Air hazard wind detection method and device, electronic equipment and readable storage medium

Abstract

The application discloses an aviation hazard wind detection method and device, electronic equipment and a readable storage medium, and belongs to the technical field of aviation safety. The method fuses physical knowledge and actual flight report positions by using a guide map, guides a self-attention module to focus on a physically reasonable area under the condition that labels are extremely scarce through a position alignment loss, and the constructed attention enhancement network has strong multi-scale feature adaptive extraction capability, overcomes the problems that a pure data-driven method fails in attention learning when positive samples are scarce, and a traditional physical standard has low dimensions and cannot capture spatial multi-scale structures, and improves the detection accuracy of aviation hazard wind.

Description

Technical Field

[0001] This invention belongs to the field of aviation safety technology, specifically relating to a method, device, electronic equipment, and readable storage medium for detecting aviation hazard winds. Background Technology

[0002] Aviation hazard winds refer to various wind phenomena that pose a safety threat to aircraft during takeoff, cruise, landing, or parking on the ground. These winds may directly affect flight stability and controllability, and may even lead to accidents, making them an important factor affecting aviation safety.

[0003] Currently, aviation hazardous wind detection methods mainly fall into two categories: The first category is based on physical standards, which identifies hazardous winds by calculating physical indicators. These methods offer strong physical interpretability and do not require a large number of labels, but they can only capture low-dimensional features and cannot cover the multi-scale characteristics of hazardous winds, resulting in limited detection accuracy. The second category is purely data-driven methods, which learn hazardous wind features from observational data based on machine learning or deep learning models. These methods can extract high-dimensional features, but in scenarios with extremely scarce labels, the model struggles to learn effective features, leading to problems such as overfitting, poor interpretability, and weak generalization ability. Both current mainstream methods struggle to balance high accuracy, strong generalization, and good interpretability, resulting in poor detection performance for aviation hazardous winds. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that the above-mentioned methods commonly used in the prior art are difficult to achieve high precision, strong generalization and good interpretability at the same time, resulting in poor detection effect of aviation hazard wind. In order to solve the above problems, the present invention provides an aviation hazard wind detection method, device, electronic device and readable storage medium.

[0005] The content of this invention includes: In a first aspect, embodiments of the present invention provide a method for detecting aviation hazard winds, including: Acquire radar observation training data and hazardous wind field locations within historical detection periods. The radar observation training data includes Doppler velocity sampling data from multiple sampling points covering the directions before and after the aircraft take-off and landing passage. The aviation hazardous wind detection model is iteratively trained using the radar observation training data and the location of the hazardous wind field to obtain a trained aviation hazardous wind detection model. The aviation hazardous wind detection model includes an encoder, a decoder, and a classifier. The encoder includes a first convolutional module and at least one self-attention module. The decoder includes a second convolutional module. The first convolutional module and the second convolutional module have a symmetrical structure. The encoder is used to extract multi-scale flow field features. The decoder is used to perform feature reconstruction based on the multi-scale flow field features to obtain the reconstruction result. The classifier is used to detect based on the multi-scale flow field features to obtain the hazard detection result. The trained aviation hazard wind detection model is used to detect the radar observation data to be detected, and the hazard wind detection results are obtained. The iterative training of the aviation hazard wind detection model using the radar observation training data and the location of the hazard wind field includes: A sample observation matrix is ​​constructed based on the radar observation training data; The sample observation matrix is ​​input into the aviation hazard wind detection model for processing to obtain the training output results; The physical parameters of the wind field at each sampling point on the aircraft take-off and landing channel are calculated based on the radar observation training data, and physical knowledge priors are constructed based on the physical parameters. The prior to the operation location is determined based on the location of the hazardous wind field. The prior to the operation location is used to describe whether each sampling point on the aircraft take-off and landing channel is a location of a hazardous wind field. A guidance map is generated based on the aforementioned physical knowledge priors and the aforementioned operation position priors; The parameters of the aviation hazard wind detection model are adjusted based on the loss value, which includes a position alignment loss. The position alignment loss is determined based on the difference between the attention weight matrix output by the attention module and the guidance map.

[0006] Optionally, the step of calculating the physical indices of the wind field at each sampling point on the aircraft takeoff and landing path based on the radar observation training data, and constructing prior physical knowledge based on the physical indices, includes: Calculate the F-factor and eddy current dissipation rate at each sampling point in the aircraft takeoff and landing tunnel; The sampling points that satisfy the first preset condition of the F factor are marked as 1, and the remaining sampling points are marked as 0, to obtain the first physical prior; The sampling points where the eddy current dissipation rate satisfies the second preset condition are marked as 1, and the remaining sampling points are marked as 0, thus obtaining the second physical prior. The physical knowledge prior is obtained by taking the union of the first physical prior and the second physical prior.

[0007] Optionally, determining the prior operation location based on the location of the hazardous wind field includes: The sampling point corresponding to the location of the hazardous wind field on the aircraft take-off and landing passage is marked as 1, and other sampling points are marked as 0, thus obtaining the prior of the operation position.

[0008] Optionally, before adjusting the parameters of the aviation hazard wind detection model based on the loss value, the method further includes: The importance vector is obtained by calculating the average value of the attention weight matrix row by row. The guidance map is downsampled using bilinear interpolation to have the same spatial resolution as the importance vector to obtain the target guidance map; The position alignment loss is calculated based on the difference between the target guidance map and the importance vector.

[0009] Optionally, the position alignment loss for: ; in, Let be the importance vector. The guide diagram is as described. The target guidance map is used to characterize the result of the downsampling operation. The square of the L2 norm is used to characterize the L2 norm.

[0010] Optionally, the loss value further includes reconstruction loss and detection loss, wherein the reconstruction loss is determined based on the difference between the reconstruction result output by the decoder and the sample observation matrix, and the detection loss is determined based on the difference between the training output and the label value.

[0011] Optionally, the step of using the trained aviation hazard wind detection model to detect the radar observation data to obtain hazard wind detection results includes: Acquire the radar observation data to be detected; Select the region of interest in the aircraft takeoff and landing corridor. There are 1 parallel line of sight, each corresponding to a virtual observation line parallel to the aircraft takeoff and landing passage. ; Setting each of the aforementioned parallel lines of sight along the direction of the aircraft takeoff and landing passage equidistant sampling points, ; The radial velocity value of each sampling point is obtained by interpolation from the radar observation data to be detected using the k-nearest neighbor method, and a two-dimensional observation matrix is ​​constructed. The two-dimensional observation matrix is ​​input into the trained aviation hazard wind detection model for detection, and the hazard wind detection result is obtained.

[0012] Secondly, embodiments of the present invention provide an aviation hazard wind detection device, comprising: The acquisition module is used to acquire radar observation training data and the location of hazardous wind fields within the historical detection period. The radar observation training data includes Doppler velocity sampling data from multiple sampling points covering the direction in front of and behind the aircraft take-off and landing passage. The training module is used to iteratively train the aviation hazardous wind detection model using the radar observation training data and the location of the hazardous wind field to obtain a trained aviation hazardous wind detection model. The aviation hazardous wind detection model includes an encoder, a decoder, and a classifier. The encoder includes a first convolutional module and at least one self-attention module. The decoder includes a second convolutional module. The first convolutional module and the second convolutional module have a symmetrical structure. The encoder is used to extract multi-scale flow field features. The decoder is used to perform feature reconstruction based on the multi-scale flow field features to obtain the reconstruction result. The classifier is used to detect based on the multi-scale flow field features to obtain the hazard detection result. The detection module is used to detect the radar observation data to be detected using the trained aviation hazardous wind detection model to obtain the hazardous wind detection result; The iterative training of the aviation hazard wind detection model using the radar observation training data and the location of the hazard wind field includes: A sample observation matrix is ​​constructed based on the radar observation training data; The sample observation matrix is ​​input into the aviation hazard wind detection model for processing to obtain the training output results; The physical parameters of the wind field at each sampling point on the aircraft take-off and landing channel are calculated based on the radar observation training data, and physical knowledge priors are constructed based on the physical parameters. The prior to the operation location is determined based on the location of the hazardous wind field. The prior to the operation location is used to describe whether each sampling point on the aircraft take-off and landing channel is a location of a hazardous wind field. A guidance map is generated based on the aforementioned physical knowledge priors and the aforementioned operation position priors; The parameters of the aviation hazard wind detection model are adjusted based on the loss value, which includes a position alignment loss. The position alignment loss is determined based on the difference between the attention weight matrix output by the attention module and the guidance map.

[0013] Thirdly, embodiments of the present invention provide an electronic device, including: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is configured to read the program in the memory to implement the steps in the aviation hazard wind detection method as described in the first aspect.

[0014] Fourthly, embodiments of the present invention provide a readable storage medium for storing a program, which, when executed by a processor, implements the steps in the aviation hazard wind detection method as described in the first aspect.

[0015] The beneficial effects of this invention are as follows: In the embodiments of this invention, when training the aviation hazardous wind detection model, on the one hand, physical knowledge priors are constructed based on the calculation of wind field physical indicators using radar Doppler velocity data, thereby reflecting the laws of meteorological dynamics; on the other hand, based on the location of the hazardous wind field, high-risk areas in the takeoff and landing channels are identified, generating operational position priors. A guiding map is generated through these two methods, serving as a soft supervision signal for attention mechanism learning. This forces the model to allocate high attention weights to sampling points that are physically high-risk and historically have frequently occurred hazardous winds, thereby enhancing the physical interpretability and operational reliability of the detection results. Using the aviation hazardous wind detection model trained in the above manner for aviation hazardous wind detection allows for better extraction and processing of information from radar observation data, resulting in more accurate detection results and providing high-precision, reliable technical support for intelligent aviation meteorological early warning systems. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the aviation hazard wind detection method provided in an embodiment of the present invention; Figure 2 A schematic diagram of the scanning geometry setup for scanning lidar to detect aircraft takeoff and landing passageways. Figure 3 This is a schematic diagram of the structure of the aviation hazard wind detection model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the processing flow of the multi-head attention module provided in an embodiment of the present invention; Figure 5a Example diagram of single-viewpoint radial velocity data along the aircraft takeoff and landing path; Figure 5b A schematic diagram showing the location distribution of the nine parallel lines of sight; Figure 5c A schematic diagram of the constructed 128×9 two-dimensional observation matrix; Figure 6a This is a priori diagram of the operating position based on the pilot's report; Figure 6b This is a schematic diagram of physical prior generation based on the F-factor; Figure 6c A schematic diagram of physical prior generation based on eddy current dissipation rate; Figure 6d This is a schematic diagram of the fused guide diagram; Figure 7 This is a schematic diagram of the structure of the aviation hazard wind detection device provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] In the embodiments of this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar. The terms "first," "second," etc., in the specification of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such terms can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are usually of the same class, without limiting the number of objects. For example, the first object can be one or multiple.

[0018] 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 a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0020] This application provides a method, apparatus, electronic device, and readable storage medium for detecting aviation hazard winds, aiming to solve the technical problems of low detection accuracy and insufficient multi-scale feature capture under extreme tag scarcity conditions.

[0021] Please see Figure 1 , Figure 1 This is a flowchart illustrating the aviation hazard wind detection method provided in an embodiment of the present invention. The method specifically includes the following steps: Step 101: Obtain radar observation training data and the location of hazardous wind fields within the historical detection period. The radar observation training data includes Doppler velocity sampling data from multiple sampling points covering the direction in front of and behind the aircraft take-off and landing passage.

[0022] Step 102: Iteratively train the aviation hazardous wind detection model using the radar observation training data and the location of the hazardous wind field to obtain a trained aviation hazardous wind detection model. The aviation hazardous wind detection model includes an encoder, a decoder, and a classifier. The encoder includes a first convolutional module and at least one self-attention module. The decoder includes a second convolutional module. The first convolutional module and the second convolutional module have a symmetrical structure. The encoder is used to extract multi-scale flow field features. The decoder is used to perform feature reconstruction based on the multi-scale flow field features to obtain the reconstruction result. The classifier is used to perform detection based on the multi-scale flow field features to obtain the hazard detection result.

[0023] Step 103: Use the trained aviation hazard wind detection model to detect the radar observation data to obtain the hazard wind detection result.

[0024] The radar observation training data and hazardous wind field locations obtained within the historical detection period are used as training datasets to train the pre-built aviation hazardous wind detection model. The specific number of radar observation training data and hazardous wind field locations obtained within the historical detection period is not limited here.

[0025] In some embodiments, during a historical detection period, a scanning lidar is used to perform a two-dimensional continuous scan of the aircraft takeoff and landing passage area to acquire radial velocity sampling data at multiple lines of sight (LOS) covering the directions before and after the aircraft takeoff and landing passage, thus obtaining radar observation training data. Radial velocity refers to the velocity component of the target relative to the lidar along the LOS direction.

[0026] In some embodiments, the location of hazardous winds is recorded based on pilot reports within a historical detection period and is used as the hazardous wind field location. In other embodiments, the hazardous wind field location is obtained based on logs from an automatic early warning system within a historical detection period.

[0027] After obtaining radar observation training data and the location of hazardous wind fields, the data corresponding to the occurrence time of hazardous winds in the radar observation training data are used as labeled hazardous wind field data, which are used as positive samples during training. From the radar observation training data that do not correspond to the occurrence time of hazardous winds, the same number of data points as the hazardous wind field data or data points selected according to a preset ratio are randomly selected as labeled non-hazardous wind field data, which are used as negative samples during training.

[0028] In specific implementation, the pre-constructed aviation hazard wind detection model is iteratively trained based on the aforementioned positive and negative samples until a preset convergence condition is met, resulting in a well-trained aviation hazard wind detection model. Specifically, the iterative training of the aviation hazard wind detection model using the radar observation training data and the location of the hazard wind field includes: A sample observation matrix is ​​constructed based on the radar observation training data; The sample observation matrix is ​​input into the aviation hazard wind detection model for processing to obtain the training output results; The physical parameters of the wind field at each sampling point on the aircraft take-off and landing channel are calculated based on the radar observation training data, and physical knowledge priors are constructed based on the physical parameters. The prior to the operation location is determined based on the location of the hazardous wind field. The prior to the operation location is used to describe whether each sampling point on the aircraft take-off and landing channel is a location of a hazardous wind field. A guidance map is generated based on the aforementioned physical knowledge priors and the aforementioned operation position priors; The parameters of the aviation hazard wind detection model are adjusted based on the loss value, which includes a position alignment loss. The position alignment loss is determined based on the difference between the attention weight matrix output by the attention module and the guidance map.

[0029] Preprocessing the radar observation training data yields the observation matrix input to the aviation hazard wind detection model. For ease of description, the observation matrix corresponding to the radar observation training data is referred to as the sample observation matrix. Taking radar observation training data at any sampling time as an example, in some embodiments, multiple parallel lines of sight are selected along the aircraft take-off and landing channel direction, and multiple equidistant sampling points are set on each parallel line of sight. Based on the radar observation training data, interpolation is performed to obtain the data of each sampling point, thereby constructing the corresponding sample observation matrix.

[0030] As a specific example, a region of interest (ROI) is selected within the aircraft takeoff and landing corridor. There are 1 parallel line of sight, and each parallel line of sight corresponds to a virtual observation line parallel to the aircraft takeoff and landing passage. Set each parallel line of sight along the direction of the aircraft takeoff and landing passage. equidistant sampling points, The radial velocity value of each sampling point is obtained by interpolating from the radar observation training data at the current sampling time using the k-nearest neighbor method, and a sample observation matrix is ​​constructed. The size of the sample observation matrix is... .

[0031] The aviation hazardous wind detection model includes an encoder, a decoder, and a classifier. The encoder contains a convolutional module and a self-attention module, responsible for extracting multi-scale flow field features from the input data. The extracted multi-scale flow field features are processed in parallel. One path reconstructs the input flow field through an upsampled decoder symmetrical to the convolutional part of the encoder, obtaining a reconstruction result used to assist feature learning. The other path is fed into the classifier, which outputs the hazardous wind detection result, with a hazardous wind output as 1 and a non-hazardous wind output as 0. In some embodiments, the encoder consists of a first convolutional module and at least one multi-head self-attention module cascaded together, wherein the first convolutional module includes at least one convolutional layer.

[0032] For the input observation matrix , observation matrix The first convolutional module extracts local features, and these local features are then input into the self-attention module for processing. Let the input features of the self-attention module be... Then query matrix Key matrix Sum matrix They are respectively: , , ; in, , , These are the query weight matrix, key weight matrix, and value weight matrix, all of which are learnable weight matrices, and the attention weight matrix output by the attention module. for:

[0033] in, This is the attention weight matrix, used to dynamically weight features at different spatial locations. for Dimensions This is used to characterize the softmax function. The final output of the self-attention module is as follows: ; It should be understood that when there are multiple self-attention modules, the input of the first self-attention module is the local features output by the first convolutional module, the input of other self-attention modules is the output of the previous self-attention module, and the output of the last self-attention module is the output of the encoder.

[0034] The encoder's output is input into the decoder for processing. The decoder includes a second convolution module. The first and second convolution modules are symmetrical convolutional structures. The second convolution module is used to perform reconstruction processing based on the encoder's output to obtain the reconstruction result.

[0035] It should be understood that the first and second convolutional modules having a symmetrical structure means that both modules contain the same number of convolutional layers, and it is assumed that the number of convolutional layers is... The first convolutional module's first... The structure of the first convolutional layer is the same as that of the second convolutional module. Each convolutional layer has the same structure. , .

[0036] In this embodiment, physical knowledge priors and operational position priors are constructed and a guidance map is generated. The guidance network allocates higher attention weights to the areas indicated by the guidance map. The physical knowledge priors are determined by calculating the physical indicators of the wind field, and the operational position priors are generated based on the actual hazardous wind locations reported by pilots or recorded by automatic warning systems.

[0037] Since physical laws are universal and stable, introducing physical priors can enable the model to maintain reliable performance when facing different airports and different weather conditions, reducing the over-reliance on historical label data in specific scenarios. Physical prior knowledge ensures that the model follows basic physical laws and has broad applicability.

[0038] Specifically, the physical parameters of the wind field are not limited here. For example, the physical parameters of the wind field include at least one of the following: wind shear index, mean wind speed, turbulence intensity, F-factor, and eddy dissipation rate (EDR). In some embodiments, the physical parameters of the wind field include the F-factor and the eddy dissipation rate.

[0039] Optionally, in some embodiments, the step of calculating the physical indices of the wind field at each sampling point on the aircraft takeoff and landing path based on the radar observation training data, and constructing prior physical knowledge based on the physical indices, includes: Calculate the F-factor and eddy current dissipation rate at each sampling point in the aircraft takeoff and landing tunnel; The sampling points that satisfy the first preset condition of the F factor are marked as 1, and the remaining sampling points are marked as 0, to obtain the first physical prior; The sampling points where the eddy current dissipation rate satisfies the second preset condition are marked as 1, and the remaining sampling points are marked as 0, thus obtaining the second physical prior. The physical knowledge prior is obtained by taking the union of the first physical prior and the second physical prior.

[0040] It should be understood that the first and second preset conditions are not limited herein. In some embodiments, the first preset condition is that the F factor is greater than a first preset value, and the second preset condition is that the eddy current dissipation rate is greater than a second preset value. In other embodiments, the top X sampling points with the highest F factor are marked as 1, and the remaining sampling points are marked as 0; the top Y sampling points with the highest eddy current dissipation rate are marked as 1, and the remaining sampling points are marked as 0.

[0041] Specifically, the F factor The calculation is as follows: ; in, For headwind speed, Vertical wind speed, For the aircraft's approach speed, This is the acceleration due to gravity.

[0042] Eddy current dissipation rate The calculation is as follows: ; in, Kolmogorov's constant, , The wavenumber corresponding to the sampling scale, This represents the standard deviation of the velocity.

[0043] The hazards defined by physical indicators may differ slightly from the hazards actually perceived by pilots. Operational priors can calibrate the model's judgment criteria, making its outputs more closely match real-world aviation safety concerns, thus improving the model's usability and reliability.

[0044] Optionally, in some embodiments, determining the prior operation location based on the location of the hazardous wind field includes: The sampling point corresponding to the location of the hazardous wind field on the aircraft take-off and landing passage is marked as 1, and other sampling points are marked as 0, thus obtaining the prior of the operation position.

[0045] In this embodiment, prior physical knowledge and prior operational position knowledge are fused to obtain the guidance map. Specifically, the prior operational position knowledge is fused... and prior knowledge of physics Weighted fusion is performed to obtain the guide graph. : ; in, As a balancing coefficient, it is used to adjust the confidence levels of prior physical knowledge and prior operational location. In some embodiments, Set it to 0.5.

[0046] In some embodiments, the observation matrix The size is ,in, This represents the number of sampling points on the aircraft takeoff and landing ramp. This represents the number of parallel lines of sight parallel to the aircraft takeoff and landing ramp within the region of interest; each element... Representative at the The sampling point and the first Radar radial velocity observations along a parallel line of sight.

[0047] As an optional implementation method, and as a specific embodiment, based on the traditional wind shear index F-factor and eddy current dissipation rate, the index values ​​at each sampling point along the aircraft takeoff and landing tunnel are calculated, and the points with the highest index values ​​are selected. Each sampling point is marked as 1, and the rest as 0, resulting in two sets of binary vectors. Taking the union of these vectors yields the prior physical knowledge. Specifically, based on the location of the hazardous wind field reported by the pilot or provided by the automatic warning system, the sampling points on the aircraft takeoff and landing tunnel corresponding to the reported location are marked as 1, and the rest are marked as 0, thus obtaining the prior operational position. By fusing prior physical knowledge and prior operational position knowledge, a guidance graph is obtained. The dimension of the guidance graph is... .

[0048] Optionally, in some embodiments, before adjusting the parameters of the aviation hazard wind detection model based on the loss value, the method further includes: The importance vector is obtained by calculating the average value of the attention weight matrix row by row. The guidance map is downsampled using bilinear interpolation to have the same spatial resolution as the importance vector to obtain the target guidance map; The position alignment loss is calculated based on the difference between the target guidance map and the importance vector. In this embodiment, averaging the attention weight matrix row by row allows for the extraction of the one-dimensional signal most directly related to flight safety, making the model's learning objective clearer. Simultaneously, averaging smooths out attention fluctuations caused by noise or local anomalies, resulting in a more stable learning signal.

[0049] Due to the configuration of the first convolutional module, the importance vector is a low-resolution representation obtained after several layers of convolution / pooling, and its dimension is related to the structure of the first convolutional module. In order to supervise the importance vector using the guiding graph, the guiding graph needs to be downsampled to have the same spatial resolution as the importance vector. Bilinear interpolation is a smooth and differentiable downsampling method that can better preserve the continuous structure and gradient information (such as the gradual region of wind shear intensity) in the prior graph.

[0050] Of course, as another optional implementation method, after calculating the dimension through the above implementation method, After the initial guide graph, it is expanded to obtain a dimension of The specific method for creating the guiding diagram is not limited. For example, based on the distance between the parallel line of sight and the aircraft takeoff and landing tunnel, and using the index values ​​of each sampling point on the aircraft takeoff and landing tunnel, the physical indexes of each sampling point on the parallel line of sight are proportionally calculated, resulting in a dimension of... Prior knowledge of physics.

[0051] Optionally, the position alignment loss for: ; in, Let be the importance vector. The guide diagram is as described. Used to characterize the target guidance map obtained after the downsampling operation. The square of the L2 norm is used to characterize the L2 norm.

[0052] Optionally, in some embodiments, the loss value further includes reconstruction loss and detection loss, wherein the reconstruction loss is determined based on the difference between the reconstruction result output by the decoder and the sample observation matrix, and the detection loss is determined based on the difference between the training output and the label value.

[0053] It should be understood that the reconstruction loss (such as mean squared error or L1 loss) constrains the decoder to reproduce the input sample observation matrix as accurately as possible, forcing the encoder to learn a physically consistent, structurally complete, and information-rich low-dimensional representation of the wind field in the latent space. This self-supervised reconstruction objective not only enhances the model's robustness to noise and missing values ​​in radar observation data but also effectively suppresses overfitting, which is particularly valuable in real-world scenarios with limited training samples.

[0054] It should be understood that detection losses (such as cross-entropy loss or triplet clustering loss) directly drive the model output to approximate the real hazard wind labels, ensuring that the learned feature representations are highly focused on discriminative information, such as key hazard signals like wind shear gradients, turbulence spatial patterns, or the divergent structure of downbursts.

[0055] In this embodiment, by simultaneously introducing reconstruction loss and detection loss into the loss function, the synergistic optimization of representation learning and task-oriented learning is achieved, significantly improving the overall performance and generalization ability of the aviation hazard wind detection model.

[0056] In some embodiments, the final loss value includes position alignment loss, reconstruction loss, and detection loss. As a specific example, the reconstruction loss is the mean squared error. : ; in, The reconstruction result is output by the decoder. The square of the L2 norm is used to characterize the L2 norm.

[0057] Detection loss A triplet clustering loss is employed, where the detection loss causes the features of hazardous samples to cluster towards a learnable cluster center in the embedding space, while simultaneously distancing the features of non-hazardous samples from this cluster center. Through the detection loss, the network learns discriminative feature representations, effectively distinguishing between hazardous and non-hazardous winds.

[0058] In this embodiment, the total loss function for: ; in, The pre-set weighting coefficients.

[0059] After training the aviation hazard wind detection model as described above, the trained model can be used to detect aviation hazard winds. Optionally, in some embodiments, step 103 includes: Acquire the radar observation data to be detected; Select the region of interest in the aircraft takeoff and landing corridor. There are 1 parallel line of sight, each corresponding to a virtual observation line parallel to the aircraft takeoff and landing passage. ; Setting each of the aforementioned parallel lines of sight along the direction of the aircraft takeoff and landing passage equidistant sampling points, ; The radial velocity value of each sampling point is obtained by interpolation from the radar observation data to be detected using the k-nearest neighbor method, and a two-dimensional observation matrix is ​​constructed. The two-dimensional observation matrix is ​​input into the trained aviation hazard wind detection model for detection, and the hazard wind detection result is obtained.

[0060] In this embodiment, by constructing a two-dimensional multi-line-of-sight observation structure, not only are the longitudinal evolution characteristics of the wind field preserved along the aircraft take-off and landing corridor, but the spatial non-uniformity and structural changes of the wind field in the lateral direction are also effectively captured by setting multiple parallel lines of sight. Compared with traditional single-line-of-sight or one-dimensional profile methods, this method significantly enhances the ability to perceive aviation hazard winds with obvious spatial gradient characteristics, such as micro-downbursts, local wind shear, and lateral turbulence zones.

[0061] The following is an example of a specific implementation. A long-range scanning coherent Doppler lidar is installed beside the airport runway to continuously scan the wind field area above the runway. For example... Figure 2 As shown, Figure 2 The central sector represents the scanning coverage area of ​​the lidar. Doppler velocity data obtained from the radar scan of the area above the runway is interpolated to obtain Doppler velocity sampling data along the aircraft takeoff and landing paths. Each scan generates multiple line-of-sight wind records along the aircraft takeoff and landing paths. Simultaneously, hazardous wind field information reported by pilots or output by the onboard automatic warning system is recorded, including the location and time of occurrence of hazardous wind fields, as training label values.

[0062] The radar observation training data obtained through the above method, corresponding to the pilot's reported time, is used as labeled hazardous wind field data (positive samples). Specifically, the training set contains 159 positive samples and 524,280 unlabeled data points. From the unlabeled data that does not correspond to hazardous times, the same number of data points (159) as the positive samples are randomly selected as labeled non-hazardous wind field data (negative samples). The test set contains 159 positive samples and 159 randomly selected negative samples.

[0063] A network containing an encoder, decoder, and classifier is pre-constructed as an aviation hazard wind detection model, the specific structure of which is as follows: Figure 3 As shown. The encoder consists of three convolutional layers and a multi-head self-attention module cascaded together, with convolutional kernel sizes of 5×5, 9×9, and 15×15, respectively. The input is a two-dimensional matrix. Local features are extracted by processing through three convolutional layers in sequence. Then, the input is processed by a multi-head self-attention module. The decoder employs a convolutional structure symmetrical to the encoder to reconstruct the input observations and obtain the reconstruction result. The structure of a multi-head self-attention module is as follows: Figure 4 As shown, The input is processed by a multi-head self-attention module, where... Input features Dimensions The value vector dimension for each attention head.

[0064] The aviation hazard wind detection model is iteratively trained using the data in the aforementioned training set until a well-trained model is obtained. In this embodiment, 6000 meters outward from the landing point along the aircraft takeoff and landing corridor, L=128 equidistant sampling points are set to form the original one-dimensional radial velocity vector, as shown below. Figure 5a As shown. To capture the lateral spatial structure of the wind field, N=9 parallel lines of sight are selected within the region of interest around the aircraft takeoff and landing tunnel, as shown. Figure 5bAs shown. Each parallel line of sight corresponds to a virtual observation line parallel to the aircraft takeoff and landing tunnel, and the wind speed value on it is obtained from the original lidar measurements using the k-nearest neighbor interpolation method. This forms the basis of the sample observation matrix. ,like Figure 5c As shown, the rows correspond to the distance index of the aircraft take-off and landing passage (128 sampling points), and the columns correspond to different viewpoints (9). This matrix retains the lateral spatial variation information of the wind field as the input of the subsequent network.

[0065] Please see Figures 6a-6d It should be noted that, Figures 6a-6d The text provides two data examples from different time periods. Figures 6a-6d The diagram on the left shows data from March 4, 2017, and the diagram on the right shows data from March 1, 2017. During training, based on hazard locations reported by pilots or provided by the automatic warning system, sampling points on the aircraft takeoff and landing ramps corresponding to the reported locations were marked as 1, and the rest as 0, resulting in a binary vector of length 128. See [link to documentation] for details. Figure 6a , Figure 6a This is a schematic diagram of the prior operating position based on the pilot's report. (a1) is a schematic diagram of the prior operating position reported by the pilot based on data collected on March 4, 2017, and (a2) is a schematic diagram of the prior operating position reported by the pilot based on data collected on March 1, 2017. The F-factor value and eddy current dissipation rate (EDR) of each sampling point are calculated along the aircraft take-off and landing corridor. The positions with the highest F-factor value (top P%) and the positions with the highest EDR value (top P%) are marked as 1, and the rest are marked as 0, resulting in two binary vectors. The union of these vectors yields the prior physical knowledge. In this embodiment, P is set to 90. For details, please refer to [link to documentation]. Figure 6b and Figure 6c , Figure 6b The diagrams illustrate the generation of physical priors for the F-factor, where (b1) is a diagram of the generation of physical priors for the F-factor based on data collected on March 4, 2017, and (b2) is a diagram of the generation of physical priors for the F-factor based on data collected on March 1, 2017. Figure 6c The diagram shows the physical prior generation based on eddy current dissipation rate, where (c1) is the physical prior generation diagram based on the data collected on 2017 / 03 / 04, and (c2) is the physical prior generation diagram based on the data collected on 2017 / 03 / 01.

[0066] The guidance graph G is obtained by further integrating prior physical knowledge with prior operational position knowledge. (See details in [link to documentation]). Figure 6d , Figure 6dThe diagrams show the fused guide graphs, where (d1) is a schematic diagram of the fused guide graph based on data collected on March 4, 2017, and (d2) is a schematic diagram of the fused guide graph based on data collected on March 1, 2017. The resolution of the guide graph G is related to the input observation matrix. Since the dimensions are the same (128×1), G is downsampled to the same spatial resolution as the feature map using bilinear interpolation to obtain the target guidance map. The attention weight matrix output from the self-attention module... In the middle, the average value is calculated row by row to obtain the importance vector. Flatten the target guidance map to match... Vectors of the same dimension.

[0067] The total loss is calculated based on the position alignment loss, reconstruction loss, and detection loss. This total loss is then used to iteratively train the hazardous wind detection model, ultimately yielding a well-trained aviation hazardous wind detection model. For the radar observation data to be detected, a corresponding two-dimensional observation matrix is ​​constructed following the steps described above. Input the trained aviation hazard wind detection model and obtain the classifier output. If the classifier output is 1, the observation is determined to contain aviation hazard wind; otherwise, it is determined to be hazardless.

[0068] In this embodiment, the two-dimensional spatial structure of the wind field is obtained through a two-dimensional multi-line observation matrix. Multi-scale feature adaptive extraction is achieved through a self-attention mechanism. A guided graph integrates physical knowledge and operational priors into the attention learning process, significantly improving detection performance even under extreme label scarcity conditions. Furthermore, the guided attention graph of this invention is directly interpretable, enhancing the model's credibility in actual aviation safety operations.

[0069] like Figure 7 As shown, this embodiment of the invention also provides an aviation hazard wind detection device 700, comprising: The acquisition module 701 is used to acquire radar observation training data and the location of hazardous wind fields within the historical detection period. The radar observation training data includes Doppler velocity sampling data of multiple sampling points covering the direction in front of and behind the aircraft take-off and landing passage. Training module 702 is used to iteratively train the aviation hazardous wind detection model using the radar observation training data and the location of the hazardous wind field to obtain a trained aviation hazardous wind detection model. The aviation hazardous wind detection model includes an encoder, a decoder, and a classifier. The encoder includes a first convolutional module and at least one self-attention module. The decoder includes a second convolutional module. The first convolutional module and the second convolutional module have a symmetrical structure. The encoder is used to extract multi-scale flow field features. The decoder is used to perform feature reconstruction based on the multi-scale flow field features to obtain a reconstruction result. The classifier is used to detect based on the multi-scale flow field features to obtain a hazard detection result. The detection module 703 is used to detect the radar observation data to be detected using the trained aviation hazard wind detection model to obtain the hazard wind detection result; The iterative training of the aviation hazard wind detection model using the radar observation training data and the location of the hazard wind field includes: A sample observation matrix is ​​constructed based on the radar observation training data; The sample observation matrix is ​​input into the aviation hazard wind detection model for processing to obtain the training output results; The physical parameters of the wind field at each sampling point on the aircraft take-off and landing channel are calculated based on the radar observation training data, and physical knowledge priors are constructed based on the physical parameters. The prior to the operation location is determined based on the location of the hazardous wind field. The prior to the operation location is used to describe whether each sampling point on the aircraft take-off and landing channel is a location of a hazardous wind field. A guidance map is generated based on the aforementioned physical knowledge priors and the aforementioned operation position priors; The parameters of the aviation hazard wind detection model are adjusted based on the loss value, which includes a position alignment loss. The position alignment loss is determined based on the difference between the attention weight matrix output by the attention module and the guidance map.

[0070] Optionally, the step of calculating the physical indices of the wind field at each sampling point on the aircraft takeoff and landing path based on the radar observation training data, and constructing prior physical knowledge based on the physical indices, includes: Calculate the F-factor and eddy current dissipation rate at each sampling point in the aircraft takeoff and landing tunnel; The sampling points that satisfy the first preset condition of the F factor are marked as 1, and the remaining sampling points are marked as 0, to obtain the first physical prior; The sampling points where the eddy current dissipation rate satisfies the second preset condition are marked as 1, and the remaining sampling points are marked as 0, thus obtaining the second physical prior. The physical knowledge prior is obtained by taking the union of the first physical prior and the second physical prior.

[0071] Optionally, determining the prior operation location based on the location of the hazardous wind field includes: The sampling point corresponding to the location of the hazardous wind field on the aircraft take-off and landing passage is marked as 1, and other sampling points are marked as 0, thus obtaining the prior of the operation position.

[0072] Optionally, before adjusting the parameters of the aviation hazard wind detection model based on the loss value, the method further includes: The importance vector is obtained by calculating the average value of the attention weight matrix row by row. The guidance map is downsampled using bilinear interpolation to have the same spatial resolution as the importance vector to obtain the target guidance map; The position alignment loss is calculated based on the difference between the target guidance map and the importance vector.

[0073] Optionally, the position alignment loss for: ; in, Let be the importance vector. The guide diagram is as described. The target guidance map is used to characterize the result of the downsampling operation. The square of the L2 norm is used to characterize the L2 norm.

[0074] Optionally, the loss value further includes reconstruction loss and detection loss, wherein the reconstruction loss is determined based on the difference between the reconstruction result output by the decoder and the sample observation matrix, and the detection loss is determined based on the difference between the training output and the label value.

[0075] Optionally, the detection module 703 is specifically used for: Acquire the radar observation data to be detected; Select the region of interest in the aircraft takeoff and landing corridor. There are 1 parallel line of sight, each corresponding to a virtual observation line parallel to the aircraft takeoff and landing passage. ; Setting each of the aforementioned parallel lines of sight along the direction of the aircraft takeoff and landing passage equidistant sampling points, ; The radial velocity value of each sampling point is obtained by interpolation from the radar observation data to be detected using the k-nearest neighbor method, and a two-dimensional observation matrix is ​​constructed. The two-dimensional observation matrix is ​​input into the trained aviation hazard wind detection model for detection, and the hazard wind detection result is obtained.

[0076] The aviation hazard wind detection device 700 provided in this application embodiment can perform the above method embodiment, and its implementation principle and technical effect are similar, so it will not be described again here.

[0077] It should be noted that the division of units in the embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0078] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0079] like Figure 8 As shown, this application embodiment provides an electronic device 800, including: a memory 802, a processor 801, and a program stored in the memory 802 and executable on the processor 801; the processor 801 is used to read the program in the memory 802 to implement the steps in the aviation hazard wind detection method described above.

[0080] This application also provides a readable storage medium storing a program. When the program is executed by a processor, it implements the various processes of the above-described aviation hazard wind detection method embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here. The readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (such as compact disks (CDs), digital video discs (DVDs), Blu-ray discs (BDs), high-definition universal discs (HVDs), etc.), and semiconductor storage (such as read-only memory (ROMs), erasable programmable read-only memory (EPROMs), electrically erasable programmable read-only memory (EEPROMs), non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0081] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0082] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0083] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of this application without departing from its spirit, and all of these modifications are within the scope of protection of this application.

Claims

1. A method for detecting aviation hazard winds, characterized in that, include: Acquire radar observation training data and hazardous wind field locations within historical detection periods. The radar observation training data includes Doppler velocity sampling data from multiple sampling points covering the direction in front of and behind the aircraft take-off and landing passage. The aviation hazardous wind detection model is iteratively trained using the radar observation training data and the location of the hazardous wind field to obtain a trained aviation hazardous wind detection model. The aviation hazardous wind detection model includes an encoder, a decoder, and a classifier. The encoder includes a first convolutional module and at least one self-attention module. The decoder includes a second convolutional module. The first convolutional module and the second convolutional module have a symmetrical structure. The encoder is used to extract multi-scale flow field features. The decoder is used to perform feature reconstruction based on the multi-scale flow field features to obtain the reconstruction result. The classifier is used to detect based on the multi-scale flow field features to obtain the hazard detection result. The trained aviation hazard wind detection model is used to detect the radar observation data to be detected, and the hazard wind detection results are obtained. The iterative training of the aviation hazard wind detection model using the radar observation training data and the location of the hazard wind field includes: A sample observation matrix is ​​constructed based on the radar observation training data; The sample observation matrix is ​​input into the aviation hazard wind detection model for processing to obtain the training output results; The physical parameters of the wind field at each sampling point on the aircraft take-off and landing channel are calculated based on the radar observation training data, and physical knowledge priors are constructed based on the physical parameters. The prior to the operation location is determined based on the location of the hazardous wind field. The prior to the operation location is used to describe whether each sampling point on the aircraft take-off and landing channel is a location of a hazardous wind field. A guidance map is generated based on the aforementioned physical knowledge priors and the aforementioned operation position priors; The parameters of the aviation hazard wind detection model are adjusted based on the loss value, which includes a position alignment loss. The position alignment loss is determined based on the difference between the attention weight matrix output by the attention module and the guidance map.

2. The method for detecting aviation hazard winds according to claim 1, characterized in that, The process involves calculating the physical parameters of the wind field at each sampling point on the aircraft takeoff and landing path based on the radar observation training data, and constructing prior physical knowledge based on these parameters, including: Calculate the F-factor and eddy current dissipation rate at each sampling point in the aircraft takeoff and landing tunnel; The sampling points that satisfy the first preset condition of the F factor are marked as 1, and the remaining sampling points are marked as 0, to obtain the first physical prior; The sampling points where the eddy current dissipation rate satisfies the second preset condition are marked as 1, and the remaining sampling points are marked as 0, thus obtaining the second physical prior. The physical knowledge prior is obtained by taking the union of the first physical prior and the second physical prior.

3. The method for detecting aviation hazard winds according to claim 1, characterized in that, The prior determination of the operational location based on the location of the hazardous wind field includes: The sampling point corresponding to the location of the hazardous wind field on the aircraft take-off and landing passage is marked as 1, and other sampling points are marked as 0, thus obtaining the prior of the operation position.

4. The method for detecting aviation hazard winds according to claim 1, characterized in that, Before adjusting the parameters of the aviation hazard wind detection model based on the loss value, the method further includes: The importance vector is obtained by calculating the average value of the attention weight matrix row by row. The guidance map is downsampled using bilinear interpolation to have the same spatial resolution as the importance vector to obtain the target guidance map; The position alignment loss is calculated based on the difference between the target guidance map and the importance vector.

5. The method for detecting aviation hazard winds according to claim 4, characterized in that, The position alignment loss for: ； in, Let be the importance vector. The guide diagram is as follows. The target guidance map is used to characterize the result of the downsampling operation. The square of the L2 norm is used to characterize the L2 norm.

6. The method for detecting aviation hazard winds according to claim 1, characterized in that, The loss values ​​also include reconstruction loss and detection loss. The reconstruction loss is determined based on the difference between the reconstruction result output by the decoder and the sample observation matrix, and the detection loss is determined based on the difference between the training output and the label value.

7. The method for detecting aviation hazard winds according to any one of claims 1-6, characterized in that, The step of using the trained aviation hazardous wind detection model to detect radar observation data and obtaining hazardous wind detection results includes: Acquire the radar observation data to be detected; Select the region of interest in the aircraft takeoff and landing corridor. There are 1 parallel line of sight, each corresponding to a virtual observation line parallel to the aircraft takeoff and landing passage. ; Setting each of the aforementioned parallel lines of sight along the direction of the aircraft takeoff and landing passage equidistant sampling points, ; The radial velocity value of each sampling point is obtained by interpolation from the radar observation data to be detected using the k-nearest neighbor method, and a two-dimensional observation matrix is ​​constructed. The two-dimensional observation matrix is ​​input into the trained aviation hazard wind detection model for detection, and the hazard wind detection result is obtained.

8. An aviation hazard wind detection device, characterized in that, include: The acquisition module is used to acquire radar observation training data and the location of hazardous wind fields within the historical detection period. The radar observation training data includes Doppler velocity sampling data from multiple sampling points covering the direction in front of and behind the aircraft take-off and landing passage. The training module is used to iteratively train the aviation hazardous wind detection model using the radar observation training data and the location of the hazardous wind field to obtain a trained aviation hazardous wind detection model. The aviation hazardous wind detection model includes an encoder, a decoder, and a classifier. The encoder includes a first convolutional module and at least one self-attention module. The decoder includes a second convolutional module. The first convolutional module and the second convolutional module have a symmetrical structure. The encoder is used to extract multi-scale flow field features. The decoder is used to perform feature reconstruction based on the multi-scale flow field features to obtain the reconstruction result. The classifier is used to detect based on the multi-scale flow field features to obtain the hazard detection result. The detection module is used to detect the radar observation data to be detected using the trained aviation hazardous wind detection model to obtain the hazardous wind detection result; The iterative training of the aviation hazard wind detection model using the radar observation training data and the location of the hazard wind field includes: A sample observation matrix is ​​constructed based on the radar observation training data; The sample observation matrix is ​​input into the aviation hazard wind detection model for processing to obtain the training output results; The physical parameters of the wind field at each sampling point on the aircraft take-off and landing channel are calculated based on the radar observation training data, and physical knowledge priors are constructed based on the physical parameters. The prior to the operation location is determined based on the location of the hazardous wind field. The prior to the operation location is used to describe whether each sampling point on the aircraft take-off and landing channel is a location of a hazardous wind field. A guidance map is generated based on the aforementioned physical knowledge priors and the aforementioned operation position priors; The parameters of the aviation hazard wind detection model are adjusted based on the loss value, which includes a position alignment loss. The position alignment loss is determined based on the difference between the attention weight matrix output by the attention module and the guidance map.

9. An electronic device, comprising: A memory, a processor, and a program stored in the memory and executable on the processor; characterized in that the processor is configured to read the program in the memory to implement the steps in the aviation hazard wind detection method as described in any one of claims 1 to 7.

10. A readable storage medium for storing a program, characterized in that, When the program is executed by the processor, it implements the steps in the aviation hazard wind detection method as described in any one of claims 1 to 7.

Images