A Deep Learning-Based Aircraft Attitude Control Method

Abstract

This invention discloses a deep learning-based aircraft attitude control method, relating to the field of video surveillance technology. The method includes: constructing an aircraft attitude control network based on a recurrent neural network, with the input layer receiving various aircraft state data and real-time monitoring sensor data. This invention deeply mines and learns the temporal information of the aircraft's flight state through a recurrent neural network, accurately predicting the trend of aircraft attitude changes. When facing complex and ever-changing flight conditions, such as strong airflow, bird strikes, or equipment malfunctions, the real-time monitoring sensors can quickly capture abnormal information. The adaptive adjustment module optimizes network parameters or structure in a timely manner according to a preset strategy, effectively improving the response speed, accuracy, and stability of aircraft attitude control under sudden conditions. This ensures that the aircraft can maintain a safe and stable flight attitude under various extreme circumstances, greatly enhancing the safety and reliability of aircraft flight and reducing the risk of flight accidents caused by sudden situations.

Description

Technical Field

[0001] This invention relates to the field of aircraft attitude control technology, specifically to an aircraft attitude control method based on deep learning. Background Technology

[0002] An airplane is a heavier-than-air aircraft with wings and one or more engines, capable of flying in the atmosphere under its own power. Aircraft can be broadly classified into military and civilian aircraft. Military aircraft are those used in various military fields, while civilian aircraft encompass all non-military uses (such as passenger planes, cargo planes, agricultural aircraft, sport aircraft, ambulances, and experimental research aircraft). As an advanced means of air transportation, airplanes have brought tremendous changes to the world since their inception and play a vital role in many fields, possessing profound significance. In the aviation field, aircraft attitude control plays a crucial role in flight safety and performance. With the development of aircraft towards informatization, integration, and intelligence, the information provided to pilots has exploded, making it almost impossible for pilots to make timely and correct flight decisions on their own. Therefore, establishing aircraft attitude control methods for pilot decision support is particularly important. With the continuous development of aviation technology, the flight environment faced by aircraft is becoming increasingly complex and diverse, and traditional aircraft attitude control methods are gradually revealing many limitations.

[0003] Traditional attitude control methods are mostly based on pre-defined fixed control rules and models. These rules and models are insufficient to fully encompass the various complex situations that aircraft may encounter in actual flight. Aircraft face numerous complex and variable situations during flight, such as strong airflow, bird strikes, and sudden equipment failures. These all place extremely high demands on aircraft attitude control. Traditional aircraft attitude control methods lack flexibility in responding to emergencies and are unable to make rapid, accurate, and effective adaptive adjustments, easily leading to the risk of loss of flight attitude and endangering flight safety. Existing deep learning-based aircraft attitude control methods mainly focus on optimization under conventional flight scenarios and have significant shortcomings in responding to emergencies. Therefore, it is necessary to propose a deep learning-based aircraft attitude control method to address the problems in existing technologies. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an aircraft attitude control method based on deep learning. It can introduce a real-time monitoring and adaptive adjustment mechanism into the aircraft attitude control network, so that when encountering sudden situations such as strong airflow, bird strikes or sudden equipment failures, it can quickly detect and automatically adjust the internal parameters of the network, and even change the network structure in a timely manner, thereby improving the efficiency of handling abnormal situations and ensuring the stability and safety of the aircraft's flight attitude.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a deep learning-based aircraft attitude control method, which includes the following steps:

[0006] An aircraft attitude control network based on a recurrent neural network is constructed. The input layer receives various state data of the aircraft and real-time monitoring sensor data, the recurrent hidden layer processes timing information, and the output layer outputs the aircraft attitude control quantity.

[0007] The aircraft attitude control network was trained by using the state and attitude control data of the aircraft under different flight conditions. The network was trained after the dataset was divided and its generalization ability was verified.

[0008] Deploy strong airflow sensors, bird strike sensors, and equipment failure monitoring sensors on the aircraft. Strong airflow sensors are installed on the leading edge of the aircraft wings and the top of the fuselage in areas susceptible to airflow. Bird strike sensors are installed on the nose, wings, and tail of the aircraft. Equipment failure monitoring sensors are connected to the aircraft's engine, hydraulic system, and electrical system.

[0009] During flight, data is collected by real-time monitoring sensors and transmitted to the aircraft attitude control network. The anomaly detection module in the network detects anomalies and triggers the adaptive adjustment module. The anomaly detection module adopts an anomaly detection method that combines threshold-based and custom machine learning algorithms.

[0010] The adaptive adjustment module optimizes the network in real time according to a preset response strategy. It adjusts the weights of key sensor data, activates backup neurons, and changes network structure parameters to cope with unexpected situations. The adjustment coefficient for the key sensor data weights is defined as follows: When a specific anomaly is detected, the corresponding key sensor data weights change. The backup neurons are activated based on preset activation conditions. When satisfied Activate the first One spare neuron; network structure parameters are changed, including adjusting the loop step size. ,in The adjustment amount and the number of hidden layer neurons are determined based on abnormal conditions. ,in This refers to the increase or decrease determined based on abnormal circumstances.

[0011] Furthermore, when constructing the aircraft attitude control network based on a recurrent neural network, the aircraft state data is defined as follows: Real-time monitoring of sensor data Input layer data is integrated into Define the loop hidden layer. Time of the first The input of each neuron is ,in For connection weights, For the previous moment The output of each neuron As a bias term, the activation function of the neuron is , For a custom non-linear activation function ,in and As trainable parameters, the recurrent hidden layer neurons of the recurrent neural network are connected through time steps to learn historical information about the aircraft's flight state. The update formula for the connection weights between the recurrent hidden layer neurons is as follows: ,in Update the step size for weights. The weight update is calculated based on the backpropagation algorithm. The calculation of the error term in the backpropagation algorithm involves a custom error propagation function. , This indicates element-wise multiplication. Activation function The derivative of the equation is used to calculate the error term at each time step through backpropagation, thereby updating the connection weights and better learning the complex relationship between historical flight status information and current attitude control variables.

[0012] Furthermore, the training process of the aircraft attitude control network employs a custom gradient descent algorithm. ,in For network parameters, For learning rate, For loss function about gradient, loss function Where N is the number of samples, This represents the actual attitude control quantity. For network-predicted attitude control variables. For regularization parameters, For the first Connection weights of layered networks.

[0013] Furthermore, the anomaly detection module employs an anomaly detection method that combines thresholding with machine learning algorithms, first setting thresholds for each sensor's data. When sensor data satisfy The data is labeled in real time, and then input into a custom machine learning algorithm for comprehensive judgment. The threshold is set according to the aircraft model, flight environment, and sensor characteristics. The machine learning algorithm adopts a cluster-based anomaly detection algorithm, which performs cluster analysis on sensor data under normal flight conditions to obtain cluster centers. and cluster radius When newly collected sensor data Distance from cluster center Greater than the cluster radius Data is identified as anomalous at certain times. All anomalous data is then labeled and fed into a decision tree-based classifier for further evaluation. The decision tree is constructed based on a large amount of historical anomalous data and the corresponding real anomalous types. The optimal feature is selected for node splitting using the information gain index, and the anomalous type is finally determined.

[0014] Furthermore, the adaptive adjustment module adjusts the connection weights of neurons in the recurrent hidden layer when there is a strong airflow anomaly. The determination of a strong airflow anomaly is based on the airflow intensity collected by the strong airflow sensor. , rate of change of direction Data, when and It was determined to be a strong airflow anomaly, among which and Assuming a preset threshold, the key sensor data weight adjustment coefficients are then determined based on the geometric relationship between the strong airflow direction and the aircraft attitude. If the strong airflow originates from the left side of the aircraft and primarily affects its roll attitude, then the data weights of the angle-of-attack sensor and pitot tube sensor on the left wing should be increased, and the adjustment coefficients adjusted accordingly. According to airflow intensity and the rate of change of direction Functional relationship Confirm, and simultaneously adjust the connection weights of neurons in the recurrent hidden layer, the adjustment amount of the connection weights. Data related to strong airflow is mapped using a custom function. Related.

[0015] Furthermore, the adaptive adjustment module activates backup neural pathways in the event of equipment failure, and the equipment failure monitoring sensor monitors the operating parameters of critical aircraft equipment, including engine temperature. Hydraulic system pressure Electrical system current When these parameters are outside the normal range When a fault is detected, a backup neuron pathway is activated based on a preset correspondence between the fault type and the backup neuron. For example, in the event of an engine overheating fault, a backup neuron pathway specifically designed for engine fault compensation is activated. The input to this backup neuron pathway consists of operating data from other normal equipment and historical flight status information of the aircraft, processed through pre-trained connection weights. Information is integrated, and the output is the attitude control compensation amount in the event of engine failure. ,in Enter information.

[0016] Furthermore, the adaptive adjustment module changes the recurrent neural network structure parameters upon bird strike, and the bird strike sensor detects changes in aircraft surface pressure. and vibration frequency Information determines bird strike events, when and When a bird strike is determined to have occurred, among which and The preset threshold is used to determine the location of the bird strike. ( , , (respectively, three-dimensional coordinates) and force Determine the changes in network structure parameters. If the bird strike location is close to the wingtip and the force is large, it may cause significant changes in the wing's aerodynamic performance. In this case, adjust the cycle step size. and the number of hidden layer neurons Cycle step size adjustment The relationship between the position and force of the bird strike is Increase or decrease in the number of neurons in the hidden layer The relationship with bird strikes is as follows This changes the way the network extracts and processes flight status features after localized damage to an aircraft.

[0017] Furthermore, the data collected by the real-time monitoring sensors needs to be preprocessed before being transmitted to the aircraft attitude control network. Data preprocessing includes noise removal, outlier handling, and data normalization. Noise removal employs a wavelet transform-based filtering method, decomposing the sensor data into wavelet coefficients of different frequencies and removing the corresponding wavelet coefficients based on the noise frequency characteristics. Outlier handling uses a statistical method to calculate the mean of the data. and standard deviation When data satisfy When an outlier is identified, it is corrected. To predetermine the multiple, data normalization uses a linear normalization formula. ,in and Let these be the minimum and maximum values ​​of the data, so that the data is between 0 and 1.

[0018] Furthermore, during the integration of input layer data in the aircraft attitude control network, different types of aircraft state data and real-time monitoring sensor data are encoded. For discrete data such as aircraft flap position status... One-hot encoding is used, where Representing different flap positions, convert them into binary vectors. For continuous data flight speed A segmented linear encoding method is adopted, based on the speed range. Convert the speed value to the corresponding encoded value. Encoded data is more conducive to network processing.

[0019] Furthermore, when training the aircraft attitude control network, a stratified sampling method is used to divide the dataset, stratifying it according to different categories of aircraft flight states, including takeoff, cruise, and landing, ensuring that each category has an equal proportion in the training, validation, and test sets. Early stopping is employed during training to prevent overfitting, and the loss function value for the validation set is defined to be continuous. Training is stopped when the number of epochs has not decreased. At the same time, model ensemble technology is used to train multiple recurrent neural network models with different structures and parameter initializations. When predicting the aircraft attitude control variables, a weighted average method is used to combine the outputs of multiple models.

[0020] Compared with existing technologies, this deep learning-based aircraft attitude control method has the following advantages:

[0021] This invention utilizes recurrent neural networks to deeply mine and learn the temporal information of aircraft flight status, accurately predicting the trend of aircraft attitude changes. When facing complex and ever-changing flight conditions, such as strong airflow, bird strikes, or equipment malfunctions, real-time monitoring sensors can quickly capture abnormal information. The adaptive adjustment module optimizes network parameters or structure in a timely manner according to preset strategies, effectively improving the response speed, accuracy, and stability of aircraft attitude control under emergencies. This ensures that the aircraft can maintain a safe and stable flight attitude under various extreme conditions, greatly enhancing the safety and reliability of aircraft flight and reducing the risk of flight accidents caused by emergencies.

[0022] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0024] Figure 1 This is a flowchart of a deep learning-based aircraft attitude control method. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1

[0027] Consider a commercial aircraft performing long-haul passenger flights, and apply this deep learning-based aircraft attitude control method throughout its flight journey.

[0028] During the aircraft design and manufacturing phase, the first step is to construct an aircraft attitude control network architecture based on recurrent neural networks, and to determine the data types received by the input layer, including the aircraft's real-time flight altitude information. Flight speed information Heading angle information and fuselage longitudinal acceleration information Conventional data reflecting the basic flight status of the aircraft, as well as airflow intensity data collected by strong airflow sensors. Airflow direction change data Impact strength data of aircraft surface monitored by bird strike sensors Impact location data Engine operating parameter data fed back by equipment fault monitoring sensors Hydraulic system pressure and flow data Real-time monitoring data is collected and integrated to form an input vector. .

[0029] The loop hidden layer processes the data according to time steps, the first... Time of the first The input to each neuron is calculated as follows: ,in For connection weights, For the previous moment The output of each neuron As a bias term, the activation function of the neuron is set to... , here ( and (These are the parameters to be trained). In this way, the network can learn the temporal features and complex relationships in flight data, and finally output the key parameters required for aircraft attitude control, such as aileron deflection angle, from the output layer. Elevator adjustment angle rudder rotation angle and engine thrust correction wait.

[0030] During the network training phase, a wide range of flight data samples were collected for this aircraft model under various real-world flight scenarios and simulated extreme conditions. These scenarios covered normal cruise flight, flight with airflow interference of varying intensities and directions (including situations ranging from minor turbulence to strong wind shear), simulated bird strikes at different locations and intensities, and simulated single or combined failures of various key equipment. The massive amount of flight status data collected... And the corresponding precise attitude control data The system performs preprocessing operations, using wavelet transform-based filtering techniques to remove noise components from the sensor data, and calculates the mean of the data. and standard deviation According to the formula ( To identify and correct outliers (based on coefficients determined by data distribution characteristics), a linear normalization formula is used. Mapping data to a specific range improves data quality and consistency, facilitating network training.

[0031] Subsequently, the processed dataset was divided into training, validation, and test sets according to a scientific stratified sampling strategy, and a custom gradient descent algorithm was used during training. The network parameters are iteratively optimized, where the loss function... ( For the sample size, This represents the actual attitude control quantity. For network-predicted attitude control variables. For regularization parameters, For the first The connection weights of the layer network are adjusted continuously to make the loss function value gradually converge and reach the optimum. At the same time, the training process is monitored using a validation set to prevent overfitting. Early stopping is adopted, that is, when the loss function value on the validation set is continuously decreasing, the network is stopped. Training is stopped when the number of epochs has not decreased. Multiple recurrent neural network models with different structures or parameter initializations are trained using model ensemble techniques. In actual attitude control prediction, the outputs of multiple models are combined by weighted averaging to enhance the network's generalization ability and control accuracy.

[0032] During the assembly and commissioning of the aircraft, strong airflow sensors are precisely installed on critical parts sensitive to airflow changes, such as the leading edge of the wing and the top of the fuselage. Bird strike sensors are evenly distributed in areas prone to bird strikes, such as the nose, leading edge of the wing, and tail. Equipment fault monitoring sensors are closely connected to key monitoring points of core equipment such as the engine, hydraulic system, and electrical system. All sensors are seamlessly connected to the aircraft attitude control network through a high-speed, reliable data bus with a custom encryption protocol. This ensures the integrity, accuracy, and security of data transmission and prevents data from being tampered with or lost during transmission, thus providing a stable and reliable data source for the aircraft attitude control network.

[0033] Data sets initial threshold When sensor data satisfy The data is labeled and then input into a cluster-based anomaly detection algorithm for deep analysis. This algorithm pre-analyzes a large amount of sensor data under normal flight conditions to obtain cluster centers. and cluster radius Calculate the newly acquired sensor data Distance from cluster center If the distance is greater than the cluster radius If the data is deemed abnormal, all abnormal data is marked and then fed into a decision tree-based classifier for further analysis of the abnormality type.

[0034] The decision tree is constructed based on massive historical anomaly data and corresponding real anomaly types. Optimal features are selected for node splitting using metrics such as information gain, thereby accurately determining the anomaly type and triggering the adaptive adjustment module in the attitude control network. If the anomaly detection module determines it to be a strong airflow anomaly, i.e., the airflow intensity collected by the strong airflow sensor... and the rate of change of direction satisfy and ( and When determining the threshold based on aircraft aerodynamic performance and flight safety standards, the key sensor data weight adjustment coefficients are determined according to the geometric relationship between the strong airflow direction and aircraft attitude, as well as factors such as airflow intensity and rate of change of direction. The weights of the corresponding key sensor data are calculated according to the formula. Adjustments were made, including adjusting the connection weights of neurons in the recurrent hidden layer, and the amount of adjustment for the connection weights. Data related to strong airflow is mapped using a custom function. Relatedly, this mapping function is determined based on a large amount of wind tunnel experimental data and flight simulation data. In this way, the network can quickly adapt to the dynamic influence of strong airflow on the aircraft attitude, ensuring stable flight of the aircraft in strong airflow environments.

[0035] When the equipment fault monitoring sensor detects engine temperature Hydraulic system pressure Electrical system current Parameters exceeding the normal range When a device malfunction is detected, the backup neuron pathway is activated based on the preset malfunction type and its corresponding backup neuron. The input to the backup neuron consists of operating data from other normal equipment and historical flight status information of the aircraft. This information is processed through pre-trained connection weights. The components are organically integrated, and the output is the attitude control compensation amount in the event of equipment failure. This effectively compensates for the functional deficiencies of faulty equipment, maintains the continuity and stability of aircraft attitude control, and prevents loss of aircraft attitude control due to equipment failure.

[0036] Once the bird strike sensor detects a change in pressure on the aircraft's surface... Vibration frequency Information such as and ( and To determine whether a bird strike event has occurred based on the threshold determined by the aircraft structural strength and bird strike damage model, the location of the bird strike is considered. ( , , (These are the three-dimensional coordinates of the aircraft's body coordinate system, representing its position and force.) Determine the changes in network structure parameters and the adjustment amount of the loop step size. The relationship between the position and force of the bird strike is This function is determined based on changes in the aircraft's aerodynamic shape and the principles of flight mechanics; the increase or decrease in the number of neurons in the hidden layer. The relationship with bird strikes is as follows This approach alters the network's method of extracting and processing flight state features after partial damage to the aircraft, enabling the network to quickly adapt to changes in the aircraft's structure and aerodynamic performance caused by bird strikes, ensuring that the aircraft can maintain a relatively stable flight attitude after a bird strike.

[0037] The effects of this embodiment are as follows: Throughout the entire flight, whether in normal flight conditions or encountering sudden situations such as strong airflow, bird strikes, or equipment failures, this deep learning-based aircraft attitude control method demonstrates superior performance. Through deep mining and learning of the temporal characteristics of flight data using recurrent neural networks, and the efficient collaborative operation of real-time monitoring and adaptive adjustment mechanisms, the accuracy, timeliness, and stability of aircraft attitude control are greatly improved. When facing strong airflow, it can quickly adjust the aircraft attitude to reduce the risk of turbulence and deviation from the flight path; after encountering bird strikes, it can quickly adapt to changes in aircraft structure to avoid attitude loss; when equipment failure occurs, it effectively compensates for functional deficiencies and maintains flight attitude stability. This significantly reduces the probability of flight accidents, improves flight safety, reduces flight delays and diversions caused by sudden situations, ensures passenger safety and comfort, and also improves the operational efficiency and economic benefits of airlines. It enhances the adaptability and reliability of aircraft in complex and ever-changing real-world flight environments, providing strong technical support for the steady development of the modern air transport industry.

[0038] Example 2

[0039] Taking a drone performing a military reconnaissance mission as an example, this paper illustrates the application of the deep learning-based aircraft attitude control method in special scenarios.

[0040] During the research and development and debugging phase of UAVs, an aircraft attitude control network based on recurrent neural networks is constructed. The input layer receives the UAV's flight attitude angle data (such as pitch angle). Yaw angle Roll angle Flight speed data Location coordinate data And specialized sensor data deployed specifically for military mission environments, such as electromagnetic interference monitoring sensor data. (Used to detect changes in the intensity and frequency of the surrounding electromagnetic environment), ground threat detection sensor data (such as radiation signals from ground-based air defense fire or radar reflection signals from moving targets), data from airframe structural stress sensors. (Monitoring the stress on the aircraft during high-speed maneuvers or when attacked), integrating the data to form an input data set. In the loop hidden layer, the first Time of the first Input of one neuron Activation function ,in By learning the temporal characteristics of flight data through a loop structure, the output layer outputs attitude control commands for the UAV, such as wing and control surface adjustment amounts. Tail fin control surface adjustment range and engine thrust adjustment amount wait.

[0041] During network training, data on UAVs under simulated military combat scenarios is collected, including flight status data and corresponding attitude control data under conditions such as flying in areas with different electromagnetic interference intensities, evading simulated ground-based air defense fire attacks, and performing high-speed maneuver reconnaissance maneuvers. The data is preprocessed, and noise removal is performed using a wavelet transform-based filtering method. Outlier handling is based on the data mean. and standard deviation Through formula Identify and correct outliers; data normalization is employed. The data was then divided into training, validation, and test sets, and a custom gradient descent algorithm was used. Training the network, where the loss function Furthermore, early stopping and model ensemble techniques are employed to improve network performance.

[0042] During the assembly of the UAV, electromagnetic interference monitoring sensors are installed in electromagnetically sensitive areas such as the wingtips and fuselage belly. Ground threat detection sensors are installed under the nose and on both sides of the fuselage. Airframe structural stress sensors are distributed at the wing roots and key connection points of the fuselage. These sensors are connected to the attitude control network through a high-speed, encrypted data transmission link to ensure the security and reliability of data transmission.

[0043] During military reconnaissance missions, sensors collect data in real time and transmit it to the attitude control network. The anomaly detection module first sets thresholds for the data from each sensor. ,

[0044] When sensor data satisfy The data is labeled and then fed into an anomaly detection algorithm based on clustering and decision trees. This algorithm first clusters the normal flight data to obtain cluster centers. and cluster radius Calculate the distance between the data and the cluster centers. If the value is greater than the radius, it is considered an anomaly. Furthermore, the decision tree classification determines the anomaly type and triggers the adaptive adjustment module. When the electromagnetic interference intensity... Exceeding the preset threshold When an electromagnetic interference anomaly is detected, it is determined based on the electromagnetic interference frequency. Determine the key sensor data weight adjustment coefficients based on orientation information. The key sensor data weights become Simultaneously, adjust the connection weights of neurons in the recurrent hidden layer, and adjust the connection weights by a certain amount. Electromagnetic interference data is mapped using a custom function. Correlation enables the network to adapt to the effects of electromagnetic interference.

[0045] When the ground threat detection sensor detects the strength of the ground threat signal Exceeding the threshold At that time, based on the threat direction D and distance L information, the network structure parameters are adjusted, and the cyclic step size is adjusted accordingly. Relationship with threat information is Increase or decrease in the number of neurons in the hidden layer Relationship with threat situation This allows the drone to quickly adjust its attitude to avoid evasion when the stress sensor on the airframe detects the stress value. Exceeding the threshold When an abnormal force is detected on the aircraft, a backup neural pathway is activated. Based on other flight data and historical information, the backup neurons utilize pre-trained connection weights... Output attitude control compensation amount To maintain the drone's attitude stability and ensure the continuation of reconnaissance missions.

[0046] The effects of this embodiment are as follows: In military reconnaissance missions, this aircraft attitude control method enables the UAV to have a strong ability to cope with complex environments and sudden threats. Through precise anomaly detection and rapid adaptive adjustment, it can fly stably and maintain reconnaissance effectiveness in electromagnetic interference environments; it can quickly take evasive action when facing ground threats, improving survivability; and it can maintain attitude stability even when the airframe is subjected to abnormal forces, ensuring mission continuity. This effectively improves the reliability, adaptability, and mission success rate of the UAV in military operations, provides solid technical support for military reconnaissance missions, enhances the intelligence gathering capability and battlefield situational awareness capability of the combat system, and helps to gain an advantage in complex military confrontation environments.

[0047] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the same elements of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A deep learning-based aircraft attitude control method, characterized in that, The method includes the following steps: An aircraft attitude control network based on a recurrent neural network is constructed. The input layer receives various aircraft state data and real-time monitoring sensor data, the recurrent hidden layer processes timing information, and the output layer outputs the aircraft attitude control variables. When constructing the aircraft attitude control network based on the recurrent neural network, the aircraft state data is defined as follows: Real-time monitoring of sensor data Input layer data is integrated into The circular hidden layer processes timing information, and the first circular hidden layer is defined. Time of the first The input of each neuron is ,in For connection weights, For the previous moment The output of each neuron As a bias term, the activation function of the neuron is For a custom non-linear activation function ,in and As trainable parameters, the recurrent hidden layer neurons of the recurrent neural network are connected through time steps to learn historical information about the aircraft's flight state. The update formula for the connection weights between the recurrent hidden layer neurons is as follows: ,in Update the step size for weights. The weight update is calculated based on the backpropagation algorithm. The calculation of the error term in the backpropagation algorithm involves a custom error propagation function. , This indicates element-wise multiplication. Activation function The derivative of the equation is used to calculate the error term at each time step through backpropagation at the time step, thereby updating the connection weights and better learning the complex relationship between historical flight status information and current attitude control variables. The aircraft attitude control network was trained by using the state and attitude control data of the aircraft under different flight conditions. The network was trained after the dataset was divided and its generalization ability was verified. Deploy strong airflow sensors, bird strike sensors, and equipment failure monitoring sensors on the aircraft. Strong airflow sensors are installed on the leading edge of the aircraft wings and the top of the fuselage in areas susceptible to airflow. Bird strike sensors are installed on the nose, wings, and tail of the aircraft. Equipment failure monitoring sensors are connected to the aircraft's engine, hydraulic system, and electrical system. During flight, data is collected by real-time monitoring sensors and transmitted to the aircraft attitude control network. The anomaly detection module in the network detects anomalies and triggers the adaptive adjustment module. The anomaly detection module adopts an anomaly detection method that combines threshold-based and custom machine learning algorithms. The adaptive adjustment module optimizes the network in real time according to a preset response strategy. It adjusts the weights of key sensor data, activates backup neurons, and changes network structure parameters to cope with unexpected situations. The adjustment coefficient for the key sensor data weights is defined as follows: When a specific anomaly is detected, the corresponding key sensor data weights change. The backup neurons are activated based on preset activation conditions. When satisfied Activate the first One spare neuron; network structure parameters are changed, including adjusting the loop step size. ,in The adjustment amount determined based on abnormal conditions; the number of neurons in the hidden layer. ,in This refers to the increase or decrease determined based on abnormal circumstances.

2. The deep learning-based aircraft attitude control method according to claim 1, characterized in that, The training process of the aircraft attitude control network employs a custom gradient descent algorithm. ,in For network parameters, For learning rate, For loss function about gradient, loss function ,in For the sample size, This represents the actual attitude control quantity. For network-predicted attitude control variables. For regularization parameters, For the first Connection weights of layered networks.

3. The deep learning-based aircraft attitude control method according to claim 1, characterized in that, The anomaly detection module employs an anomaly detection method combining thresholding and machine learning algorithms. The threshold is set based on the aircraft model, flight environment, and sensor characteristics. The machine learning algorithm uses a clustering-based anomaly detection algorithm to perform cluster analysis on sensor data under normal flight conditions to obtain cluster centers. and cluster radius When newly collected sensor data Distance from cluster center Greater than the cluster radius Data is identified as anomalous at certain times. All anomalous data is then labeled and fed into a decision tree-based classifier for further evaluation. The decision tree is constructed based on a large amount of historical anomalous data and the corresponding real anomalous types. The optimal feature is selected for node splitting using the information gain index, and the anomalous type is finally determined.

4. The aircraft attitude control method based on deep learning according to claim 1, characterized in that, The adaptive adjustment module adjusts the connection weights of neurons in the recurrent hidden layer when there is a strong airflow anomaly. The determination of a strong airflow anomaly is based on the airflow intensity collected by the strong airflow sensor. , rate of change of direction Data, when and It was determined to be a strong airflow anomaly, among which and Assuming a preset threshold, the key sensor data weight adjustment coefficients are then determined based on the geometric relationship between the strong airflow direction and the aircraft attitude. If the strong airflow originates from the left side of the aircraft and primarily affects its roll attitude, then the data weights of the angle-of-attack sensor and pitot tube sensor on the left wing should be increased, and the adjustment coefficients adjusted accordingly. According to airflow intensity and the rate of change of direction Functional relationship Confirm, and simultaneously adjust the connection weights of neurons in the recurrent hidden layer, the adjustment amount of the connection weights. Data related to strong airflow is mapped using a custom function. Related.

5. The deep learning-based aircraft attitude control method according to claim 1, characterized in that, The adaptive adjustment module activates backup neural pathways in the event of equipment failure, and the equipment failure monitoring sensor monitors the operating parameters of critical aircraft equipment, including engine temperature. Hydraulic system pressure Electrical system current When these parameters exceed the normal range, a device malfunction is determined. The backup neuron pathway is activated based on a pre-defined correspondence between the fault type and the backup neuron. For example, in the case of an engine overheating fault, a backup neuron pathway specifically designed for engine fault compensation is activated. The backup neuron in this pathway receives inputs from operating data of other normal equipment and historical information about the aircraft's flight status, processed through pre-trained connection weights. Information is integrated, and the output is the attitude control compensation amount in the event of engine failure. ,in Enter information.

6. The deep learning-based aircraft attitude control method according to claim 1, characterized in that, The data collected by the real-time monitoring sensors needs to be preprocessed before being transmitted to the aircraft attitude control network. Data preprocessing includes noise removal, outlier handling, and data normalization. Noise removal employs a wavelet transform-based filtering method, decomposing the sensor data into wavelet coefficients of different frequencies and removing the corresponding wavelet coefficients based on the noise frequency characteristics. Outlier handling uses a statistical method to calculate the data mean. and standard deviation When the collected data satisfy When an outlier is identified, it is corrected. To predetermine the multiple, data normalization uses a linear normalization formula. ,in and Let these be the minimum and maximum values ​​of the data, so that the data is between 0 and 1.

7. The deep learning-based aircraft attitude control method according to claim 1, characterized in that, When training the aircraft attitude control network, a stratified sampling method is used to divide the dataset, stratifying it according to different categories of aircraft flight states, including takeoff, cruise, and landing, ensuring that each category has an equal proportion in the training, validation, and test sets. Early stopping is employed during training to prevent overfitting, and the loss function value for the validation set is defined to be continuous. Training is stopped when the number of epochs has not decreased. At the same time, model ensemble technology is used to train multiple recurrent neural network models with different structures and parameter initializations. When predicting the aircraft attitude control variables, a weighted average method is used to combine the outputs of multiple models.

Images