Flight parameter prediction and confidence evaluation method
By integrating dual-output Bayesian neural networks and isomorphic dual-head network models, the cognition and random uncertainties are quantified, solving the problem of lack of comprehensive analysis in existing technologies. This enables reliable confidence evaluation of flight parameter predictions and improves the safety of navigation systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SICHUAN UNIV
- Filing Date
- 2023-04-23
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies only analyze cognitive uncertainty and accidental uncertainty in isolation, lacking a comprehensive analysis of flight parameter prediction results, resulting in insufficient safety of navigation systems under abnormal or malfunctioning conditions.
An integrated approach combining a dual-output Bayesian neural network and a homogeneous dual-head network model is adopted. By using sliding window sampling and a two-stage training algorithm, cognitive and random uncertainties are quantified to establish a flight parameter prediction and confidence evaluation model. The confidence level is calculated by combining the weights of cognitive and random uncertainties.
It enables reliable confidence evaluation of flight parameter prediction results, improves the safety performance of the navigation system, and allows it to operate stably in extreme environments.
Smart Images
Figure CN116468174B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of flight parameter prediction technology, specifically to methods for flight parameter prediction and confidence evaluation. Background Technology
[0002] Modern aircraft, such as manned fighter jets and drones, often operate in extreme environments including low pressure, extreme cold, and low visibility. Their navigation systems are complex systems comprised of interconnected subsystems, responsible for real-time monitoring of the aircraft's flight status and providing decision-making support for pilots and ground control personnel. They are a crucial foundation for the safe and stable operation of aircraft. When the navigation system is in an abnormal or malfunctioning state, it will send abnormal navigation data to the flight control system—a phenomenon known as navigation deception—which can severely impact aircraft safety performance.
[0003] Deep learning technology has been widely applied to various tasks such as image recognition, machine translation, and reinforcement learning, achieving great success. However, deep neural networks cannot quantify the uncertainty of their results, and the predictions they provide are not entirely reliable. In critical fields with high safety requirements, such as aviation and aerospace, relying entirely on the predictions of deep learning models could lead to catastrophic consequences. Uncertainty quantification is key to enabling deep neural networks to understand unknown information. In navigation fraud detection tasks, quantifying the widespread multi-source uncertainties and evaluating the confidence level of flight parameter data prediction models can effectively improve the overall safety performance of navigation systems and ensure the safe and stable operation of aircraft.
[0004] Two types of uncertainty are prevalent in deep learning: cognitive uncertainty and random uncertainty. Cognitive uncertainty arises from insufficient historical data and incomplete training; however, it can be effectively reduced with sufficient datasets. Random uncertainty, on the other hand, is caused by noise in the collected data due to environmental or sensor inherent characteristics, leading to randomness in prediction and anomaly localization results. Random uncertainty is unavoidable in practical engineering applications.
[0005] Currently, methods for quantifying cognitive uncertainty in deep learning can be divided into two categories: Bayesian network-based methods and non-Bayesian network-based methods. Bayesian network-based methods quantify prediction uncertainty by pre-specifying the prior distribution of neural network parameters and calculating the posterior distribution of those parameters given training data. These methods use weight and bias distributions instead of the fixed weights and biases found in traditional neural networks, effectively capturing the impact of subtle changes in model parameters on prediction results, thereby calculating the confidence level of the prediction and improving the model's robustness. Commonly used Bayesian network-based uncertainty quantification methods include Monte Carlo, Laplace approximation, and variational autoencoders. However, Bayesian network-based methods typically require calculating an accurate model posterior, which is difficult to compute in practical applications. Therefore, variational methods are often used to approximate the model's posterior distribution. The other category of methods for quantifying cognitive uncertainty is non-Bayesian methods, with representative examples being measurement algorithms based on model ensembles and evidence theory. Model ensemble-based methods reduce variance by synthesizing multiple predictions that are prone to error individually, and generate distribution estimates of model uncertainty. However, ensembles require higher memory and computational costs, making model deployment difficult in many situations, and this approach may obfuscate the sources of uncertainty. Currently, research on the random uncertainty inherent in the distribution of quantified data is relatively limited. Random uncertainty arises when test data classes overlap or when noise is present in the data. Compared to cognitive uncertainty, although the formula for random uncertainty is naturally measured through maximum likelihood training, most works still do not consider the inherent uncertainty of the data. However, the noise, inconsistency, and multimodal nature of real-world data make random uncertainty non-negligible.
[0006] In summary, current research on uncertainty quantification based on neural network prediction models only analyzes cognitive uncertainty and accidental uncertainty separately, lacking a comprehensive analysis of the impact of both uncertainties on the prediction results. Summary of the Invention
[0007] To address the technical problem of existing technologies that only analyze cognitive uncertainty and random uncertainty in isolation, lacking a comprehensive analysis of the impact of both uncertainties on prediction results, this application provides a method for predicting and evaluating flight parameters that can both predict the values of missing flight parameters and derive confidence levels from the perspectives of both cognitive and random uncertainties. Specifically:
[0008] The method for predicting flight parameters and evaluating confidence levels includes the following steps:
[0009] S1 collects multiple sets of flight data after the aircraft has lost the same flight parameter within a certain period of time. Each set of flight data includes N-1 flight parameters.
[0010] S2 performs normalization processing and sliding window sampling on the collected flight data. After sliding windowing, the model input samples are represented as follows: , This represents the values of N-1 flight parameters at time t, and D represents the window width;
[0011] S3, Input the model obtained in step S2 into the sample. The input is fed into a confidence evaluation network model to obtain a sequence of predicted results for M missing flight parameters. and M random uncertainty sequences ,in Let represent the value of the m-th prediction among M predictions at time t. Let m represent the value of the m-th random uncertainty among M random uncertainties at time t, where 1 ≤ m ≤ M, and m ∈ [1, M].
[0012] S4, based on the prediction result sequence obtained in step S3 and random uncertainty sequences The final prediction results of missing flight parameters were obtained. Quantification of cognitive uncertainty and the quantification of random uncertainty ;
[0013] S5, based on the quantification value of cognitive uncertainty and the quantification of random uncertainty The confidence level of the final prediction result for the missing parameters is calculated. .
[0014] Furthermore, in step S4, the predicted result sequence is taken. The mean of the missing parameters is used as the final prediction result. .
[0015] Furthermore, in step S4, the predicted result sequence is taken. The variance is used as a quantification of cognitive uncertainty. .
[0016] Furthermore, in step S4, a random uncertainty sequence is taken. The mean is used as a quantification value of random uncertainty. .
[0017] Furthermore, in step S5, the confidence level The calculation formula is: ,in, Weights for random uncertainty To understand the weight of uncertainty, This is the error measurement coefficient.
[0018] Furthermore, the steps for establishing the confidence evaluation network model in step S3 are as follows:
[0019] S31, build a dual-head network;
[0020] S311, Establish a comprehensive feature extraction network This network is used to extract comprehensive high-level abstract features and perform feature dimensionality reduction from the input data. The input to this network is the model input sample at time t. The output is the comprehensive feature at time t. , The weights for the comprehensive feature extraction network;
[0021] S312, Establish the prediction result output network and random uncertain networks ;
[0022] Prediction results output network The input is the comprehensive feature at time t. The output is the predicted value of the missing parameter at time t. , Output the network weights to provide the prediction results;
[0023] Random Uncertain Network To capture the random uncertainty in the data, the network input is the comprehensive features at time t. The output is the random uncertainty at time t. , The weights are for random, uncertain networks;
[0024] S313, connect the three networks established in steps S311 and S312 to form a homogeneous dual-head network;
[0025] S32, repeat step S31 to obtain M isomorphic bi-headed networks, and connect the M isomorphic bi-headed networks to obtain the confidence evaluation network model.
[0026] Furthermore, the network model is trained according to the established confidence level using the following steps:
[0027] S100, data acquisition, collects flight data from multiple normal flights of the aircraft, with each flight data containing N flight parameters;
[0028] S200, select one of N flight parameters as the parameter to be predicted, and denote the true value of the selected parameter to be predicted as... ;
[0029] S201, normalizes the flight parameters (excluding the predicted parameters) and performs sliding window sampling to obtain the model input samples. , This represents the values of N-1 flight parameters at time t, and D represents the window width;
[0030] S202, Training the Comprehensive Feature Extraction Network With the prediction result output network :
[0031] Input the model obtained in step S201 into the sample. The input is fed into the confidence evaluation network model constructed according to claim 1, and the result is obtained through the mean squared error loss function. Output the predicted sequence of flight parameters Where b is the number of samples in a single batch, and the mean of the predicted result sequence is taken as the final predicted result. The variance of the predicted result sequence is taken as the final cognitive uncertainty; thus, the final prediction result of the flight parameter data prediction model with multi-source uncertainty is... Approximating the true value of the parameter to be predicted Determine the weights of the comprehensive feature extraction network for the parameters to be predicted. The weights of the network that produce the prediction results ;
[0032] S203, Training a network with random uncertainty :
[0033] Freeze-through-feature extraction network With the prediction result output network Pre-trained parameter weights and Input the model obtained in step S201 into the sample. The input is fed into the confidence evaluation network model constructed according to claim 1, and then processed through the random uncertainty loss function. Output the predicted values of the parameters to be predicted. and random uncertainty sequences Where b is the number of samples in a single batch, and the mean of the random uncertainty sequence is taken as the final random uncertainty. This leads to the random uncertainty of flight parameter data prediction models with multi-source uncertainty. It can fully describe the difference between predicted and actual values and determine the weights of the random uncertainty network for the flight parameters to be predicted.
[0034] S300, sequentially select other flight parameters from the N flight parameters as parameters to be predicted, and repeat steps S201 to S203 to determine the weights of the comprehensive feature extraction network for each flight parameter. The prediction results are output as the network weights. Weights of random and uncertain networks .
[0035] Furthermore, before performing steps S2 and S201, the collected flight parameter data is first cleaned.
[0036] The beneficial effects of this invention are:
[0037] 1. This invention comprehensively considers the impact of cognitive uncertainty and random uncertainty on the prediction model, quantifies the cognitive uncertainty and random uncertainty in the model respectively, and combines the method of integrating dual-output Bayesian neural network and isomorphic dual-head network model to realize flight parameter prediction and confidence evaluation of the predicted value. The final confidence index is also more reliable.
[0038] 2. This invention uses a homogeneous dual-head network model integration method to quantify cognitive uncertainty. Compared with other Bayesian network-based methods, the homogeneous dual-head network can obtain the distribution of prediction data. At the same time, this method can perform parallel computation, which can accelerate the computational efficiency.
[0039] 3. This invention establishes a dual-output neural network, designs a random uncertainty loss function, captures noise at the data level, trains cognitive uncertainty and random uncertainty separately, improves the transferability and scalability of the uncertainty quantification model, and constructs a simple and universal uncertainty quantification method. Attached Figure Description
[0040] Figure 1 This is a flowchart of the flight parameter prediction and confidence evaluation method of this application;
[0041] Figure 2 This is a schematic diagram of the sliding window processing in this application;
[0042] Figure 3 This is a schematic diagram of the isomorphic dual-head network structure in this application;
[0043] Figure 4 This is a schematic diagram of the confidence evaluation network model structure in this application;
[0044] Figure 5 This embodiment of the application shows the final prediction result and confidence level of the pitch angle after injecting 50dB of noise data;
[0045] Figure 6 This application embodiment shows the final pitch angle prediction result and confidence level obtained after injecting 20dB of noise data;
[0046] Figure 7 This embodiment of the application obtains the final prediction result and confidence level of the pitch angle after injecting 10dB of noise data. Detailed Implementation
[0047] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings, so as to provide a better understanding of the concept of the present invention, the technical problem solved, the technical features constituting the technical solution, and the technical effects brought about. However, it should be noted that the description of these embodiments is illustrative and does not constitute a specific limitation of the present invention.
[0048] Flight parameters are parameters that represent the aircraft's state during flight. These include axial overload, lateral overload, normal overload, pitch rate, roll rate, yaw rate, angle of attack, sideslip angle, flight path angle, path inclination angle, true heading angle, pitch angle, roll angle, free speed, and ground speed. Flight parameters provide a basis for decision-making by pilots or ground controllers. When the aircraft's navigation system malfunctions or fails, the flight system will send abnormal flight data to the flight control system, which will seriously affect flight safety.
[0049] With the development of science and technology, deep learning technology has been widely used for flight parameter prediction. However, the predictions made by deep learning networks are not entirely reliable, and relying solely on their predictions could lead to catastrophic consequences. Uncertainty quantification is key to enabling deep learning neural networks to understand unknown information. In navigation fraud detection tasks, quantifying the widespread multi-source uncertainties and evaluating the confidence level of flight parameter data prediction models can effectively improve the overall safety performance of navigation systems and ensure the safe and stable operation of aircraft. Two types of uncertainty are prevalent in deep learning: cognitive uncertainty and random uncertainty. However, current research on uncertainty quantification based on neural network prediction models only analyzes cognitive uncertainty and random uncertainty separately, lacking a comprehensive analysis of the impact of both uncertainties on the prediction model.
[0050] Based on the above, it is necessary to provide a method for predicting flight parameters and evaluating confidence levels that simultaneously considers cognitive uncertainty and random uncertainty, specifically including the following steps:
[0051] First, S1 collects multiple sets of flight data after the aircraft has lost the same flight parameter within a certain period of time. Each set of flight data includes N-1 flight parameters.
[0052] Then, in S2, the collected flight data is normalized and a sliding window sampling operation is performed. After the sliding window is applied, the model input sample is represented as follows: , This represents the values of N-1 flight parameters at time t, and D represents the window width. The sliding window processing is as follows: Figure 2As shown. Those skilled in the art should understand that this method utilizes historical flight data after multiple sets of data lacking the same flight parameter to predict the value of the missing flight parameter at the next moment, and to determine the final prediction result for the missing flight parameter. Make a confidence assessment. The model input samples for this method... It is a two-dimensional data set of (N-1)×D.
[0053] In this application, "missing" includes missing flight parameters due to abnormal circumstances, as well as two situations: parameter abnormalities and the provision of redundant parameter information.
[0054] After completing step S2, step S3 is required, where the model obtained in step S2 is input into the samples. The input is fed into a confidence evaluation network model to obtain a sequence of predicted results for M missing flight parameters. and M random uncertainty sequences ,in Let represent the value of the m-th prediction among M predictions at time t. Let m represent the value of the m-th random uncertainty among M random uncertainties at time t, where 1 ≤ m ≤ M, and m ∈ [1, M].
[0055] S4, based on the prediction result sequence obtained in step S3 and random uncertainty sequences The final prediction results of missing flight parameters were obtained. Quantification of cognitive uncertainty and the quantification of random uncertainty .
[0056] S5, based on the quantification value of cognitive uncertainty and the quantification of random uncertainty The final prediction result of the missing parameters is calculated. confidence level .
[0057] In step S3, multiple prediction results with missing flight parameters and multiple random uncertainties are obtained. In this application, the prediction result sequence is taken. The mean of the missing parameters is used as the final prediction result. Take the prediction result sequence The variance is used as a quantification of cognitive uncertainty. Take the random uncertainty sequence The mean is used as a quantification value of random uncertainty. .
[0058] In step S5, confidence level The calculation formula is: ,in, Weights for random uncertainty To understand the weight of uncertainty, This is the error measurement coefficient.
[0059] Confidence The calculation formula involves and Five parameters, among which and The calculation was performed according to the steps described above. Weights for random uncertainty To understand the weight of uncertainty, These three values are error measurement coefficients, and the user shall make the appropriate selection based on the actual situation. No limitation is made in this application.
[0060] The confidence level of the final prediction result of the missing parameters calculated by the method provided in this application is given. The value ranges from (0, 100). The closer the value is to 0, the lower the confidence level, indicating a lower reliability of the final prediction result for the missing parameter; the closer the value is to 100, the higher the confidence level, indicating a higher reliability of the final prediction result for the missing parameter. If calculated... If the value is less than 0, it is forcibly reset to 0.
[0061] As shown above, this method comprehensively considers the impact of cognitive uncertainty and random uncertainty on the prediction results, quantifies the cognitive uncertainty and random uncertainty in the model respectively, and finally obtains the confidence level. It is more accurate and objective.
[0062] In step S2, the confidence evaluation network model will output the prediction results for the M missing flight parameters. and M random uncertainties The structure of the confidence evaluation network model is as follows: Figure 3 As shown, Figure 3 The confidence evaluation network model shown is structured by multiple such... Figure 2 The network consists of a homogeneous dual-headed network as shown.
[0063] First, such as Figure 2As shown, a homogeneous dual-head network includes a comprehensive feature extraction network, a prediction result output network, and a random uncertainty network. Data is input to the comprehensive feature extraction network, which outputs its result to both the prediction result output network and the random uncertainty output network. The prediction result output network outputs the prediction result, and the random uncertainty network outputs the random uncertainty. Those skilled in the art should understand that a homogeneous dual-head network ultimately outputs only one prediction result and one random uncertainty. Since step S3 of this application requires the output of M prediction results and random uncertainties, M homogeneous dual-head networks should be constructed.
[0064] Specifically, the confidence evaluation model in step S3 can be established according to the following steps:
[0065] S31, build a homogeneous dual-head network.
[0066] S311, Establish a comprehensive feature extraction network This network is used to extract comprehensive high-level abstract features and perform feature dimensionality reduction from the input data. The input to this network is the model input sample at time t. The output is the comprehensive feature at time t. , The weights are used for the comprehensive feature extraction network.
[0067] S312, Establish the prediction result output network and random uncertain networks ;
[0068] Prediction results output network The input is the comprehensive feature at time t. The output is the predicted value of the missing parameter at time t. , Output the network weights to provide the prediction results;
[0069] Random Uncertain Network To capture the random uncertainty in the data, the network input is the comprehensive features at time t. The output is the random uncertainty at time t. , The weights are for random, uncertain networks.
[0070] S313, connect the three networks established in steps S211 and S212 to form a homogeneous dual-head network.
[0071] S32, repeat step S21 to obtain M isomorphic dual-head networks, and connect the M isomorphic dual-head networks to obtain the confidence evaluation network model.
[0072] Those skilled in the art should understand that the above This represents the comprehensive feature extraction network of the m-th isomorphic dual-head network among M isomorphic dual-head networks; This represents the prediction output network of the m-th isomorphic two-head network among M isomorphic two-head networks; This represents the random uncertain network of the m-th isomorphic bi-headed network among M isomorphic bi-headed networks.
[0073] Once the confidence evaluation model is established, it needs to be trained. The purpose of training is to determine the comprehensive feature extraction network in each isomorphic dual-head network. Weights of the comprehensive feature extraction network Prediction result output network The prediction results output the network weights. Random uncertainty network Weights of random uncertainty networks Specifically, the confidence evaluation model established according to the above steps is trained using the following steps:
[0074] S100 collects data from multiple normal flights of the aircraft, with each flight data set containing N flight parameters.
[0075] It is easy to understand that the training process should use flight parameters under normal flight conditions rather than abnormal conditions in order to achieve the training objective.
[0076] S200, select one of N flight parameters as the parameter to be predicted, and denote the true value of the selected parameter to be predicted as... .
[0077] S201, normalizes the flight parameters (excluding the predicted parameters) and performs sliding window sampling to obtain the model input samples. , This represents the values of N-1 flight parameters at time t, and D represents the window width.
[0078] This application employs a two-stage training algorithm. First, in step S202, the comprehensive feature extraction network is trained. With the prediction result output network :
[0079] Input the model obtained in step S201 into the sample. The input is fed into the confidence evaluation network model constructed as described above, and the result is obtained through the mean squared error loss function. Output the predicted sequence of flight parameters Where b is the number of samples in a single batch, and the mean of the predicted result sequence is taken as the final predicted result. The variance of the predicted result sequence is taken as the final cognitive uncertainty; thus, the predicted values of the flight parameter data prediction model with multi-source uncertainty are... Approximating the true value of the parameter to be predicted Determine the weights of the comprehensive feature extraction network for the parameters to be predicted. The weights of the network that produce the prediction results ;
[0080] S203, Training a network with random uncertainty :
[0081] Freeze-through-feature extraction network With the prediction result output network Pre-trained parameter weights and Input the model obtained in step S201 into the sample. The input is fed into the confidence evaluation network model constructed as described above, and then processed through the random uncertainty loss function. Output the predicted values of the parameters to be predicted. and random uncertainty sequences Where b is the number of samples in a single batch, and the mean of the random uncertainty sequence is taken as the final random uncertainty. This leads to the random uncertainty of flight parameter data prediction models with multi-source uncertainty. It can fully describe the difference between predicted and actual values and determine the weights of the random uncertainty network for the flight parameters to be predicted. .
[0082] S300, sequentially select other flight parameters from the N flight parameters as parameters to be predicted, and repeat steps S201 to S203 to determine the weights of the comprehensive feature extraction network for each flight parameter. The prediction results are output as the network weights. Weights of random and uncertain networks .
[0083] It should be understood that when training the confidence evaluation model, different initial weights should be set for the M isomorphic dual-head networks. The training process is random, and the parameters learned by the M isomorphic models are different, resulting in different weights.
[0084] The traditional heteroscedasticity function design loss function is as follows: Its loss function exists and The two adversarial terms make it difficult for the model to converge during training.
[0085] This application employs a two-stage training algorithm, namely, first training the comprehensive feature extraction network. With the prediction result output network The loss function used in training is Then train the random uncertainty network. In training a network with random uncertainty Freezing the integrated feature extraction network during the process With the prediction result output network With the trained weights, the loss function used for training is: .
[0086] Compared to traditional methods using heteroscedasticity functions, the two-stage training algorithm does not have... and With two adversarial terms, the confidence evaluation model is more likely to converge during training. Furthermore, the accidental uncertainty loss function in this application... Designed with observation noise in mind, it improves the transferability and scalability of uncertainty quantification models. In the accidental uncertainty loss function... Derived from a cognitive uncertainty network, this value is fixed for fixed input data.
[0087] Preferably, the collected flight parameter data is cleaned before performing steps S2 and S201. Cleaning specifically includes operations such as handling missing values, handling duplicate values, and data resampling.
[0088] After a neural network model is built, it undergoes a training and testing process. To further illustrate the steps of training the confidence evaluation network model in this application, and also to demonstrate the effectiveness of this application, the following embodiments are provided:
[0089] This embodiment first collects sensor data during UAV flight missions, obtaining 12 flight data sets under normal conditions as 12 training sets. Flight parameter data from four different aircraft models are used as seven test sets. Each flight data set contains 32 flight parameters, specifically: axial overload, lateral overload, normal overload, pitch rate, roll rate, yaw rate, angle of attack, sideslip angle, flight path angle, path inclination angle, true heading angle, pitch angle, roll angle, vacuum speed, ground speed, static pressure, dynamic pressure, yaw speed, landing gear, aircraft center of gravity, aircraft weight, left canard control surface position, left ETU angle, left inboard aileron control surface position, left leading flap control surface inboard position, left outboard aileron control surface position, left rudder control surface position, right canard control surface position, right ETU angle, right inboard aileron control surface position, right leading flap control surface inboard position, right outboard aileron control surface position, and right rudder control surface position.
[0090] Then, normalization and sliding window processing were performed on the training and test sets containing 33 flight parameters. First, the maximum value of each flight parameter was used... and minimum value , the original data The data is scaled proportionally to the [0,1] interval to obtain the normalized flight data. Then, the training and test sets are divided into samples using a combination of sliding window and slicing, with a sliding step size of 100. Historical slice data is used to predict the navigation parameters to be detected at the current moment, completing the sliding window sampling operation. A single sample after sliding windowing is... Where t is the timestamp of the flight data and D is the window width. The data input to the model at time t The target output value of the model is the predicted value of the parameter to be detected at time t, such as... Figure 4 As shown.
[0091] Five isomorphic head-binding networks are constructed, and then these five isomorphic head-binding networks are connected to form a confidence evaluation network model, such as... Figure 3 As shown.
[0092] Next, the constructed confidence evaluation network model is trained using a two-stage training algorithm. The first stage trains the comprehensive feature extraction network. With the prediction result output network The processed flight parameter data training set As input, through the mean squared error loss function Output 5 prediction results and 5 cognitive uncertainties, and take the mean of the prediction result sequence as the final prediction result. The variance of the predicted result sequence is taken as the final cognitive uncertainty; thus, the predicted values of the flight parameter data prediction model with multi-source uncertainty are... Approximating the true value of the parameter to be predicted Determine the weights of the comprehensive feature extraction network for the parameters to be predicted. The weights of the network that produce the prediction results .
[0093] The second phase of training involves random uncertainty in the output network. During training, the parameter weights of the integrated feature extraction network and the prediction output network are frozen, and a random uncertainty loss function is designed considering observation noise. Given the same training set, the system outputs five prediction results and five random uncertainties. The variance of the prediction result sequence is taken as the final cognitive uncertainty; the mean of the random uncertainty sequence is taken as the final random uncertainty. This leads to the random uncertainty of flight parameter data prediction models with multi-source uncertainty. It can fully describe the difference between predicted and actual values and determine the weights of the random uncertainty network for the flight parameters to be predicted. .
[0094] Next, process the test set data. Implanting different levels of noise The data after noise implantation is + Specifically, taking pitch angle as an example, noise data of 50dB, 20dB, and 10dB are injected into 32 flight parameters other than pitch angle, respectively. The noise-injected data is then input into the model to obtain the prediction results and prediction confidence of pitch angle under different noise levels, such as... Figure 5 , Figure 6 , Figure 7 As shown. From Figure 5 , Figure 6 , Figure 7 It can be observed that as the signal-to-noise ratio decreases, the predicted data gradually deviates from the actual data. At the same time, the confidence index also decreases as the signal-to-noise ratio decreases, indicating that the confidence evaluation index proposed in this paper can effectively represent the accuracy of the prediction results.
[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method of flight parameter prediction and confidence assessment, characterized in that, Includes the following steps: S1 collects multiple sets of flight data after the aircraft has lost the same flight parameter within a certain period of time. Each set of flight data includes N-1 flight parameters. S2, normalizing the collected flight data and performing a sliding window sampling operation, and the model input sample after the sliding window is represented as , represents the value of N-1 flight parameters at time t, and D represents the window width. S3, Input the model obtained in step S2 into the sample. The input is fed into a confidence evaluation network model to obtain a sequence of predicted results for M missing flight parameters. and M random uncertainty sequences ,in Let represent the value of the m-th prediction among M predictions at time t. Let m represent the value of the m-th random uncertainty among M random uncertainties at time t, where 1 ≤ m ≤ M, and m ∈ [1, M]. S4, based on the prediction result sequence obtained in step S3 and random uncertainty sequences The final prediction results of missing flight parameters were obtained. Quantification of cognitive uncertainty and the quantification of random uncertainty ; S5, based on the quantification value of cognitive uncertainty and the quantification of random uncertainty The confidence level of the final prediction result for the missing parameters is calculated. ; In step S4, the prediction result sequence is obtained. The mean of the missing parameters is used as the final prediction result. Take the prediction result sequence The variance is used as a quantification of cognitive uncertainty. Take a random, uncertain sequence The mean is used as a quantification of random uncertainty. ; In step S5, confidence level The calculation formula is: ,in, Weights for random uncertainty To understand the weight of uncertainty, This is the error measurement coefficient; The steps for establishing the confidence evaluation network model in step S3 are as follows: S31, build a dual-head network; S311, Establish a comprehensive feature extraction network This network is used to extract comprehensive high-level abstract features and perform feature dimensionality reduction from the input data. The input to this network is the model input sample at time t. The output is the comprehensive feature at time t. , The weights for the comprehensive feature extraction network; S312, Establish the prediction result output network and random uncertain networks ; Prediction results output network The input is the comprehensive feature at time t. The output is the predicted value of the missing parameter at time t. , Output the network weights to provide the prediction results; Random Uncertain Network To capture the random uncertainty in the data, the network input is the comprehensive features at time t. The output is the random uncertainty at time t. , The weights are for random, uncertain networks; S313, connect the three networks established in steps S311 and S312 to form a homogeneous dual-head network; S32, repeat step S31 to obtain M isomorphic bi-headed networks, and connect the M isomorphic bi-headed networks to obtain the confidence evaluation network model.
2. The method for predicting flight parameters and evaluating confidence level according to claim 1, characterized in that, The confidence evaluation network model established according to claim 1 is trained using the following steps: S100, data acquisition, collects flight data from multiple normal flights of the aircraft, with each flight data containing N flight parameters; S200, select one of N flight parameters as the parameter to be predicted, and denote the true value of the selected parameter to be predicted as... ; S201, normalizes the flight parameters (excluding the predicted parameters) and performs sliding window sampling to obtain the model input samples. , This represents the values of N-1 flight parameters at time t, and D represents the window width; S202, Training the Comprehensive Feature Extraction Network With the prediction result output network : Input the model obtained in step S201 into the sample. The input is fed into the confidence evaluation network model constructed according to claim 1, and the mean squared error loss function is applied. Output the predicted sequence of flight parameters Where b is the number of samples in a single batch, and the mean of the predicted result sequence is taken as the final predicted result. The variance of the predicted result sequence is taken as the final cognitive uncertainty; thus, the final prediction result of the flight parameter data prediction model with multi-source uncertainty is... Approximating the true value of the parameter to be predicted Determine the weights of the comprehensive feature extraction network for the parameters to be predicted. The weights of the network that produce the prediction results ; S203, Training a network with random uncertainty : Freeze-through-feature extraction network With the prediction result output network Pre-trained parameter weights and Input the model obtained in step S201 into the sample. The input is fed into the confidence evaluation network model constructed according to claim 1, and then processed through the random uncertainty loss function. Output the predicted values of the parameters to be predicted. and random uncertainty sequences Where b is the number of samples in a single batch, and the mean of the random uncertainty sequence is taken as the final random uncertainty. This leads to the random uncertainty of flight parameter data prediction models with multi-source uncertainty. It can fully describe the difference between predicted and actual values and determine the weights of the random uncertainty network for the flight parameters to be predicted. ; S300, sequentially select other flight parameters from the N flight parameters as parameters to be predicted, and repeat steps S201 to S203 to determine the weights of the comprehensive feature extraction network for each flight parameter. The prediction results are output as the network weights. Weights of random and uncertain networks .
3. The method for predicting flight parameters and evaluating confidence level according to claim 1 or 2, characterized in that: Before proceeding to steps S2 and S201, the collected flight parameter data is first cleaned.