Jolt prediction correction method and system, electronic equipment and storage medium

Abstract

The invention relates to the technical field of aviation safety prediction, and discloses a bump prediction correction method and system, an electronic device and a storage medium. The method comprises the steps that multi-dimensional flight state data is acquired for classification preprocessing, and anomaly detection and missing value recovery are carried out on a bump intensity value and a vertical overload value; converting a model identifier into a low-dimensional embedded vector, and performing distribution standardization processing on a vertical flight path angle and a horizontal stabilizer distance; constructing a time window based on the preprocessed data, inputting the time-space feature extraction model to carry out multi-dimensional feature fusion and time correlation analysis, and outputting a correction coefficient sequence corresponding to the time window; multiplying the bumping intensity value of each time point in the time window by the corresponding correction coefficient sequence point by point to generate a bumping intensity predicted value sequence in a non-level flight state; and training the spatial-temporal feature extraction model based on a composite loss function comprising a key scene enhancement mechanism and a non-negative guarantee mechanism. The bump prediction precision and reliability can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of aviation safety prediction, and in particular relates to a turbulence prediction correction method, system, electronic equipment and storage medium. Background Art

[0002] Currently, aviation turbulence prediction is mainly based on the equivalent turbulence intensity value (EDR) in level flight. However, in non-level flight attitudes such as climb / descent, changes in the aircraft's vertical path angle will introduce attitude interference components, resulting in significant deviations when the level flight EDR is directly applied to non-level flight attitudes.

[0003] In the existing technology, the traditional correction method uses a linear compensation model. However, in practice, the relationship between the attitude component and the actual turbulence component is affected by the nonlinear coupling of multiple factors such as the aerodynamic characteristics of the aircraft model, the size of the path angle, and the turbulence intensity. The linear model is difficult to accurately represent the complex relationship.

[0004] Therefore, the present application provides a bump prediction correction method to solve the above technical problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a turbulence prediction correction method, system, electronic device and storage medium to solve the technical problem in the prior art that the factors represented during prediction are relatively simple and multi-dimensional prediction cannot be performed, resulting in low prediction accuracy.

[0006] In order to solve the above technical problems, the present invention provides a turbulence prediction and correction method, comprising: Acquiring multi-dimensional flight status data, wherein the multi-dimensional flight status data includes a turbulence intensity value in a level flight state, a vertical overload value in a non-level flight state, a vertical path angle, an aircraft model identifier, and a horizontal stabilizer distance; performing classification preprocessing on the multi-dimensional flight status data, including performing anomaly detection and missing value recovery on the turbulence intensity value and the vertical overload value, converting the aircraft model identifier into a low-dimensional embedding vector, and performing distribution normalization processing on the vertical path angle and the horizontal stabilizer distance; Constructing a time window based on the pre-processed multi-dimensional flight status data, inputting the spatiotemporal feature extraction model to perform multi-dimensional feature fusion and time correlation analysis, and outputting a correction coefficient sequence corresponding to the time window; Multiply the turbulence intensity value at each time point in the time window by the corresponding correction coefficient sequence at each time point to generate a turbulence intensity prediction value sequence in the non-level flight state; The spatiotemporal feature extraction model is trained based on a composite loss function including a key scene reinforcement mechanism and a non-negativity guarantee mechanism to optimize the prediction accuracy of the correction coefficient sequence.

[0007] In some specific embodiments, the multi-dimensional flight status data is subjected to classification preprocessing, including anomaly detection and missing value recovery of the turbulence intensity value and the vertical overload value, converting the aircraft model identifier into a low-dimensional embedding vector, and performing distribution normalization processing on the vertical path angle and the horizontal stabilizer distance, further comprising: Filter out bump intensity values ​​greater than a preset intensity range and vertical overload values ​​greater than a preset load range; Based on the time series continuity interpolation method, the missing turbulence intensity values ​​and vertical overload values ​​are restored through the data correlation of adjacent time points; Map the model identifier to a continuous vector space of fixed dimension to generate an embedding vector; A linear transformation with zero mean and unit variance is performed on the vertical path angle and the horizontal stabilizer distance.

[0008] In some specific embodiments, constructing a time window based on the pre-processed multi-dimensional flight status data further includes: Set the length of the continuous time window that covers the complete turbulence event life cycle; A sliding window mechanism with a preset overlap ratio is used to segment the time series, maintaining the temporal continuity of adjacent windows; splicing the pre-processed multi-dimensional flight status data along the feature dimension to generate a combined feature vector; Constructs a time window of three-dimensional tensor samples, with dimensions including time step dimension, combined feature dimension, and single channel dimension.

[0009] In some specific embodiments, the input spatiotemporal feature extraction model is used to perform multi-dimensional feature fusion and time correlation analysis, and the correction coefficient sequence corresponding to the time window is output, further comprising: The feature compression path involves extracting long-term trend features through multi-level time dimension downsampling operations, where each level contains convolution operations at different time scales, gradually expanding the temporal receptive field; Feature fusion layer, including aggregating the long-term trend features and eliminating information redundancy to form global features; The feature recovery path includes reconstructing the temporal resolution through upsampling operations in the temporal dimension, fusing global features with local transient features using a cross-level feature connection mechanism, and generating spatiotemporal features. The coefficient mapping layer converts the spatiotemporal features into a complete correction coefficient sequence in the time dimension.

[0010] In some specific embodiments, the feature compression path includes extracting long-term trend features through multi-level time dimension downsampling operations, each level including convolution operations at different time scales to gradually expand the temporal receptive field, and further includes: Perform a sequence of temporal convolution operations with increasing dilation rates, gradually expanding the temporal receptive field by adjusting the convolution kernel sampling interval; Introducing nonlinear activation functions and batch normalization after each level of convolution operation to enhance feature expression capabilities; Inserting a time dimension compression operation after a specified convolutional layer reduces the temporal resolution while increasing the feature depth; The output is a composite feature tensor including the long-term aerodynamic impact of the aircraft model, the evolution trend of the vertical path angle, and the baseline level of the turbulence intensity value.

[0011] In some specific embodiments, multiplying the turbulence intensity value at each time point in the time window by the corresponding correction coefficient sequence time point by time point to generate a turbulence intensity prediction value sequence for a non-level flight state further includes: Get the turbulence intensity value at each time point in the current time window; Call the trained spatiotemporal feature extraction model and output the correction coefficient sequence corresponding to the time window; Perform a dot multiplication operation on the turbulence intensity value at the same time point and the corresponding correction coefficient to generate a turbulence intensity prediction value; Outputs a real-time updated sequence of turbulence intensity prediction values.

[0012] In some specific embodiments, the spatiotemporal feature extraction model is trained based on a composite loss function including a key scene reinforcement mechanism and a non-negativity guarantee mechanism to optimize the prediction accuracy of the correction coefficient sequence, further comprising: Scenario weighting mechanism, including dynamic calculation of error weight coefficients at each time point based on turbulence intensity and absolute value of vertical path angle, to enhance the error contribution of key scenarios; The error term includes the mean squared error between the predicted turbulence intensity value after the scenario weighting mechanism is implemented and the actual turbulence intensity value; Constraints, including penalty functions for negative values ​​in the correction coefficient series; The comprehensive optimization term includes combining the error term and the constraint term in a preset ratio to generate a composite loss function.

[0013] Based on the same concept, the present invention also provides a turbulence prediction and correction system, comprising: a multi-dimensional data acquisition module configured to acquire multi-dimensional flight status data, wherein the multi-dimensional flight status data includes a turbulence intensity value in a level flight state, a vertical overload value in a non-level flight state, a vertical path angle, an aircraft type identifier, and a horizontal stabilizer distance; a multi-dimensional data processing module configured to perform classification pre-processing on the multi-dimensional flight status data, including performing anomaly detection and missing value recovery on the turbulence intensity value and the vertical overload value, converting the aircraft model identifier into a low-dimensional embedding vector, and performing distribution normalization processing on the vertical path angle and the horizontal stabilizer distance; a correction coefficient sequence generation module configured to construct a time window based on the preprocessed multi-dimensional flight status data, input the spatiotemporal feature extraction model to perform multi-dimensional feature fusion and time correlation analysis, and output a correction coefficient sequence corresponding to the time window; a turbulence intensity prediction value sequence generation module configured to multiply the turbulence intensity value at each time point within the time window by the corresponding correction coefficient sequence at each time point to generate a turbulence intensity prediction value sequence for a non-level flight state; The spatiotemporal feature extraction model optimization module is configured to train the spatiotemporal feature extraction model based on a composite loss function including a key scene reinforcement mechanism and a non-negativity guarantee mechanism to optimize the prediction accuracy of the correction coefficient sequence.

[0014] Based on the same concept, the present invention also provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of a turbulence prediction and correction method.

[0015] Based on the same concept, the present invention also provides a computer-readable storage medium storing a computer program executable by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of a turbulence prediction and correction method.

[0016] Compared with the prior art, the beneficial effects are: The present invention discloses a bump prediction and correction method, system, electronic device and storage medium, which adopt multi-dimensional features for expression, optimize computational efficiency while ensuring physical rationality, and improve prediction accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings: Figure 1 This is a flowchart of a bump prediction and correction method according to some specific embodiments of the present invention; Figure 2 This is a schematic structural diagram of a bump prediction and correction system according to some specific embodiments of the present invention; Figure 3is a schematic structural diagram of an electronic device according to some specific embodiments of the present invention; In the figure, 710 is a processor; 720 is a memory; 730 is an input device; and 740 is an output device. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of this application more clear, this application will be further described in detail below with reference to the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0019] The terms used in the examples of this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms "a," "the," and "the" used in the examples of this application and the appended claims are also intended to include plural forms, and unless the context clearly indicates otherwise, "a plurality" generally includes at least two.

[0020] It should be understood that the term "and / or" as used herein is merely a description of the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0021] It should be understood that although the terms first, second, third, etc. may be used to describe in the embodiments of the present application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, without departing from the scope of the embodiments of the present application, the first may also be referred to as the second, and similarly, the second may also be referred to as the first.

[0022] As used herein, the words "if" and "if" may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to the determination" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)," depending on the context.

[0023] It should also be noted that the terms "include," "comprises," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a product or device comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such product or device. In the absence of further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the product or device comprising the element.

[0024] It should be noted in particular that any symbols and / or numbers in the specification that are not marked in the accompanying drawings are not drawing marks.

[0025] Reference Figure 1 , a turbulence prediction correction method, comprising: S101, acquiring multi-dimensional flight status data, wherein the multi-dimensional flight status data includes a turbulence intensity value in level flight, a vertical overload value in non-level flight, a vertical path angle, an aircraft type identifier, and a horizontal stabilizer distance; S102, performing classification preprocessing on the multi-dimensional flight status data, including performing anomaly detection and missing value recovery on the turbulence intensity value and the vertical overload value, converting the aircraft model identifier into a low-dimensional embedding vector, and performing distribution normalization processing on the vertical path angle and the horizontal stabilizer distance; S103, constructing a time window based on the pre-processed multi-dimensional flight status data, inputting the time-space feature extraction model to perform multi-dimensional feature fusion and time correlation analysis, and outputting a correction coefficient sequence corresponding to the time window; S104, multiplying the turbulence intensity value at each time point in the time window by the corresponding correction coefficient sequence at each time point to generate a turbulence intensity prediction value sequence for a non-level flight state; S105 , training the spatiotemporal feature extraction model based on a composite loss function including a key scene reinforcement mechanism and a non-negativity guarantee mechanism to optimize the prediction accuracy of the correction coefficient sequence.

[0026] Specifically, in an embodiment of the present invention, multi-dimensional flight status data including level flight turbulence intensity value, non-level flight vertical overload value, vertical path angle, aircraft type identification and horizontal stabilizer distance are obtained; abnormal points of turbulence intensity value and vertical overload value are eliminated based on physical threshold range and missing values ​​are supplemented by time adjacent interpolation; aircraft type identification is converted into a dense embedding vector of fixed dimension through vector space mapping to retain aerodynamic similarity; vertical path angle and horizontal stabilizer distance are normalized by performing linear transformation with zero mean and unit variance; based on covering the complete turbulence cycle A three-dimensional sample tensor is constructed based on the time window length and the preset overlap rate; the samples are input into the spatiotemporal feature extraction model, and the global features including the long-term characteristics of the aircraft model and the path trend are extracted through the compression path. After fusion through the bottleneck layer, the time resolution is reconstructed by the recovery path combined with the jump connection, and the final output is a correction coefficient sequence; the level flight turbulence intensity value at each time point is multiplied by the corresponding correction coefficient to generate the non-level flight prediction value; the model is trained using a composite loss function, which contains dynamic weight terms determined by the turbulence intensity and the path angle to ensure the accuracy of key scenarios, and at the same time, the physical rationality of the prediction value is guaranteed by the non-negative constraint terms.

[0027] For example, in the climb phase: 60 seconds of continuous data are acquired, including the level flight EDR sequence of 0.3g, 0.4g, 0.5g... overload values ​​of 1.2g, 1.3g, 1.4g... path angles of 4°, 5°, 6°... aircraft model identification B737 horizontal stabilizer distance of 12 meters; outliers with EDR greater than 2.0g or overload greater than 3.0g are filtered out, and missing values ​​are interpolated using the mean of the previous and next moments; the B737 is mapped to an 8-dimensional vector [-0. 2,0.5,1.1,-0.3,0.7,0.4,-1.2,0.9]; the original mean of the flight path angle is 5.3° and the standard deviation is 1.8. After normalization, the third second value of 6° is converted to (6-5.3) / 1.8=0.39; the 60-second window sample tensor dimension is constructed as time step 60×feature 12×channel 1, where the feature contains 8-dimensional aircraft model vector + normalized flight path angle + normalized stabilizer distance + EDR + overload value; after inputting the spatiotemporal model, the compression path performs three-level processing. The first level uses a 3×1 convolution kernel to sample every 1 unit and output 30×12×2 features. The second level outputs 15×12×4 every 2 units and the third level outputs 15×12×8 every 4 units; the recovery path upsamples the third level output to 30×12×4 and concatenates it with the first level activation feature. After 3×1 convolution, the output is 30×12×4. The original feature is further upsampled and concatenated to output a 60×1 correction coefficient sequence, such as [1.1,1 .2,1.3...]; Multiply the EDR 0.5g in level flight at the 3rd second by the coefficient 1.3 to obtain the predicted value 0.65g; In the loss function calculation, the weight of the 3rd second = 0.5×0.5g+0.3×|6°|+0.2=0.25+1.8+0.2=2.25 The error term is 2.25×(0.65g-true value 0.62g)²=0.002 The non-negative constraint term imposes a penalty on negative coefficients The total loss is the sum of the weighted error and the constraint term.

[0028] In some of the applications, the multi-dimensional flight status data is classified and pre-processed, including anomaly detection and missing value recovery of the turbulence intensity values ​​and the vertical overload values, converting the aircraft model identification into a low-dimensional embedding vector, and performing distribution normalization processing on the vertical path angle and the horizontal stabilizer distance, including filtering out turbulence intensity values ​​greater than a preset intensity range and vertical overload values ​​greater than a preset load range; based on the time series continuity interpolation method, the missing turbulence intensity values ​​and vertical overload values ​​are restored through the data correlation of adjacent time points; the aircraft model identification is mapped to a continuous vector space of fixed dimension to generate an embedding vector; and a zero-mean unit variance linear transformation is performed on the vertical path angle and the horizontal stabilizer distance.

[0029] It can be understood that reasonable intervals for turbulence intensity and vertical overload are set, and abnormal data points exceeding the intervals are filtered out; an interpolation method based on the continuity of time series is adopted, and the numerical correlation characteristics of adjacent time points are used to restore the missing turbulence intensity values ​​and vertical overload values; the discrete aircraft model identification is converted into a continuous vector representation of fixed dimension through vector space mapping technology, and a low-dimensional embedding vector that retains the similarity relationship between aircraft models is generated; a standardized linear transformation is performed on the vertical path angle and the horizontal stabilizer distance, and a distribution conversion with zero mean and unit variance is achieved by subtracting the mean and dividing by the standard deviation.

[0030] For example, using the B737 data, the turbulence intensity value of 2.5g at the 15th second, exceeding the preset 2.0g threshold, is detected as an anomaly and discarded. The turbulence intensity value at the 20th second is missing, so the arithmetic mean of the adjacent 1.2g at the 19th second and 1.3g at the 21st second is used for interpolation (1.25g). The aircraft model identifier B737 is mapped into an 8-dimensional embedding vector [0.2, -0.5, 1.1, 0.7, -0.3, 0.9, 0.4, -0.1]. The original vertical path angle data sequence is 3.5°, 4.2°, 5.1°... The 60-second window mean is 4.3°, with a standard deviation of 0.8. The 3rd second value of 5.1° is normalized to (5.1-4.3) / 0.8=1.0. The horizontal stabilizer distance is fixed at 12 meters, with a mean of 10.5 meters and a standard deviation of 2.5 meters for the entire data set, resulting in a normalized value of (12-10.5) / 2.5=0.6.

[0031] In some of these applications, a time window is constructed based on the preprocessed multi-dimensional flight status data, including setting the length of a continuous time window that covers the complete life cycle of a turbulence event; a sliding window mechanism with a preset overlap ratio is used to split the time series to maintain the time continuity of adjacent windows; the preprocessed multi-dimensional flight status data is spliced ​​along the feature dimension to generate a combined feature vector; and a time window of three-dimensional tensor samples is constructed, where the dimensions include the time step dimension, the combined feature dimension, and the single channel dimension.

[0032] It can be understood that the length of the continuous time window covering the entire life cycle of the turbulence event is set to ensure the capture of the entire process of turbulence initiation, development and disappearance; the sliding window mechanism with a preset overlapping ratio is used to segment the time series data, and the time continuity of adjacent windows is maintained by controlling the window movement step to avoid information fragmentation; the standardized vertical path angle, standardized horizontal stabilizer distance, low-dimensional embedding vector, level flight turbulence intensity value and vertical overload value are spliced ​​and combined along the feature dimension to form a multi-dimensional feature vector; a three-dimensional tensor sample structure including the time step dimension, combined feature dimension and single channel dimension is constructed to retain the time series characteristics and feature correlation of the original data.

[0033] For example, taking continuous flight data as an example: a 60-second time window length is set to cover the typical turbulence event cycle; a 50% overlapping sliding window is used, that is, it slides every 30 seconds, the first window contains data from the 1st to 60th seconds, and the secondary window contains data from the 31st to 90th seconds; the pre-processed features include an 8-dimensional aircraft model embedding vector [0.2, -0.5, 1.1, 0.7, -0.3, 0.9, 0.4, -0.1], a standardized path angle value of 1.0, a standardized stabilizer distance value of 0.6, a level flight turbulence intensity value of 1.25g, and a vertical overload value of 1.3g, forming a 12-dimensional combined feature vector; a 3-dimensional tensor sample is constructed whose time step dimension 60 corresponds to a 60-second sampling point, the combined feature dimension 12 corresponds to 12 feature parameters, the single channel dimension 1 represents a single data stream, and the first window sample is represented as a 60×12×1 tensor structure.

[0034] In some of these applications, a spatiotemporal feature extraction model is input for multi-dimensional feature fusion and temporal correlation analysis, and a correction coefficient sequence corresponding to the time window is output, including a feature compression path, which extracts long-term trend features through multi-level time dimension downsampling operations, each level contains convolution operations at different time scales, and gradually expands the temporal receptive field; a feature fusion layer, which aggregates the long-term trend features and eliminates information redundancy to form global features; a feature recovery path, which reconstructs the temporal resolution through time dimension upsampling operations, and combines the cross-level feature connection mechanism to fuse global features and local transient features to generate spatiotemporal features; a coefficient mapping layer, which converts spatiotemporal features into a correction coefficient sequence with a complete time dimension.

[0035] It can be understood that the feature compression path extracts long-term trend features through multi-level time dimension downsampling operations, and each level uses convolution operations of different time scales to gradually expand the time receptive field to capture the aerodynamic characteristics of the aircraft model and the evolution trend of the path angle; the feature fusion layer aggregates the long-term trend features output by the compression path and eliminates redundancy through the information concentration mechanism to form a fusion feature that characterizes the global flight state; the feature recovery path reconstructs the time resolution step by step through the time dimension upsampling operation and combines the cross-level feature connection mechanism to splice and fuse the fusion features with the local transient features of the compression path to generate spatiotemporal features containing global constraints and local details; the coefficient mapping layer converts the spatiotemporal features into a correction coefficient sequence of the same length as the original time window through a one-dimensional convolution operation, and each coefficient corresponds to a nonlinear correction factor at a specific time point.

[0036] For example, taking a 12-dimensional feature tensor with a 60-second time window as an example: the first stage of the feature compression path uses a 3-time-point scale convolution kernel with a sampling interval of 1 unit, and inputs a 60x12x1 tensor to output a 30x12x2 feature tensor; the second stage uses a 3-time-point scale convolution kernel with a sampling interval of 2 units, and inputs a 30x12x2 to output a 15x12x4 feature tensor; the third stage uses a 3-time-point scale convolution kernel with a sampling interval of 4 units to output a 15x12x8 feature tensor; the feature fusion layer performs 4-interval convolution on the output of the third stage to generate a 15x12x8 global feature; the feature recovery path first upsamples the global feature to a 30x12x4 tensor and concatenates it with the output of the first stage to form a 30x12x6 tensor, and outputs a 30x12x4 feature after 2-interval convolution; continue to upsample to 60x12x2 and concatenate with the original input to form a 60x12x3 tensor, and output a 60x12x2 spatiotemporal feature after one-interval convolution; the coefficient mapping layer converts the spatiotemporal feature into a 60x1 correction coefficient sequence through a 1x1 convolution, such as a time point 3 corresponding to a coefficient 1.32 and a time point 58 corresponding to a coefficient 0.98.

[0037] In some applications, the feature compression path includes extracting long-term trend features through multi-level time dimension downsampling operations, each level containing convolution operations of different time scales, gradually expanding the time receptive field, including performing a sequence of time convolution operations with increasing dilation rates, gradually expanding the time receptive field range by adjusting the convolution kernel sampling interval; introduce a nonlinear activation function and batch normalization processing after each convolution operation to enhance feature expression capability; insert a time dimension compression operation after a specified convolution level, reduce the time resolution while increasing the feature depth; output a composite feature tensor containing the long-term aerodynamic influence of the model, the vertical diameter angle evolution trend, and the baseline level of the intensity value.

[0038] It can be understood that the sequence of time convolution operations with increasing dilation rates is implemented, the receptive field is expanded exponentially by adjusting the convolution kernel sampling interval; introduce a nonlinear activation function after each convolution operation for feature transformation and stabilize data distribution through batch normalization processing to enhance model expression capability; insert a time dimension compression operation after a specified convolution level, simultaneously reduce the time resolution and increase the feature depth; finally output a multi-dimensional composite feature tensor containing the long-term aerodynamic characteristics of the model, the vertical diameter angle evolution trend, and the baseline level of the intensity value.

[0039] For example, take the input 60x12x1 sample tensor as an example: the first level convolution adopts a 3x1 convolution kernel with a sampling interval of 1 unit, and the receptive field covers 3 time points, and the output dimension is 30x12x2; the second level adopts a 3x1 convolution kernel with a sampling interval of 2 units, and the receptive field is expanded to 7 time points, and the output is 15x12x4 features; the third level adopts a 3x1 convolution kernel with a sampling interval of 4 units, and the receptive field reaches 15 time points, and the output is 15x12x8 features; after each level of convolution, a ReLU activation function is executed and batch normalization is performed, such as the feature value of the 5th time point in the second level output being adjusted from 1.25 to 0.82 after BN; a max pooling operation is inserted after the third level to fix the time resolution to 15 points; the final output composite feature tensor contains model influence components such as the 3rd channel value 0.73 representing the B737 aerodynamic characteristics, the flight path angle trend component such as the 7th channel value -0.15 indicating a continuous climbing state, and the turbulence baseline component such as the 12th channel value 1.28 reflecting the current flight segment strong turbulence background.

[0040] In some applications, the turbulence intensity values at each time point within the time window are multiplied by the corresponding correction coefficient sequence point by point to generate a turbulence intensity prediction value sequence in a non-level flight state, including obtaining the turbulence intensity values at each time point in the current time window, calling the trained spatio-temporal feature extraction model, and outputting the correction coefficient sequence of the corresponding time window; the turbulence intensity value at the same time point is multiplied by the corresponding correction coefficient to generate a turbulence intensity prediction value; and the real-time updated turbulence intensity prediction value sequence is output.

[0041] It can be understood that the level flight turbulence intensity value sequence at all sampling times in the current time window is obtained; the trained spatio-temporal feature extraction model is called to output the correction coefficient sequence that is completely aligned with the time window; the level flight turbulence intensity value at each same sampling time is multiplied by the correction coefficient to generate an instantaneous prediction value; and the non-level flight turbulence intensity prediction sequence in the continuous time domain is output through a sliding window update mechanism.

[0042] For example, take the 31-90 second time window as an example: obtain the level flight turbulence intensity value sequence [0.3g, 0.4g, 0.5g...] (60 values); call the trained model to output the correction coefficient sequence [1.1, 1.2, 1.3...] (60 dimensions); perform calculation at the 3rd second position (33rd second in the window): 0.5g x 1.3 = 0.65g; calculate at the 58th second position (88th second in the window): 0.8g x 0.98 = 0.784g; output the complete prediction sequence [0.33g, 0.48g, 0.65g...]; when the window slides to the 61-120 second, the new prediction sequence is output in real time.

[0043] In some of these applications, the spatiotemporal feature extraction model is trained based on a composite loss function that includes a key scenario reinforcement mechanism and a non-negativity guarantee mechanism to optimize the prediction accuracy of the correction coefficient sequence, including a scenario weighting mechanism, which dynamically calculates the error weight coefficient at each time point based on the turbulence intensity value and the absolute value of the vertical path angle to enhance the error contribution of the key scenario; an error term, which includes the mean squared error between the turbulence intensity predicted value and the actual turbulence intensity value after executing the scenario weighting mechanism; a constraint term, which includes applying a penalty function to negative values ​​in the correction coefficient series; and a comprehensive optimization term, which includes combining the error term and the constraint term in a preset proportion to generate a composite loss function.

[0044] It can be understood that the scenario weighting mechanism dynamically calculates the error weight coefficient at each time point through the turbulence intensity value and the absolute value of the vertical path angle, thereby strengthening the error contribution weight of the key flight scenario; the error term calculates the mean square error between the turbulence intensity prediction value and the true value after scenario weighting; the constraint term applies a penalty function to the negative values ​​in the correction coefficient sequence to ensure physical rationality; the comprehensive optimization term linearly combines the error term and the constraint term in a preset proportion to generate a differentiable composite loss function.

[0045] For example, at the 3-second point in time, the level flight turbulence intensity is 0.5g, and the absolute vertical path angle is 6°. The dynamic weight coefficient is calculated as 0.5 × 0.5g + 0.3 × 6° + 0.2 × 1 = 0.25 + 1.8 + 0.2 = 2.25. The predicted value at this point is 0.65g, and the true value is 0.62g. The squared error is (0.65-0.62)² = 0.0009, and the weighted error is 2.25 × 0.0009 = 0.002025. If the correction factor at this point is -0.1, the constraint term is 50 × max(0,-(-0.1)) = 5.0. Assuming the sum of the error terms is 0.15 and the sum of the constraint terms is 8.0, the composite loss is 0.15 + 8.0 = 8.15. The loss for the entire time window is the average of all points.

[0046] Another embodiment of the present invention is described below: Input features: Level flight EDR ( ): Turbulence signal of event data recorder in level flight (sampling frequency 1Hz, unit ); Non-level flight auxiliary data: overload data ( ): Vertical overload sensor data (reflects instantaneous force, sampling frequency 1Hz); vertical path angle ( ): vertical speed and ground speed The ratio of , dimensionless, sampling frequency 1Hz); aircraft model: such as B737, A320, C919, etc. (categorical variables); distance from horizontal stabilizer to lift center ( ): Model fixed structure parameters (unit: meters). Label data: True non-level flight EDR ( ): Tumble signal recorded by EDR in non-level flight state (climb / descent) (sampling frequency 1 Hz, aligned with level flight EDR time).

[0047] Outlier and missing value processing: Outlier: remove values beyond the physical reasonable range in EDR and overload data (such as EDR > 2.0g, overload > 3.0g); Missing value: complete by linear interpolation method (since the data is a time series, the correlation between adjacent time points is strong). Model code optimization: embedded coding scheme: map the model to an integer index (such as B737→1, A320→2, C919→3…); convert the index to an 8-dimensional dense vector through the embedding layer ( ), and the vector parameters are optimized through model training, so that the vector distance of similar models (such as the same manufacturer) is closer; feature standardization: standardize numerical features ( ).

[0048] where is the mean, is the standard deviation (calculated based on the training set to avoid data leakage).

[0049] Time window division (sample construction): window size (60 sampling points), sliding step (30 sampling points), sample tensor construction: each sample is a three-dimensional tensor: ; where (time step), (feature dimension: 8-dimensional model embedding vector + 4 numerical features ); the label is the true non-level flight EDR corresponding to the time window: .

[0050] The U-shaped expansion convolution module takes the time window sample as input and outputs the correction coefficient of each time step, and finally predicts the non-level flight EDR through the “ ” layer.

[0051] The contraction path (downsampling + dilated convolution) compresses the time dimension, expands the receptive field, and captures global features (such as the long-term impact of the model, trend changes), which provides a basis for distinguishing between “attitude additional components” and “true turbulence components”.

[0052] Global features specifically include: the persistent influence of model aerodynamic characteristics (such as the attitude additional components of large aircraft are more gentle than those of small aircraft); the trend of vertical range angle (such as when continuously climbing, If the turbulence is positive and stable, the correction factor needs to be offset as a whole); the baseline level of the level flight EDR (if the overall turbulence in a certain segment is strong, the correction factor needs to be adapted to the baseline).

[0053] Structural and dimensional design (with layer as an example): Layer 1: enter: ; is the input tensor, the dimension is Among them, 60 represents the time step, that is, the data of 60 consecutive time points; 12 represents the number of features, including different features such as level flight EDR, overload, vertical path angle, etc.; 1 represents the number of channels, which is a single channel at the initial input. Dilated convolution: kernel size 3×1 (3 points in the time direction, 1 point in the feature direction), dilation rate , input channel 1 → output channel 2; kernel size 3×1 considers 3 adjacent time points in the time direction and only considers 1 feature point in the feature direction. Indicates that during the convolution process, the interval between convolution kernel elements is 1. The number of input channels is 1, and the number of output channels becomes 2 after convolution. Activation and batch normalization:

[0054] First, batch normalization (BN) is performed on the result of the dilated convolution. Then the ReLU activation function is used. Max pooling (step size 2): Output (time step is halved and number of channels is doubled).

[0055] Layer 2: Dilated convolution: kernel size 3×1, dilation rate , input channel 2 → output channel 4; the number of input channels is 2, and the number of output channels becomes 4, further increasing the diversity of features. Output after activation, batch normalization, and pooling: .

[0056] Layer 3: Dilated convolution: kernel size 3×1, dilation rate , input channel 4 → output channel 8; the number of input channels is 4, the number of output channels becomes 8, and the expressive power of features is further increased. Output after activation and batch normalization (no pooling, as bottleneck layer input): .

[0057] The bottleneck layer fuses the global features of the contraction path and further compresses redundant information. Calculation: ; : It is the output feature map of the third layer of the contraction path. It contains the feature information extracted by the previous layers. :Here Indicates the expansion rate, the value is 4. When , the interval between convolution kernel elements is 4. : Indicates the dilated convolution operation. Use the dilation rate The dilated convolution of the input Convolution is performed to extract broader features. : Indicates batch normalization operation. : represents the activation function. is the output feature map of the bottleneck layer. : represents the dimension of the output feature map. Among them, 15 represents the time step, 12 represents the number of features, and 8 represents the number of channels.

[0058] The expansion path (upsampling + skip connection) restores the global features of the bottleneck layer to the original time scale through upsampling, and the local features of the corresponding layer of the contraction path (such as instantaneous overload peak, mutations) splicing to achieve the fusion of "global constraints + local details".

[0059] Structural and dimensional design (corresponding to the opposite direction of shrinkage path): Layer 2: Transposed convolution (upsampling): kernel size 3×1, stride 2, input channels 8 → output channels 4, Upsample to The kernel size of 3×1 means that 3 adjacent time points are considered in the time direction and 1 feature point is considered in the feature direction. The step size of 2 is the step size of upsampling. By setting the step size to 2, the input (Dimensions are ) upsampled to , the time step is changed from 15 to 30, and the number of channels is changed from 8 to 4. Skip connection: activation features of the first layer of the contraction path Splicing (number of channels 4+2=6); dilated convolution: kernel size 3×1, dilation rate , input channel 6 → output channel 4, output The kernel size 3×1 is used to capture local correlation in the temporal direction. The interval between the convolution kernel elements is 2, thereby expanding the receptive field. The number of input channels is 6, and the number of output channels becomes 4 after convolution, and the output is , whose dimensions are .

[0060] Layer 1: Transposed convolution: stride 2, input channel 4 → output channel 2, upsampling to ; The step size is 2, and the input features (dimension is ) upsampled to , the time step is changed from 30 to 60, and the number of channels is changed from 4 to 2. Skip connection: the input feature of the 0th layer of the contraction path Concatenation (number of channels 2+1=3); input features of the 0th layer of the contraction path The dimension is , and compare it with the upsampled features (dimension is ) is concatenated in the channel dimension to obtain a new feature with a channel number of 2+1=3. Dilated convolution: kernel size 3×1, dilation rate , input channel 3 → output channel 2, output The kernel size 3×1 is used to capture local temporal correlations. This is a standard convolution operation. The number of input channels is 3, and after convolution the number of output channels becomes 2, resulting in the output , whose dimensions are .

[0061] Layer 0: Dilated convolution: kernel size 3×1, dilation rate , input channel 2 → output channel 1, output The kernel size 3×1 is used to capture local temporal correlations. This is a standard convolution operation. The number of input channels is 2, and after convolution the number of output channels becomes 1, resulting in the final output , whose dimensions are , the same dimension as the original input data.

[0062] Correction coefficient output layer, which maps the fused features into correction coefficients for each time step .calculate: ; : It is the output feature map of the 0th layer of the expansion path, which is the fusion feature obtained after all the previous processing steps. : represents a 1×1 convolution operation. : is the output of the correction coefficient output layer, which is of dimension Where 60 represents the time step, which is the same as the time step of the original input data; 1 represents the number of channels, and each time step corresponds to a correction coefficient. : Represents the dimension of the output tensor, that is, 60 time steps, each time step corresponds to a correction coefficient.

[0063] The predicted non-level flight EDR is: ; : Indicates the predicted non-level flight EDR, which is about time function. : Indicates EDR in level flight, that is, the original level flight turbulence data. : represents the correction coefficient, which is the output of the correction coefficient output layer and time The corresponding scalar value.

[0064] Model training and optimization: The data set is divided into flight segments (to avoid the same flight data appearing in both training and test sets): 70% for training set, 15% for validation set, and 15% for test set.

[0065] Loss function: ; : Represents the loss function, which is used to measure the difference between the model prediction results and the actual results. : Indicates the number of samples, that is, the total number of samples included in the training set. : represents the number of time steps for each sample, each sample has time steps. :Indicates that The first sample The weight of each time step is used to perform differentiated error penalties on different samples and time steps. : represents the first The sample in The non-level flight EDR value of the time step. :Indicates the The sample in The actual non-level flight EDR value for each time step. : Represents the weight coefficient of the non-negative constraint term, which is used to control the strength of the non-negative constraint. : represents the correction coefficient, which is the output of the correction coefficient output layer and time The corresponding scalar value. :represents a non-negative constraint function, when When , the value of this item is , otherwise 0.

[0066] Weighted error term: weight ,in (Strong Bump Weight), (Large path angle weight), (Basic weight), strengthens the error penalty in key scenarios; :It is Sample No. The weight of each time step. : is the strong turbulence weight, the value is 0.5. When the EDR ( ) is large, indicating that the model is in a strong turbulence scenario. At this time, the weight will increase accordingly and the penalty for error will be heavier to improve the prediction accuracy of the model in strong turbulence scenarios. : is the maximum path angle weight, value is 0.3. The absolute value of the path angle When it is larger, it means that it is in a large path angle scenario, the weight will increase, and the penalty for errors in this scenario will be strengthened. : is the basic weight, with a value of 0.2, which provides a basic weight for each time step to ensure that there is a certain error penalty even in non-critical scenarios. :It is Sample No. The EDR value of level flight at each time step. :It is Sample No. The absolute value of the path angle in time steps. Non-negative constraint: ,avoid (Ensure that the predicted non-level flight EDR is non-negative and conforms to the physical meaning of turbulence intensity). : is the weight coefficient of the non-negative constraint term, with a value of 50, which is used to control the strength of the non-negative constraint. : is the correction coefficient. The purpose of the non-negative constraint is to ensure .

[0067] Optimizer and regularization, optimizer: Adam (learning rate ); Regularization: Dropout (probability 0.2, to prevent overfitting) + weight decay ( , constraining the parameter size). Training process: Iterate 50 rounds, calculating the training set and validation set loss in each round; early stopping strategy: stop training when the validation set loss increases for 5 consecutive rounds, and save the optimal model (the model with the smallest validation set loss).

[0068] Model validation and evaluation: EDR prediction error: Root Mean Square Error (RMSE): ; : represents the root mean square error. : Indicates the number of samples. : represents the number of time steps for each sample, each sample has time steps. : Indicates the EDR value predicted by the model. : Indicates the actual EDR value. : represents the summation operation, which sums the differences between the predicted values ​​and the true values ​​of all time steps of all samples. : represents the square of the difference between the predicted value and the true value. : represents taking the average of the sum result, getting the average square error. : represents taking the square root of the average square error, getting the root mean square error, whose unit is the same as EDR, making it easy to intuitively understand the size of the error.

[0069] Mean Absolute Error (MAE): ; : represents the mean absolute error. : number of samples. : number of time steps for each sample. : EDR value predicted by the model. : true EDR value. : sum operation, summing the absolute difference between the predicted value and the true value for all samples and all time steps. : represents the absolute difference between the predicted value and the true value, and the absolute operation makes the positive and negative of the error not affect the result, only focusing on the size of the error. : taking the average of the sum result, getting the mean absolute error, reflecting the average deviation between the predicted value and the true value. Determination coefficient (R^2): .

[0070] : represents the determination coefficient. : represents the residual sum of squares, which is the sum of the squares of the difference between the model's predicted value and the true value, reflecting the size of the model's prediction error. : represents the total sum of squares, where is the mean of the true value, and the total sum of squares reflects the dispersion of the true value. : represents the proportion of the variation that the model can explain to the total variation. If the model's prediction is completely accurate, then the residual sum of squares is 0, ; if the model's prediction is the same as the mean of the true value, then ; if the model's prediction is worse than the mean of the true value, then may be negative. The requirement .

[0071] Key scenario performance: strong turbulence sample: , requiring its to be 10% higher than the overall level; : represents the sample whose EDR in level flight is greater than defined as a strong turbulence sample. Here, is the acceleration of gravity, is the EDR value in level flight. ​

[0072] Large angle sample: , requires it 10% higher than the overall level.

[0073] : samples with the absolute value of the flight path angle greater than 5° are defined as large angle samples. is the flight path angle.

[0074] This embodiment is based on the flat EDR, combined with the overload, vertical flight path angle, aircraft type, horizontal stabilizer distance and other data in the non-flat state. First, the model characteristics are optimized by embedding coding, and the samples are constructed through standardization and time window division. Then, the U-shaped expansion convolution module is used to capture global features, fuse information in the bottleneck layer, and recover details by combining the jump connection in the expansion path, and output the correction coefficient. Through the weighted loss function (strengthening the strong roughness and large flight path angle scene error penalty) and the non-negative constraint training model, the non-flat EDR is finally predicted by "flat EDR x correction coefficient", which realizes the accurate correction of the roughness data in the non-flat state and improves the prediction accuracy in the key scene.

[0075] For the method steps disclosed in the above embodiments, the method steps are described as a series of action combinations for the purpose of simple description, but those skilled in the art should know that the embodiments of the present application are not limited by the order of the described actions, because according to the embodiments of the present application, certain steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present application.

[0076] As shown in Figure 2 , the present application also provides a roughness prediction correction system, comprising: A multi-dimensional data acquisition module 201 is configured to acquire multi-dimensional flight state data, wherein the multi-dimensional flight state data includes the roughness intensity value in the flat state, the vertical overload value, the vertical flight path angle, the aircraft type identification and the horizontal stabilizer distance in the non-flat state. A multi-dimensional data processing module 202 is configured to classify and preprocess the multi-dimensional flight state data, including abnormal detection and missing value recovery of the roughness intensity value and the vertical overload value, conversion of the aircraft type identification into a low-dimensional embedding vector, and distribution standardization processing of the vertical flight path angle and the horizontal stabilizer distance. A correction coefficient sequence generation module 203 is configured to construct a time window based on the preprocessed multi-dimensional flight state data, input a spatio-temporal feature extraction model for multi-dimensional feature fusion and time correlation analysis, and output a correction coefficient sequence corresponding to the time window. The jolt intensity prediction value sequence generation module 204 is configured to multiply the jolt intensity values at each time point in the time window with the corresponding correction coefficient sequence point by point to generate a jolt intensity prediction value sequence in the non-level flight state. The space-time feature extraction model optimization module 205 is configured to train the space-time feature extraction model based on a composite loss function including a key scene reinforcement mechanism and a non-negativity guarantee mechanism to optimize the prediction accuracy of the correction coefficient sequence.

[0077] It is worth noting that, although only some basic functional modules are disclosed in the embodiments of the present application, it does not mean that the composition of the system is limited to the above-mentioned basic functional modules. On the contrary, the meaning expressed in the embodiments is that on the basis of the above-mentioned basic functional modules, those skilled in the art can add one or more functional modules to form infinite embodiments or technical solutions in combination with the prior art. That is, the system is open rather than closed, and the protection scope of the present application claim cannot be limited to the disclosed basic functional modules because only individual basic functional modules are disclosed in the embodiments. At the same time, for the convenience of description, the above device is described as various units and modules. Of course, the functions of each unit and module can be realized in the same software and / or hardware when implementing the present application.

[0078] As shown in Figure 3 The present application also provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a jolt prediction correction method.

[0079] Figure 3 is a structural schematic diagram of an electronic device provided by the embodiments of the present application. As shown in the structure Figure 3 The electronic device provided in the embodiments of the present application includes one or more processors 710 and memories 720; the processor 710 in the electronic device can be one or more, Figure 3 The memory 720 is used for storing one or more programs; the one or more programs are executed by the one or more processors 710, so that the one or more processors 710 implement a jolt prediction correction method according to any one of the embodiments of the present application.

[0080] The electronic device can also include an input device 730 and an output device 740.

[0081] The processor 710, memory 720, input device 730 and output device 740 in the electronic device can be connected via a bus or other means. Figure 3 The bus connection is taken as an example.

[0082] The memory 720 in the electronic device serves as a computer-readable storage medium and can be used to store one or more programs. These programs can be software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the turbulence prediction and correction method provided in the embodiments of the present invention. The processor 710 executes the software programs, instructions, and modules stored in the memory 720 to execute various functional applications and data processing functions of the electronic device, thereby implementing the turbulence prediction and correction method provided in the aforementioned method embodiments.

[0083] The memory 720 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device, etc. In addition, the memory 720 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 720 may further include a memory remotely located relative to the processor 710, and these remote memories may be connected to the device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0084] The input device 730 may be used to receive input digital or character information and generate key signal input related to user settings and function control of the electronic device. The output device 740 may include a display device such as a display screen.

[0085] The present invention also provides a computer-readable storage medium storing a computer program executable by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of a turbulence prediction and correction method.

[0086] Specifically, the computer storage medium of the embodiment of the present invention can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in combination with an instruction execution system, device or device.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting and correcting turbulence, characterized in that: include: Acquiring multi-dimensional flight status data, wherein the multi-dimensional flight status data includes a turbulence intensity value in a level flight state, a vertical overload value in a non-level flight state, a vertical path angle, an aircraft model identifier, and a horizontal stabilizer distance; performing classification preprocessing on the multi-dimensional flight status data, including performing anomaly detection and missing value recovery on the turbulence intensity value and the vertical overload value, converting the aircraft model identifier into a low-dimensional embedding vector, and performing distribution normalization processing on the vertical path angle and the horizontal stabilizer distance; Constructing a time window based on the pre-processed multi-dimensional flight status data, inputting the spatiotemporal feature extraction model to perform multi-dimensional feature fusion and time correlation analysis, and outputting a correction coefficient sequence corresponding to the time window; Multiply the turbulence intensity value at each time point in the time window by the corresponding correction coefficient sequence at each time point to generate a turbulence intensity prediction value sequence in the non-level flight state; The spatiotemporal feature extraction model is trained based on a composite loss function including a key scene reinforcement mechanism and a non-negativity guarantee mechanism to optimize the prediction accuracy of the correction coefficient sequence.

2. The method for predicting and correcting turbulence according to claim 1, characterized in that: The multi-dimensional flight status data is subjected to classification preprocessing, including performing anomaly detection and missing value recovery on the turbulence intensity value and the vertical overload value, converting the aircraft model identifier into a low-dimensional embedding vector, and performing distribution normalization processing on the vertical path angle and the horizontal stabilizer distance, further comprising: Filter out bump intensity values ​​greater than a preset intensity range and vertical overload values ​​greater than a preset load range; Based on the time series continuity interpolation method, the missing turbulence intensity values ​​and vertical overload values ​​are restored through the data correlation of adjacent time points; Map the model identifier to a continuous vector space of fixed dimension to generate an embedding vector; A zero-mean-unit-variance linear transformation is performed on the vertical path angle and the horizontal stabilizer distance.

3. The method for predicting and correcting turbulence according to claim 1, characterized in that: Constructing a time window based on the pre-processed multi-dimensional flight status data further includes: Set the length of the continuous time window that covers the complete turbulence event life cycle; A sliding window mechanism with a preset overlap ratio is used to segment the time series, maintaining the temporal continuity of adjacent windows; splicing the pre-processed multi-dimensional flight status data along the feature dimension to generate a combined feature vector; Constructs a time window of three-dimensional tensor samples, with dimensions including time step dimension, combined feature dimension, and single channel dimension.

4. The method for predicting and correcting turbulence according to claim 1, characterized in that: Input the spatiotemporal feature extraction model to perform multi-dimensional feature fusion and time correlation analysis, and output the correction coefficient sequence corresponding to the time window, further including: The feature compression path involves extracting long-term trend features through multi-level time dimension downsampling operations, where each level contains convolution operations at different time scales, gradually expanding the temporal receptive field; Feature fusion layer, including aggregating the long-term trend features and eliminating information redundancy to form global features; The feature recovery path includes reconstructing the temporal resolution through upsampling operations in the temporal dimension, fusing global features with local transient features using a cross-level feature connection mechanism, and generating spatiotemporal features. The coefficient mapping layer converts the spatiotemporal features into a complete correction coefficient sequence in the time dimension.

5. The method for predicting and correcting turbulence according to claim 4, characterized in that: The feature compression path involves extracting long-term trend features through multi-level time dimension downsampling operations. Each level contains convolution operations at different time scales, gradually expanding the temporal receptive field. It further includes: Perform a sequence of temporal convolution operations with increasing dilation rates, gradually expanding the temporal receptive field by adjusting the convolution kernel sampling interval; Introducing nonlinear activation functions and batch normalization after each level of convolution operation to enhance feature expression capabilities; Inserting a time dimension compression operation after a specified convolutional layer reduces the temporal resolution while increasing the feature depth; The output is a composite feature tensor including the long-term aerodynamic impact of the aircraft model, the evolution trend of the vertical path angle, and the baseline level of the turbulence intensity value.

6. The method for predicting and correcting turbulence according to claim 1, characterized in that: Multiplying the turbulence intensity value at each time point in the time window by the corresponding correction coefficient sequence at each time point to generate a turbulence intensity prediction value sequence in a non-level flight state, further comprising: Get the turbulence intensity value at each time point in the current time window; Call the trained spatiotemporal feature extraction model and output the correction coefficient sequence corresponding to the time window; Perform a dot multiplication operation on the turbulence intensity value at the same time point and the corresponding correction coefficient to generate a turbulence intensity prediction value; Outputs a real-time updated sequence of turbulence intensity prediction values.

7. The method for predicting and correcting turbulence according to claim 1, characterized in that: The spatiotemporal feature extraction model is trained based on a composite loss function including a key scene reinforcement mechanism and a non-negativity guarantee mechanism to optimize the prediction accuracy of the correction coefficient sequence, further comprising: Scenario weighting mechanism, including dynamic calculation of error weight coefficients at each time point based on turbulence intensity and absolute value of vertical path angle, to enhance the error contribution of key scenarios; The error term includes the mean squared error between the predicted turbulence intensity value after the scenario weighting mechanism is implemented and the actual turbulence intensity value; Constraints, including penalty functions for negative values ​​in the correction coefficient series; The comprehensive optimization term includes combining the error term and the constraint term in a preset ratio to generate a composite loss function.

8. A turbulence prediction and correction system, characterized in that: include: a multi-dimensional data acquisition module configured to acquire multi-dimensional flight status data, wherein the multi-dimensional flight status data includes a turbulence intensity value in a level flight state, a vertical overload value in a non-level flight state, a vertical path angle, an aircraft type identifier, and a horizontal stabilizer distance; a multi-dimensional data processing module configured to perform classification pre-processing on the multi-dimensional flight status data, including performing anomaly detection and missing value recovery on the turbulence intensity value and the vertical overload value, converting the aircraft model identifier into a low-dimensional embedding vector, and performing distribution normalization processing on the vertical path angle and the horizontal stabilizer distance; a correction coefficient sequence generation module configured to construct a time window based on the preprocessed multi-dimensional flight status data, input the spatiotemporal feature extraction model to perform multi-dimensional feature fusion and time correlation analysis, and output a correction coefficient sequence corresponding to the time window; a turbulence intensity prediction value sequence generation module configured to multiply the turbulence intensity value at each time point within the time window by the corresponding correction coefficient sequence at each time point to generate a turbulence intensity prediction value sequence for a non-level flight state; The spatiotemporal feature extraction model optimization module is configured to train the spatiotemporal feature extraction model based on a composite loss function including a key scene reinforcement mechanism and a non-negativity guarantee mechanism to optimize the prediction accuracy of the correction coefficient sequence.

9. An electronic device, characterized in that: include: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that It stores a computer program that can be executed by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of the method according to any one of claims 1 to 7.

Images