Generalized generation type icing prediction method for establishing mapping relation between environment parameters and interface displacement based on deep learning
By mapping environmental parameters and interface displacement based on deep learning, the accuracy and real-time performance issues of traditional wing icing prediction are solved. It enables the prediction of icing pattern characteristics under arbitrary environmental variables, and has physical interpretability and generalization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2025-12-09
- Publication Date
- 2026-04-24
AI Technical Summary
Traditional methods for predicting wing icing are difficult to achieve accurate, real-time, and economical predictions, and are greatly affected by environmental factors, lacking real-time performance and accuracy.
A deep learning-based approach is used to establish a mapping relationship between environmental parameters and interface displacement. By obtaining the coordinate displacement vectors of two-dimensional ice-type interface points, data association and training are performed to construct a multi-neural network, thereby obtaining the mapping relationship between environmental factors and interface displacement vectors and realizing the prediction of ice-type characteristics.
It enables the prediction of icing-type features under arbitrary environmental variables, possesses physical interpretability and generalization ability, and improves the accuracy and real-time performance of prediction.
Smart Images
Figure CN121920184A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for predicting wing icing, and more particularly to a generalized generative icing prediction method based on deep learning to establish a mapping relationship between environmental parameters and interface displacement. Background Technology
[0002] Aircraft icing is a major factor affecting aviation safety, and in severe cases, it can lead to catastrophic consequences such as stall. Accurate detection and prediction of aircraft icing, especially airfoil icing characteristics, is crucial for ensuring the flight safety of both military and civilian aircraft. However, due to the specificity of the environment, complex dynamics, multiphase coupling, and the diversity of icing patterns, the physical mechanisms underlying aircraft icing are not yet fully understood, and it is likely that a complete understanding will remain elusive for a considerable period. This presents a significant bottleneck in the development of aircraft icing technology.
[0003] Traditional methods for predicting icing patterns obtain ice patterns from a physical perspective, including flight tests, icing wind tunnel tests, and numerical simulations. However, each of these methods has significant limitations and is closely related to environmental factors, making it difficult to obtain objective laws that are real-time, accurate, and economical. Summary of the Invention
[0004] This invention aims to provide a generalized generative icing prediction method based on deep learning to establish a mapping relationship between environmental parameters and interface displacement, in order to solve the problem that traditional wing icing prediction methods are difficult to accurately, in real time and economically predict aircraft icing patterns.
[0005] To achieve the above objectives, the specific plan is as follows:
[0006] A generalized generative icing prediction method based on deep learning to establish a mapping relationship between environmental parameters and interface displacement includes the following steps:
[0007] S1: Obtain the displacement vector of the two-dimensional ice-shaped interface point coordinates, convert both the initial airfoil image and the icing airfoil image into a set of point coordinates, sort them according to the outer contour features and calculate the difference;
[0008] S2: Associate the environmental factor dataset with the interface displacement vector point coordinate set one by one;
[0009] S3: Conduct deep learning training to obtain the environmental factors-interface displacement vector mapping relationship;
[0010] S4: Calculate the coordinate displacement vector of the two-dimensional icy interface point in the test set state using the trained weights and bias matrix.
[0011] S5, Input any environmental variable to obtain the predicted icing type: Predict the coordinate displacement of the interface point of the icing type by using the environmental parameters trained in step S3 - the interface displacement vector weight and the bias matrix, and then correct the prediction result in step S4, thereby realizing the icing type feature prediction analysis under any environmental variable.
[0012] Further, in step S1, the processing steps are as follows:
[0013] S11: Binarize the icing airfoil image, assign 0 to the wing and icing area, set the background to 1, and set an area threshold to remove small spot areas.
[0014] S12: Extract the icing airfoil boundary based on the binary image, generate a two-dimensional coordinate point set, obtain the icing airfoil point cloud coordinate set, and ensure that the point set density can describe the key geometric features of the ice shape.
[0015] S13: The initial airfoil is processed using S11 and S12 with the same set of point coordinates. The difference between the S11 and S12 is calculated and the difference is calculated with the point set of the icing airfoil to characterize the displacement of the two-dimensional ice interface point coordinates.
[0016] S14: Correspond one-to-one with the coordinate displacement vectors of the two-dimensional ice-shaped interface points to the environmental parameter dataset;
[0017] S15: Batch process all image data according to steps S11~S14 to ensure that the data information is converted from image to interface point coordinate displacement form.
[0018] Further, in step S2, the processing steps are as follows:
[0019] S21: Divide the acquired data into training set: test set ratios of 7:3;
[0020] S22, Constructing a Multiple Neural Network: The environmental parameter dataset is used as the input to the multiple neural network, and the coordinate set of interface displacement vector points is used as the output. Multiple neural networks are used to connect the points in between. The number of hidden layers and the number of neurons are determined according to the number of points.
[0021] Further, in step S3, the processing steps are as follows:
[0022] S31, Parameter initialization: Initialize the weight matrix of the neural network to a normal distribution and initialize the bias matrix to zero;
[0023] S32, Forward Propagation: The environmental parameter dataset is used as the input of a multi-neural network, and the coordinate set of interface displacement vector points is used as the output for forward propagation;
[0024] S33, Loss Calculation and Backpropagation: Using the mean squared error (MSE) as the loss function, backpropagation updates the weights and bias matrix until the error calculation requirement is met. The maximum number of rounds, target error, and algorithm can be set according to training requirements.
[0025] S34, Convergence criterion: Continue training until the accuracy of the computational error meets the requirements;
[0026] S35, Feature Extraction: Save the weight matrix ωi and bias matrix bi of each intermediate layer.
[0027] Further, in step S4, the processing steps are as follows:
[0028] S41, Network Construction: Construct a two-layer hidden fully connected network. The input dimension is the predicted coordinate displacement vector of the ice-shaped interface point in the test set, and the output dimension is the true value of the coordinate displacement vector of the interface point in the test set.
[0029] S42, Parameter Update: Calculate the mean squared error MSE2, and update the weights and biases by backpropagation based on the MSE2 result;
[0030] S43, Convergence criterion: Continue training until the accuracy meets the requirements.
[0031] In summary, the present invention has the following advantages over the prior art:
[0032] This invention filters out noise and interference information of the initial airfoil by subtracting the outer contour of the point set, while retaining the description of the ice shape physical growth process through the displacement vector of the point set. This makes the learning parameters have potential physical meaning and physical interpretability. Through training and learning with the database, an ice shape growth weight matrix and bias matrix are generated, which can be extended to any initial airfoil. The ice shape prediction method has generalization ability. Attached Figure Description
[0033] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:
[0034] Figure 1 A flowchart of the generalized generative icing prediction method based on deep learning to establish the mapping relationship between environmental parameters and interface displacement provided by the present invention.
[0035] Figure 2 The principle logic diagram of the generalized generative icing prediction method provided by the present invention is shown. Detailed Implementation
[0036] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0037] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form may also include the plural form unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0038] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.
[0039] This implementation method is based on the mapping logic of "environmental parameters - interface displacement vector". Two types of core data need to be prepared in advance: First, an environmental parameter dataset covering different flight scenarios (including incoming flow velocity v, flight altitude h, flight angle of attack α, ambient temperature T, liquid water content LWC, average diameter of supercooled water droplets MVD, and icing time t. Each parameter needs to be labeled with measurement accuracy and value range, such as temperature T ranging from -40℃ to 0℃ with an accuracy of ±0.5℃); Second, an airfoil image dataset for the corresponding scenario (it needs to include paired images of the initial airfoil and the airfoil after icing, with an image resolution of no less than 1024×1024 pixels and consistent shooting angles to avoid coordinate calculation errors caused by viewing angle deviations).
[0040] All data must undergo preprocessing and screening: samples with a missing environmental parameter rate exceeding 5% are removed, and invalid images with an icing area ratio of less than 0.1% are deleted.
[0041] See Figure 1 and Figure 2 As shown, this invention provides a generalized generative icing prediction method based on deep learning to establish a mapping relationship between environmental parameters and interface displacement, including the following steps:
[0042] S1: Obtain the displacement vector of the two-dimensional ice-shaped interface points, convert both the initial airfoil and the icing airfoil into point coordinate sets (X1,X2,…,XN, Y1,Y2,…,YN), sort them according to the outer contour features, and calculate the difference.
[0043] For example, when the initial data is an image, the processing steps are as follows:
[0044] S11, Binarization Processing: An adaptive threshold segmentation algorithm (such as the Otsu algorithm) is used to replace the fixed threshold assignment. The grayscale difference between the wing, icing area and background in the image is automatically identified. The wing and icing area are set to 0 (black) and the background is set to 1 (white) to avoid segmentation deviation caused by changes in lighting. Then, an area threshold is set to remove small spot areas.
[0045] S12, extract the icing airfoil boundary based on the binary image, generate a two-dimensional coordinate point set, obtain the icing airfoil point cloud coordinate set, and ensure that the point set density can describe the key geometric features of the ice shape;
[0046] S13: Repeat steps S11~S12 on the initial airfoil using the same set of point coordinates, and calculate the difference with the point set of the icing airfoil to characterize the displacement of the two-dimensional ice interface point coordinates.
[0047] S14: Correspond one-to-one with the coordinate displacement vector of the two-dimensional ice interface point to the environmental parameter dataset (inflow velocity v, flight altitude h, flight angle of attack α, ambient temperature T, liquid water content LWC, average diameter of supercooled water droplets MVD, and freezing time t);
[0048] S15: Batch process all image data according to steps S11~S14 to ensure that the data information is converted from image to interface point coordinate displacement form.
[0049] S2: Associate the environmental factor dataset with the interface displacement vector point coordinate set one by one;
[0050] S21, Divide the data into training and test sets: Divide the acquired data into training and test sets according to a certain ratio, generally 70% for training and 30% for test.
[0051] S22, Constructing a Multiple Neural Network: The environmental parameter dataset is used as the input of the multiple neural network, and the coordinate set of the interface displacement vector points is used as the output. Multiple neural networks are used to connect the points in the middle. The number of hidden layers and the number of neurons are determined according to the number of points.
[0052] S3: Conduct deep learning training to obtain the environmental factors-interface displacement vector mapping relationship, specifically including:
[0053] S31, Parameter initialization: Initialize the weight matrix of the neural network to a normal distribution and initialize the bias matrix to zero;
[0054] S32, Forward Propagation: The environmental parameter dataset is used as the input of a multi-neural network, and the coordinate set of interface displacement vector points is used as the output for forward propagation;
[0055] S33, Loss Calculation and Backpropagation: Using the mean squared error (MSE) as the loss function, backpropagation updates the weights and bias matrix until the error calculation requirement is met. The maximum number of rounds, target error, and algorithm can be set according to training requirements.
[0056] S34, Convergence criterion: Continue training until the accuracy of the computational error meets the requirements;
[0057] S35, Feature Extraction: Save the weight matrix ωi and bias matrix bi of each intermediate layer.
[0058] S4: Calculate the coordinate displacement vectors of the two-dimensional icy interface points in the test set state using the trained weights and bias matrices, specifically including:
[0059] S41, Network Construction: Construct a two-layer hidden fully connected network. The input dimension is the predicted coordinate displacement vector of the ice-shaped interface point in the test set, and the output dimension is the true value of the coordinate displacement vector of the interface point in the test set.
[0060] S42, Parameter Update: Calculate the mean squared error MSE2, and update the weights and biases by backpropagation based on the MSE2 result;
[0061] S43, Convergence criterion: Continue training until the accuracy meets the requirements;
[0062] S5: Input any environmental variable to obtain the predicted icing type: Predict the coordinate displacement of the interface point of the icing type by using the environmental parameters trained in step S3 - the interface displacement vector weight and the bias matrix, and then correct the prediction result by step S4, thereby realizing the icing type feature prediction analysis under any environmental variable.
[0063] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A generalized generative icing prediction method based on deep learning to establish a mapping relationship between environmental parameters and interface displacement, characterized in that, Including the following steps: S1: Obtain the displacement vector of the two-dimensional ice-shaped interface point coordinates, convert both the initial airfoil image and the icing airfoil image into a set of point coordinates, and sort them according to the outer contour features to calculate the difference. S2: Associate the environmental factor dataset with the interface displacement vector point coordinate set one by one; S3: Conduct deep learning training to obtain the environmental factors-interface displacement vector mapping relationship; S4: Calculate the coordinate displacement vector of the two-dimensional icy interface point in the test set state using the trained weights and bias matrix. S5. Input any environmental variable to obtain the predicted icing type: predict the coordinate displacement of the interface point of the icing type by using the environmental parameters trained in step S3 - the interface displacement vector weight and the bias matrix, and then correct the prediction result in step S4, thereby realizing the icing type feature prediction analysis under any environmental variable.
2. The generalized generative icing prediction method based on deep learning to establish a mapping relationship between environmental parameters and interface displacement, as described in claim 1, is characterized in that... In step S1, the processing steps are as follows: S11: Binarize the icing airfoil image, assign 0 to the wing and icing area, set the background to 1, and set an area threshold to remove small spot areas. S12: Extract the icing airfoil boundary based on the binary image, generate a two-dimensional coordinate point set, obtain the icing airfoil point cloud coordinate set, and ensure that the point set density can describe the key geometric features of the ice shape. S13: The initial airfoil is processed using S11 and S12 with the same set of point coordinates. The difference between the S11 and S12 is calculated and the difference is calculated with the point set of the icing airfoil to characterize the displacement of the two-dimensional ice interface point coordinates. S14: Correspond one-to-one with the coordinate displacement vectors of the two-dimensional ice-shaped interface points to the environmental parameter dataset; S15: Batch process all image data according to steps S11~S14 to ensure that the data information is converted from image to interface point coordinate displacement form.
3. The generalized generative icing prediction method based on deep learning to establish a mapping relationship between environmental parameters and interface displacement, as described in claim 1, is characterized in that... In step S2, the processing steps are as follows: S21: Divide the acquired data into training set: test set ratios of 7:3; S22, Constructing a Multiple Neural Network: The environmental parameter dataset is used as the input to the multiple neural network, and the coordinate set of interface displacement vector points is used as the output. Multiple neural networks are used to connect the points in between. The number of hidden layers and the number of neurons are determined according to the number of points.
4. The generalized generative icing prediction method based on deep learning to establish a mapping relationship between environmental parameters and interface displacement, as described in claim 1, is characterized in that... In step S3, the processing steps are as follows: S31, Parameter initialization: Initialize the weight matrix of the neural network to a normal distribution and initialize the bias matrix to zero; S32, Forward Propagation: The environmental parameter dataset is used as the input of a multi-neural network, and the coordinate set of interface displacement vector points is used as the output for forward propagation; S33, Loss Calculation and Backpropagation: Using the mean squared error (MSE) as the loss function, backpropagation updates the weights and bias matrices until the error calculation requirement is met. The maximum number of rounds, target error, and algorithm are set according to the training requirements. S34, Convergence criterion: Continue training until the accuracy of the computational error meets the requirements; S35, Feature Extraction: Save the weight matrix ωi and bias matrix bi of each intermediate layer.
5. The generalized generative icing prediction method based on deep learning to establish a mapping relationship between environmental parameters and interface displacement, as described in claim 1, is characterized in that... In step S4, the processing steps are as follows: S41, Network Construction: Construct a two-layer hidden fully connected network. The input dimension is the predicted coordinate displacement vector of the ice-shaped interface point in the test set, and the output dimension is the true value of the coordinate displacement vector of the interface point in the test set. S42, Parameter Update: Calculate the mean squared error MSE2, and update the weights and biases by backpropagation based on the MSE2 result; S43, Convergence criterion: Continue training until the accuracy meets the requirements.