized rockets, and artillery shells, offering a wider operational range and more precise catalyst delivery, thus receiving widespread attention.

[0003] Accurately identifying suitable cloud layers for seeding is a core prerequisite for improving the effectiveness of weather modification operations. Cold-layer seeding requires finding suitable supercooled water areas, i.e., cloud layers considered suitable for seeding. Currently, determining suitable cloud seeding conditions mainly relies on professionals observing from meteorological aircraft in real time. This involves integrating multi-source information such as aircraft observation data, satellite cloud images, radar echoes, and numerical weather prediction products, combined with historical seeding experience, and using statistical methods for manual judgment. Although this data is directly used in the command and control of weather modification operations in my country, much of the process still depends on human judgment.

[0004] In recent years, with the rapid development of machine learning technology, some research has begun to explore data-driven methods to achieve automatic identification of suitable cloud layers for seeding, providing a new approach to the objectification and intelligentization of decision-making in this field. Long Short-Term Memory (LSTM) networks, through their gating mechanism, capture long-term dependencies and short-term fluctuations in time series, enabling robust modeling of the evolution patterns of probe data during flight. However, existing discrimination methods have the following problems: 1. Discrimination methods relying on expert experience are easily affected by human factors in complex meteorological scenarios, leading to insufficient reliability and consistency in the delineation of rain enhancement potential areas, making it difficult to meet the needs of large-scale, accurate, and timely operations. 2. Discrimination methods relying on machine learning models still have significant shortcomings in effectively modeling the complex spatial topological relationships and continuously evolving temporal dynamics within cloud systems. In particular, the complex phase states, spatial distribution, and interaction mechanisms between particles within clouds and the environmental field further increase the difficulty of operational condition discrimination. Summary of the Invention

[0005] This application is proposed based on the aforementioned needs of the prior art. The technical problem to be solved by this application is to provide a method and system for intelligent identification of cold cloud aircraft cloud seeding conditions in order to improve the identification accuracy and meet the needs.

[0006] To address the above problems, the technical solution provided in this application includes: A method for intelligent identification of cold cloud conditions by aircraft is provided, comprising: S1 acquiring detection data and preprocessing the detection data to obtain preprocessed data; S2 embedding location and seasonal codes into the preprocessed data, and obtaining embedded feature data through an activation layer and a dropout layer; S3 establishing an adjacency matrix of a graph structure based on the dimensions of the embedded feature data, with parameters serving as nodes of the graph structure, traversing the embedded feature data, updating edge weights and aggregating node vectors at each time step to obtain spatial feature representation, interspersed with dropout layers and residual connections, and obtaining spatial feature data after traversing all samples; S4 traversing the embedded feature data, inputting the embedded feature data and spatial feature data at each time step into a long short-term memory unit to obtain the temporal feature representation at that time step, and outputting temporal feature data after traversing all samples; S5 aggregating the temporal feature data and spatial feature data through a feature aggregation unit to obtain spatiotemporal feature data; S6 classifying the spatiotemporal feature data based on a classifier to obtain a discrimination result.

[0007] Preferably, the preprocessing of the probe data to obtain preprocessed data includes dividing the data into fixed time windows, creating a mask for the default values ​​after Z-score normalization, and copying the results on the channels to increase the dimensionality to obtain the preprocessed data.

[0008] Preferably, the selection of the fixed time window is based on prior knowledge of the typical timescales of cloud microphysical processes.

[0009] Preferably, after partitioning, the data is Z-score normalized, as follows: , in This represents the value after normalization. Represents the original data. and These are the mean and standard deviation of the feature calculated on the training set, respectively. Normalization aims to eliminate the influence of different physical dimensions and numerical ranges, making model training more stable. μ and σ are fixed and saved as normalization parameters and applied to new data during the model inference stage to ensure the consistency of data distribution between the training and inference stages.

[0010] Preferably, the location encoding method is to add location encoding information to the data at each time point by incrementing from 1 according to the time window divided in the preprocessing module.

[0011] Preferably, the seasonal encoding method is to initialize it to 0000, representing the four seasons of spring, summer, autumn and winter in sequence, and set the value at the corresponding position of the detected season of the data sample to 1.

[0012] Preferably, the step of traversing the embedded feature data and updating the edge weights and aggregating the node vectors at each time step includes traversing a sample data in positional encoding order, and in each iteration, calculating the attention weights between the sample and all its neighboring nodes, expressed as: in, This represents the attention weights for nodes u and v. and Let u and v represent the eigenvalues ​​of node u and node v at time t, respectively. and It is a linear transformation matrix. Learnable attention vector The transpose of the matrix, This indicates a feature concatenation operation. It is a function.

[0013] Preferably, the step of traversing the embedded feature data and updating the edge weights and aggregating the node vectors at each time step further includes performing aggregation updates based on the attention weights and edge weights of the node and its neighboring nodes. The expression for the node update is: ,in, Let M(u) be the updated node feature value, M(u) represent the set of neighboring nodes of node u, σ(·) be the non-linear activation function, and W represent the edge weight.

[0014] Preferably, the classifier includes two linear layers and one activation layer. The first linear layer maintains the feature dimensions of the input and output unchanged and is activated by an activation function, outputting to the second linear layer. The output of the second linear layer is two categories: "suitable for the task" and "unsuitable for the task". The cross-entropy loss function is used to optimize the cloud microphysical feature extraction effect. The cross-entropy loss function is expressed as: ,in, Here, N represents the loss function, and N represents the number of classes. This represents the true label of the i-th category. This represents the probability that the sample belongs to the i-th category, which is the probability that it is suitable for the task, where i is the category number.

[0015] A cold cloud aircraft cloud seeding condition intelligent recognition system is also provided, comprising: a data acquisition and preprocessing module for acquiring detection data and preprocessing the detection data to obtain preprocessed data; a data embedding module for embedding location and seasonal codes on the preprocessed data, and obtaining embedded feature data through activation and discard layers; a spatial feature extraction module for establishing an adjacency matrix of a graph structure based on the dimensions of the embedded feature data, with parameters as nodes of the graph structure, traversing the embedded feature data, updating edge weights and aggregating node vectors at each time step to obtain a spatial feature representation, interspersed with discard layers and residual connections, and obtaining spatial feature data after traversing all samples; a temporal feature extraction module for traversing the embedded feature data, inputting the embedded feature data and spatial feature data at each time step into a long short-term memory unit to obtain a temporal feature representation at that time step, and outputting temporal feature data after traversing all samples; a feature aggregation module for aggregating temporal and spatial feature data through a feature aggregation unit to obtain spatiotemporal feature data; and a classifier for classifying the spatiotemporal feature data based on the classifier to obtain a discrimination result.

[0016] Compared with existing technologies, this application enhances the model's adaptability to different time periods and regional climate backgrounds by embedding the seasonal characteristics of macroclimate into the preprocessed data. A graph attention layer is introduced, using an attention mechanism to adaptively learn and detect irregular spatial dependencies between parameters, improving the model's spatial representation capability for multi-probe observation systems. Dropout layers and residual connections are interspersed during the traversal of complete samples to alleviate excessive smoothing of information during propagation, enhancing the model's ability to capture key microphysical processes of cloud particles, thereby achieving more accurate and reliable intelligent identification of cloud propagation conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart illustrating the steps of an intelligent identification method for cold cloud aircraft cloud seeding conditions in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a cold cloud aircraft cloud seeding condition intelligent identification system according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the specific processing procedure of S1 in Embodiment 1 of this application; Figure 4 This is a schematic diagram illustrating the specific processing procedure of S2 in Embodiment 1 of this application; Figure 5 This is a schematic diagram illustrating the specific processing steps of S3 and S4 in Embodiment 1 of this application; Figure 6 This is a schematic diagram of the specific processing procedure of S5 in Embodiment 1 of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In the description of the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, the term "connected" should be interpreted broadly. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0021] Throughout the text, the terms “top,” “bottom,” “above,” “below,” and “on top” refer to the relative positions of components of the device, such as the relative positions of the top and bottom substrates within the device. It is understood that the device is multifunctional and independent of its spatial orientation.

[0022] To facilitate understanding of the embodiments of this application, the following will provide further explanation and description with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of this application.

[0023] Example 1 This embodiment provides a method for intelligent identification of cold cloud aircraft cloud seeding conditions, such as... Figure 1 As shown.

[0024] The intelligent identification method for cold cloud aircraft cloud seeding conditions includes: S1 acquires the detection data and preprocesses it to obtain preprocessed data, such as... Figure 3 As shown.

[0025] The detection data includes second-level airborne detection data acquired from multiple sources.

[0026] The preprocessing includes dividing the data into fixed time windows, creating a mask for the default values ​​after Z-score normalization, and then copying the results across channels to increase the dimensionality, resulting in preprocessed data. The preprocessed data is then adapted to the input format of subsequent steps.

[0027] The selection of a fixed time window is based on prior knowledge of the typical time scales of cloud microphysical processes, ensuring that it can both capture the rapid changes in clouds and form a statistically significant analytical unit.

[0028] Specifically, the airborne detection data is stored in tabular form, with each column representing the parameters detected by the aircraft and each row containing the data detected per second. Specifically, the parameters used include longitude, latitude, altitude, temperature, relative humidity, atmospheric pressure, wind speed, liquid water content, cloud droplet number concentration, precipitation number concentration, cloud condensation nucleus number concentration, and large cloud droplet number concentration—a total of 12 data parameters. The raw data can be continuous or discontinuous during detection; a 1-minute time window is used as an example for division, and the data is then normalized using Z-score. The normalization formula is: in This represents the value after normalization. Represents the original data. and These are the mean and standard deviation of the feature calculated on the training set, respectively. Normalization aims to eliminate the influence of different physical dimensions and numerical ranges, making model training more stable. μ and σ are fixed and saved as normalization parameters and applied to new data during the model inference stage to ensure the consistency of data distribution between the training and inference stages.

[0029] Establish 60 The system uses a table of 12 and sequentially masks the probe data, creating mask markers for discontinuous areas. In the mask matrix, valid data locations are marked with 1, and missing / invalid data locations are marked with 0. The masking mechanism is crucial for handling discontinuous probe data in this system; it clearly informs the model of missing data locations, preventing the model from learning or propagating invalid values.

[0030] After the repeat expansion layer, the padded data is copied 4 times in the channels to obtain preprocessed data. This operation aims to provide a richer initial representation space for subsequent feature extraction by increasing the channel dimension, and to provide more transformation possibilities for subsequent convolution or attention mechanisms.

[0031] S2 embeds positional and seasonal codes into the preprocessed data, and then passes it through activation and dropout layers to obtain embedded feature data. For example... Figure 4 As shown.

[0032] The location encoding method involves incrementing the position code from 1 according to the time window defined in the preprocessing module, adding location code information to the data at each time point. The seasonal encoding method initializes to 0000, representing the four seasons: spring, summer, autumn, and winter. The value at the corresponding position is set to 1 based on the season in which the data sample was detected. For example, spring data is encoded as [1,0,0,0], and summer data as [0,1,0,0]. This encoding is concatenated with the features at each time point to provide the model with seasonal climate background information. After feature embedding, the data passes through an activation function and a dropout layer to obtain the feature embedding data. The dropout layer randomly sets the output of some neurons to zero during training to prevent overfitting. The activation function is the ReLU activation function.

[0033] S3 constructs an adjacency matrix of a graph structure based on the dimensions of the embedded feature data. The parameters serve as nodes in the graph structure. It traverses the embedded feature data, updating edge weights and aggregating node vectors at each time step to obtain a spatial feature representation. This involves interspersed dropout layers and residual connections. After traversing all samples, the spatial feature data is obtained. For example... Figure 5 As shown.

[0034] Specifically, the adjacency matrix size of the graph structure after adding seasonal encoding is 13. 13. The adjacency matrix of the graph structure is initialized to 1. That is, the 12 physical parameter nodes and 1 seasonal information node are regarded as a fully connected graph. This ensures that the model can fully explore the potential coupling relationship between the nodes in the initial stage. This initialization strategy avoids the deviation that may be caused by artificially pre-setting physical relationships, and allows the model to adaptively learn from the data the complex nonlinear interactions between key parameters such as temperature and liquid water content, cloud droplets and precipitation particle concentration.

[0035] The process iterates through a sample dataset in positional encoding order. In each iteration, it calculates the attention weights between the sample and all its neighboring nodes. Then, based on these weights, it aggregates the features of the neighboring nodes and combines them with its own features. After a non-linear transformation, a new node representation that aggregates the features of its neighbors is formed. The formula for calculating the attention weights of neighboring nodes is as follows: in, This represents the attention weights for nodes u and v. and Let u and v represent the eigenvalues ​​of node u and node v at time t, respectively. and It is a linear transformation matrix. Learnable attention vector The transpose of the matrix, This indicates a feature concatenation operation. Let be a function. Node u is updated by aggregating the attention weights of its neighboring nodes and the edge weights. The calculation formula for node update is as follows: in, The updated node feature values ​​are defined by M(u), which represents the set of neighboring nodes of node u, σ(·) is the non-linear activation function, and W represents the edge weights. The entire graph structure shares these values. , , And W. Residual connections and dropout layers are interspersed during the update process of the graph structure. Residual connections add the input and output of the module to ensure effective gradient propagation, alleviate the gradient vanishing problem in deep network training, and prevent the loss of key physical processes due to excessive smoothing of information during propagation. For example, a residual connection and dropout layer are added every 20 time points. The output data at each time point is concatenated with the embedded feature data at that time as the input of the temporal feature extraction module, and the output data of the last step is taken as the spatial feature data. The spatial feature data is essentially a high-order abstract representation of the comprehensive interaction relationship between all microphysical parameters within the sample.

[0036] S4 iterates through the embedded feature data. At each time step, the embedded feature data and spatial feature data are input together into the Long Short-Term Memory (LSTM) unit to obtain the temporal feature representation at that time step. It then iterates through the complete sample to output the temporal feature data. For example... Figure 5 As shown.

[0037] The temporal feature extraction process involves sequential processing through multiple interconnected Long Short-Term Memory (LSTM) units. These LSM units selectively remember and forget information through their input, forget, and output gates, making them suitable for processing time-series data with long-term dependencies, such as airborne probe data.

[0038] The embedded feature data is traversed, and the embedded feature data at each time step is concatenated with the spatial feature data corresponding to that time step and input into the long short-term memory unit. This allows the temporal feature extraction process to capture the state of each physical quantity and its interaction relationship at the current time step when analyzing the temporal evolution. This process continues until the entire sample traversal is completed. The output data of the last step is taken as the temporal feature data output, which characterizes the overall dynamic evolution process of the sample.

[0039] S5 temporal and spatial feature data are aggregated through a feature aggregation unit to obtain spatiotemporal feature data. For example... Figure 6 As shown.

[0040] The system receives temporal and spatial feature data, each processed through a linear layer before concatenation. The linear layer maps the temporal and spatial feature data to a unified feature space, facilitating fusion. Next, mask aggregation is performed, where missing locations in the original data are re-marked using a mask. During aggregation, the weights of the mask locations are significantly reduced or zeroed to ensure the final feature representation is unaffected by missing data. Finally, layer normalization outputs the spatiotemporal feature data. Layer normalization stabilizes the mean and variance of the feature distribution, adjusting the aggregated feature vector to zero mean and unit variance. This helps mitigate internal covariate bias, making subsequent classifier training more stable and accelerating model convergence.

[0041] S6 classifies spatiotemporal feature data based on a classifier to obtain the discrimination result.

[0042] The classifier comprises two linear layers and one activation layer. The first linear layer maintains the feature dimensions of the input and output unchanged and is activated by an activation function, outputting to the second linear layer. The output of the second linear layer is either "suitable for the task" or "unsuitable for the task." The first linear layer and the activation function constitute a non-linear transformation layer, used to further refine the fused spatiotemporal features. The second linear layer maps the high-dimensional features to a two-dimensional classification space. The activation function is the ReLU activation function.

[0043] The cloud microphysical feature extraction model is optimized using the cross-entropy loss function. The calculation method of the cross-entropy loss function is as follows: in, Here, N represents the loss function, and N represents the number of classes. This represents the true label of the i-th category. This represents the probability that the sample belongs to the i-th category, i.e., the probability that it is suitable for the task, where i is the category number. The cross-entropy function directly measures the difference between the predicted probability distribution and the true distribution; the greater the deviation, the higher the penalty, which directly optimizes classification accuracy. Cross-entropy can generate clear, non-saturated gradient signals, effectively avoiding the gradient vanishing problem and ensuring that the model can converge efficiently and stably during training, thereby reliably learning discrimination rules from complex spatiotemporal features.

[0044] The intelligent identification model for cold cloud aircraft cloud seeding conditions uses the Adam optimizer for backpropagation and parameter updates, and employs a validation set to monitor model performance and prevent overfitting. Training is stopped early when the validation set loss no longer decreases to avoid overfitting the model on the training set and losing generalization ability, and the best-performing model parameters are saved for subsequent inference.

[0045] The airborne detection data was input into the intelligent identification model for cold cloud aircraft cloud seeding conditions. The model was independently trained five times under the same data division, and the model parameters with the best discrimination effect were selected. This was intended to reduce the performance fluctuation caused by the randomness of model initialization and improve the reliability of the results.

[0046] Example 2 This embodiment provides an intelligent identification system for cold cloud aircraft cloud seeding conditions, such as... Figure 2 As shown.

[0047] The intelligent identification system for cold cloud aircraft cloud seeding conditions includes a data acquisition and preprocessing module, a cloud microphysical feature extraction model, and a classifier. The cloud microphysical feature extraction model includes a data embedding module, a spatial feature extraction module, a temporal feature extraction module, and a feature aggregation module.

[0048] The data acquisition and preprocessing module is used to acquire detection data and preprocess the detection data to obtain preprocessed data.

[0049] The data embedding module is used to embed positional and seasonal codes on the preprocessed data, and obtain embedded feature data through an activation layer and a discard layer.

[0050] The spatial feature extraction module is used to establish an adjacency matrix of a graph structure based on the dimension of the embedded feature data. The parameters are used as nodes of the graph structure. The embedded feature data is traversed, and the edge weights are updated and the node vectors are aggregated at each time step to obtain the spatial feature representation. The module intersperses dropout layers and residual connections. After traversing all samples, the spatial feature data is obtained.

[0051] The time feature extraction module is used to traverse the embedded feature data. The embedded feature data and spatial feature data at each time step are input into the long short-term memory unit to obtain the time feature representation at that time step. The time feature data is output after traversing the complete sample.

[0052] The feature aggregation module is used to aggregate temporal and spatial feature data through the feature aggregation unit to obtain spatiotemporal feature data.

[0053] The classifier is used to classify spatiotemporal feature data based on the classifier to obtain the discrimination result.

[0054] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent identification of cold cloud conditions for aircraft-based cloud seeding, characterized in that, include: S1 acquires detection data and preprocesses the detection data to obtain preprocessed data; the detection data includes second-level airborne detection data acquired from multiple sources. The airborne detection data is stored in tabular form, with each column representing the parameters detected by the aircraft and each row representing the data detected per second. The parameters used include longitude, latitude, altitude, temperature, relative humidity, atmospheric pressure, wind speed, liquid water content, cloud droplet number concentration, precipitation number concentration, cloud condensation nucleus number concentration, and large cloud droplet number concentration, totaling 12 data parameters. S2 embeds positional and seasonal codes into the preprocessed data, and then passes through an activation layer and a dropout layer to obtain embedded feature data. S3 establishes an adjacency matrix of a graph structure based on the dimension of the embedded feature data. The parameters serve as nodes of the graph structure. It traverses the embedded feature data, updates the edge weights and aggregates the node vectors at each time step to obtain the spatial feature representation. Dropout layers and residual connections are interspersed throughout the process. After traversing all samples, the spatial feature data is obtained. S4 iterates through the embedded feature data, and the embedded feature data and spatial feature data at each time step are input into the long short-term memory unit to obtain the temporal feature representation at that time step. It iterates through the complete sample to output the temporal feature data. S5 temporal and spatial feature data are aggregated through a feature aggregation unit to obtain spatiotemporal feature data; S6 classifies the spatiotemporal feature data based on a classifier to obtain a discrimination result, which is divided into two categories: "suitable for the job" and "unsuitable for the job".

2. The intelligent identification method for cold cloud aircraft cloud seeding conditions according to claim 1, characterized in that, The process of preprocessing the probe data to obtain preprocessed data includes dividing the data into fixed time windows, creating a mask for the default values ​​after Z-score normalization, and then copying the results on the channels to increase the dimensionality to obtain the preprocessed data.

3. The intelligent identification method for cold cloud aircraft cloud seeding conditions according to claim 2, characterized in that, The selection of a fixed time window is based on prior knowledge of typical timescales of cloud microphysical processes.

4. The intelligent identification method for cold cloud aircraft cloud seeding conditions according to claim 2, characterized in that, After partitioning, the data is normalized using Z-score, and represented as: in This represents the value after normalization. Represents the original data. and These are the mean and standard deviation of the feature calculated on the training set, respectively. Normalization aims to eliminate the influence of different physical dimensions and numerical ranges, making model training more stable. μ and σ are fixed and saved as normalization parameters and applied to new data during the model inference stage to ensure the consistency of data distribution between the training and inference stages.

5. The intelligent identification method for cold cloud aircraft cloud seeding conditions according to claim 1, characterized in that, The location encoding method is to add location encoding information to the data at each moment by incrementing from 1 according to the time window divided in the preprocessing module.

6. The intelligent identification method for cold cloud aircraft cloud seeding conditions according to claim 1, characterized in that, The seasonal encoding method is as follows: it is initialized to 0000, which represents the four seasons of spring, summer, autumn and winter in sequence, and the value at the corresponding position of the detected season of the data sample is set to 1.

7. The intelligent identification method for cold cloud aircraft cloud seeding conditions according to claim 1, characterized in that, The process of traversing the embedded feature data and updating edge weights and aggregating node vectors at each time step includes traversing a sample data in positional encoding order, and in each iteration, calculating the attention weights between the sample and all its neighboring nodes, expressed as: in, This represents the attention weights for nodes u and v. and Let u and v represent the eigenvalues ​​of node u and node v at time t, respectively. and It is a linear transformation matrix. Learnable attention vector The transpose of the matrix, This indicates a feature concatenation operation. It is a function.

8. The intelligent identification method for cold cloud aircraft cloud seeding conditions according to claim 7, characterized in that, The step of traversing the embedded feature data and updating edge weights and aggregating node vectors at each time step also includes aggregating and updating based on the attention weights and edge weights of the node and its neighboring nodes. The expression for node update is: in, Let M(u) be the updated node feature value, M(u) represent the set of neighboring nodes of node u, σ(·) be the non-linear activation function, and W represent the edge weight.

9. The intelligent identification method for cold cloud aircraft cloud seeding conditions according to claim 1, characterized in that, The classifier consists of two linear layers and one activation layer. The first linear layer keeps the feature dimensions of the input and output unchanged and is activated by an activation function. The output is fed into the second linear layer, which outputs two categories: "suitable for the job" and "unsuitable for the job". The cross-entropy loss function is used to optimize the extraction of cloud microphysical features. The cross-entropy loss function is expressed as: in, Let N be the loss function, and N represent the number of samples. This represents the true label of the i-th sample. This represents the probability that the i-th sample is a positive example, i.e., suitable for the task, where i is the sample sequence.

10. A cold cloud aircraft cloud seeding condition intelligent recognition system, characterized in that, include: The data acquisition and preprocessing module is used to acquire probe data and preprocess the probe data to obtain preprocessed data. The detection data includes second-level airborne detection data acquired from multiple sources; the airborne detection data is stored in tabular form, with each column representing the parameters detected by the aircraft and each row representing the data detected per second; the parameters used include longitude, latitude, altitude, temperature, relative humidity, atmospheric pressure, wind speed, liquid water content, cloud droplet number concentration, precipitation number concentration, cloud condensation nucleus number concentration, and large cloud droplet number concentration, totaling 12 data parameters; The data embedding module is used to embed positional and seasonal codes into preprocessed data, and then pass through an activation layer and a dropout layer to obtain embedded feature data. The spatial feature extraction module is used to build an adjacency matrix of a graph structure based on the dimension of the embedded feature data. The parameters are used as nodes of the graph structure. The module traverses the embedded feature data, updates the edge weights and aggregates the node vectors at each time step to obtain the spatial feature representation. The module intersperses dropout layers and residual connections. After traversing the samples, the spatial feature data is obtained. The temporal feature extraction module is used to traverse the embedded feature data. The embedded feature data and spatial feature data at each time step are input into the long short-term memory unit to obtain the temporal feature representation at that time step. The module traverses the complete sample and outputs the temporal feature data. The feature aggregation module is used to aggregate temporal and spatial feature data through feature aggregation units to obtain spatiotemporal feature data; A classifier is used to classify spatiotemporal feature data based on the classifier to obtain a discrimination result, which is divided into two categories: "suitable for the job" and "unsuitable for the job".

Images