Aircraft attitude control method based on deep learning

Abstract

The invention discloses an aircraft attitude control method based on deep learning, and relates to the technical field of video monitoring, and the method comprises the steps: constructing an aircraft attitude control network based on a recurrent neural network, and receiving various state data of an aircraft and real-time monitoring sensor data by an input layer; the time sequence information of the flight state of the aircraft is deeply mined and learned through the recurrent neural network, the attitude change trend of the aircraft is accurately predicted, the real-time monitoring sensor can rapidly capture abnormal information under the complex and changeable flight conditions such as strong airflow, bird strike or equipment faults, and the real-time monitoring sensor can accurately monitor the attitude change trend of the aircraft. The adaptive adjustment module optimizes network parameters or structures in time according to a preset strategy, effectively improves the response speed, accuracy and stability of aircraft attitude control in an emergency, ensures that an aircraft can still keep a safe and stable flight attitude in various extreme conditions, greatly enhances the flight safety and reliability of the aircraft, and improves the flight safety and reliability of the aircraft. And the risk of flight accidents caused by emergencies is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft attitude control, and in particular to an aircraft attitude control method based on deep learning. Background Art

[0002] An airplane is a heavier-than-air aircraft with wings and one or more engines that can fly in the atmosphere by its own power. Aircraft can be divided into two categories according to their use: military aircraft and civil aircraft. Military aircraft refers to aircraft used in various military fields, while civil aircraft refers to all aircraft used for non-military purposes (such as passenger aircraft, cargo aircraft, agricultural aircraft, sports aircraft, ambulance aircraft, and experimental research aircraft, etc.). As an advanced air transportation tool, airplanes have brought tremendous changes to the world since their birth, and have played an important role in many fields with far-reaching significance. In the aviation field, aircraft attitude control plays a vital role in flight safety and performance. As aircraft operations develop towards informatization, integration, and intelligence, the information provided to pilots has exploded. It is almost impossible for pilots to make timely and correct flight decisions by themselves. Therefore, it is particularly important to establish an aircraft attitude control method for pilots' auxiliary decision-making. With the continuous development of aviation technology, the flight environment faced by aircraft has become increasingly complex and diverse, and traditional aircraft attitude control methods have gradually exposed many limitations.

[0003] Traditional attitude control methods are mostly based on pre-set fixed control rules and models. These rules and models are difficult to fully cover the various complex conditions that aircraft may encounter in actual flight. Aircraft face many complex and changeable conditions during flight, such as strong airflow, bird strikes, and sudden equipment failures. These all put forward extremely high requirements for aircraft attitude control. Traditional aircraft attitude control methods are not flexible enough to deal with emergencies, and it is difficult to make accurate and effective adaptive adjustments quickly, which can easily lead to the risk of flight attitude loss of control and endanger flight safety. Existing aircraft attitude control methods based on deep learning mostly focus on optimization in conventional flight scenarios, and have obvious shortcomings in dealing with emergencies. Therefore, it is necessary to propose an aircraft attitude control method based on deep learning to solve the problems in the existing technology. Summary of the invention

[0004] The purpose of the present invention is to make up for the shortcomings of the existing technology and provide an aircraft attitude control method based on deep learning. It can introduce real-time monitoring and adaptive adjustment mechanisms in the aircraft attitude control network, so that when encountering emergencies 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 time, thereby improving the efficiency of handling abnormal situations and ensuring that the aircraft's flight attitude is stable and safe.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for aircraft attitude control based on deep learning, the method comprising the following steps: Construct an aircraft attitude control network based on a recurrent neural network. The input layer receives various aircraft status data and real-time monitoring sensor data, the recurrent hidden layer processes timing information, and the output layer outputs the aircraft attitude control value. Train the aircraft attitude control network, using the aircraft's state and attitude control data under different flight conditions, divide the data set, train the network and verify its generalization ability; Strong airflow sensors, bird strike sensors and equipment fault monitoring sensors are deployed on the aircraft. Strong airflow sensors are installed on the leading edge of the aircraft wing and the top of the fuselage, which are susceptible to airflow. Bird strike sensors are installed on the aircraft head, wings and tail. Equipment fault monitoring sensors are connected to the aircraft's engine, hydraulic system and electrical system. During the flight, real-time monitoring sensors collect data and transmit it 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 uses an anomaly detection method based on a combination of thresholds and custom machine learning algorithms. The adaptive adjustment module optimizes the network in real time according to the preset response strategy, and adjusts the weight of key sensor data, activates spare neurons and changes the network structure parameters to deal with emergencies. The key sensor data weight adjustment coefficient is defined as , when a specific anomaly is detected, the corresponding key sensor data weight becomes , the spare neurons are activated according to the preset activation conditions , when satisfied Activate spare neurons, network structure parameter changes include adjusting the cycle step length ,in is the adjustment amount and number of hidden layer neurons determined according to the abnormal situation ,in The increase or decrease is determined based on abnormal circumstances.

[0006] Furthermore, when constructing the aircraft attitude control network based on the recurrent neural network, the aircraft state data is defined as , real-time monitoring sensor data is , the input layer data is integrated into , define the recurrent hidden layer Moment The input of a neuron is ,in is the connection weight, For the previous moment The output of a neuron, is the bias term, and the activation function of the neuron is , For a custom nonlinear activation function ,in and is a trainable parameter. The neurons in the recurrent hidden layer of the recurrent neural network are connected through time steps to learn the historical information of the aircraft's flight status. The connection weight update formula between the neurons in the recurrent hidden layer is: ,in is the weight update step size, The weight update calculated by the back-propagation algorithm. The calculation of the error term in the back-propagation algorithm involves a custom error propagation function , represents the multiplication of corresponding elements, is the activation function The derivative of the flight state is used to calculate the error term at each moment through time step back propagation, and then the connection weights are updated to better learn the complex relationship between the flight state history information and the current attitude control amount.

[0007] Furthermore, the training process of the training aircraft attitude control network adopts a custom gradient descent algorithm ,in are network parameters, is the learning rate, is the loss function about The gradient of the loss function , where N is the number of samples, is the real attitude control quantity, Predict the posture control amount for the network, is the regularization parameter, For the The connection weights of the layer network.

[0008] Furthermore, the anomaly detection module adopts an anomaly detection method based on a combination of threshold and machine learning algorithm, first setting a threshold for each sensor data , when the sensor data satisfy The marked data is then input into a custom machine learning algorithm for comprehensive judgment. The threshold setting is determined according to the aircraft model, flight environment and sensor characteristics. The machine learning algorithm uses a clustering-based anomaly detection algorithm to cluster the sensor data under normal flight conditions and obtain the cluster center. and cluster radius , when the newly collected sensor data Distance from cluster center Larger than the cluster radius When an abnormal data is identified, all abnormal data are marked and input into the decision tree-based classifier for further judgment. The decision tree is constructed based on a large amount of historical abnormal data and the corresponding real abnormal type. The optimal feature is selected through the information gain index to split the node and finally determine the abnormal type.

[0009] Furthermore, the adaptive adjustment module adjusts the connection weights of the neurons in the cyclic hidden layer when the strong airflow is abnormal. The strong airflow abnormality is determined based on the airflow intensity collected by the strong airflow sensor. , Direction change rate data, when and It is judged as strong airflow abnormality when and is the preset threshold. At this time, the weight adjustment coefficient of key sensor data is determined according to the geometric relationship between the strong airflow direction and the aircraft attitude. , the strong airflow comes from the left side of the aircraft and mainly affects the aircraft's roll attitude, then increase the data weight of the angle of attack sensor and pitot sensor on the left wing, and adjust the coefficient According to the air flow strength and the rate of change of direction Functional relationship Determine, and adjust the connection weights of the neurons in the cyclic hidden layer at the same time, the connection weight adjustment amount Data related to strong airflow is mapped using a custom mapping function Associated.

[0010] Furthermore, the adaptive adjustment module activates the backup neuron pathway when equipment fails, and the equipment failure monitoring sensor monitors the operating parameters of key aircraft equipment, including engine temperature. , hydraulic system pressure , Electrical system current When these parameters are outside the normal range When the equipment fails, the spare neuron pathway is activated according to the preset correspondence between the fault type and the spare neuron. When the engine temperature is too high, the spare neuron pathway specially designed for engine failure compensation is activated. The spare neuron input in this pathway is other normal equipment operation data and aircraft flight status history information. Through the pre-trained connection weights Integrate information and output the attitude control compensation in case of engine failure ,in To input information.

[0011] Furthermore, the adaptive adjustment module changes the recurrent neural network structure parameters when a bird strikes, and the bird strike sensor detects the pressure change on the aircraft surface. and vibration frequency Information to determine bird strike events, when and A bird strike is determined when and is the preset threshold value, based on the bird strike location ( , , are the three-dimensional coordinate position) and the force Determine the amount of change in network structure parameters. The bird strike position is close to the wing tip and the force is large, which may cause a large change in the wing aerodynamic performance, so adjust the cycle step length and the number of neurons in the hidden layer , cycle step adjustment amount The relationship between the bird strike position and force is , the increase or decrease in the number of neurons in the hidden layer The relationship with bird strike is , changing the network’s flight status feature extraction and processing logic after partial damage to the aircraft.

[0012] Furthermore, the data collected by the real-time monitoring sensor needs to be preprocessed before being transmitted to the aircraft attitude control network. The data preprocessing includes noise removal, outlier processing and data normalization. Noise removal adopts a filtering method based on wavelet transform to decompose the sensor data into wavelet coefficients of different frequencies, and the corresponding wavelet coefficients are removed according to the noise frequency characteristics. Outlier processing adopts a statistical method to calculate the mean of the data. and standard deviation , when the data satisfy When it is determined as an abnormal value and corrected, The data is normalized using the linear normalization formula. ,in and The minimum and maximum values ​​of the data are set so that the data is between 0 and 1.

[0013] Furthermore, when the input layer data of the aircraft attitude control network is integrated, different types of aircraft status data and real-time monitoring sensor data are encoded and processed, and discrete data such as aircraft flap position status are encoded. One-hot encoding is used, where Represents different flap positions, converted into a binary vector , for continuous data flight speed Using piecewise linear encoding, according to the speed range Convert the speed value to the corresponding encoding value , the encoded data is more conducive to network processing.

[0014] Furthermore, when training the aircraft attitude control network, the data set is divided by a stratified sampling method, and the stratification is performed according to different categories of aircraft flight status, including takeoff, cruising and landing, so that the proportion of each category in the training set, validation set and test set is the same. The early stopping method is used to prevent overfitting during the training process, and the loss function value of the validation set is defined to be continuous. The training is stopped when the accuracy has not decreased within epochs. At the same time, the model integration technology is used to train multiple recurrent neural network models with different structures and parameter initializations. When predicting the aircraft attitude control value, the output of multiple models is integrated by weighted average.

[0015] Compared with the existing technology, this aircraft attitude control method based on deep learning has the following beneficial effects: The present invention uses a recurrent neural network to deeply mine and learn the time series information of the aircraft's flight status, accurately predict the trend of aircraft attitude changes, and when facing complex and changeable flight conditions, such as strong airflow, bird strikes or equipment failures, real-time monitoring sensors can quickly capture abnormal information, and the adaptive adjustment module timely optimizes network parameters or structures according to preset strategies, effectively improving the response speed, accuracy and stability of aircraft attitude control in emergency situations, ensuring 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 due to emergencies.

[0016] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 The figure is a flow chart of an aircraft attitude control method based on deep learning. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] Embodiment 1 Consider a commercial aircraft performing long-distance passenger flights and apply this deep learning-based aircraft attitude control method throughout its entire flight journey.

[0021] During the aircraft design and manufacturing phase, we first began to build an aircraft attitude control network architecture based on a recurrent neural network and determine the type of data received by the input layer, including the aircraft’s real-time flight altitude information. , Flight speed information , heading angle information And the longitudinal acceleration information of the fuselage 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 , aircraft surface impact strength data monitored by bird strike sensors , Impact position data , engine operating parameter data fed back by equipment fault monitoring sensors , hydraulic system pressure and flow data Real-time monitoring data such as , integrate these data into an input vector .

[0022] The recurrent hidden layer processes the data according to the time step. Moment The input of a neuron is calculated as ,in is the connection weight, For the previous moment The output of a neuron, is the bias term, and the activation function of the neuron is set to , here ( and In this way, the network can learn the temporal characteristics and complex correlations in the flight data, and finally the output layer outputs the key parameters required for aircraft attitude control, such as the aileron deflection angle. 、Elevator adjustment angle , rudder rotation angle and engine thrust correction wait.

[0023] In the network training phase, we collected a wide range of flight data samples of this type of aircraft in a variety of actual flight scenarios and simulated extreme situations. These scenarios include normal cruise flight, airflow interference flight of different strengths and directions (including from slight turbulence to strong wind shear, etc.), bird strike simulation flight of different positions and strengths, single or combined failure simulation flight of various key equipment, etc. The massive amount of flight status data collected is And the corresponding precise attitude control data , perform system preprocessing operations, use wavelet transform-based filtering technology to remove noise components in sensor data, and calculate the mean of the data and standard deviation , according to the formula ( To identify and correct outliers based on the coefficients determined by the data distribution characteristics, a linear normalization formula is used Map data uniformly to a specific interval to improve data quality and consistency and facilitate network training.

[0024] Subsequently, the processed data set is divided into training set, validation set and test set according to a scientific stratified sampling strategy, and a custom gradient descent algorithm is used in the training process. The network parameters are iteratively optimized, where the loss function ( is the sample size, is the real attitude control quantity, Predict the posture control amount for the network, is the regularization parameter, For the The connection weights of the layer network) are adjusted continuously to make the loss function value gradually converge and reach the optimal value. At the same time, the validation set is used to monitor the training process to prevent overfitting. The early stopping method is used, that is, when the validation set loss function value is continuously The training is stopped when the accuracy has not decreased within epochs, and the model integration technology is used to train multiple recurrent neural network models with different structures or parameter initializations. In the actual posture control prediction, the outputs of multiple models are combined by weighted average to enhance the generalization ability and control accuracy of the network.

[0025] During the assembly and debugging of the aircraft, strong airflow sensors are precisely installed on key locations that are sensitive to airflow changes, such as the leading edge of the aircraft's wings and the top of the fuselage. Bird strike sensors are evenly distributed on the aircraft's head, leading edge of the wings, tail wing and other areas susceptible to bird strikes. 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 that uses a custom encryption protocol, to ensure the integrity, accuracy and security of data transmission, and prevent data tampering or loss during transmission, thereby providing a stable and reliable data source for the aircraft attitude control network.

[0026] Data sets initial threshold , when the sensor data satisfy The data is marked when the flight is started, and then the marked data is input into the clustering-based anomaly detection algorithm for in-depth analysis. The algorithm pre-clustered a large amount of sensor data under normal flight conditions to obtain the cluster center. and cluster radius , calculate the newly collected sensor data Distance from cluster center , if the distance is greater than the cluster radius It is determined to be abnormal data, and all abnormal data are marked and input into the decision tree-based classifier to further determine the abnormal type.

[0027] The construction of the decision tree is based on massive historical abnormal data and the corresponding real abnormal types. The optimal features are selected for node splitting through indicators such as information gain, so as to accurately determine the abnormal type and trigger the adaptive adjustment module in the attitude control network. If the abnormal detection module determines that it is a strong airflow abnormality, that is, the airflow intensity collected by the strong airflow sensor and the rate of change of direction satisfy and ( and When the threshold is determined based on the aircraft aerodynamic performance and flight safety standards), the weight adjustment coefficient of key sensor data is determined according to the geometric relationship between the strong airflow direction and the aircraft attitude, as well as factors such as airflow intensity and direction change rate. , the corresponding key sensor data weights are calculated according to the formula Make adjustments and adjust the connection weights of the neurons in the cyclic hidden layer at the same time. Data related to strong airflow is mapped using a custom mapping function The mapping function is determined based on a large amount of wind tunnel test data and flight simulation data. In this way, the network can quickly adapt to the dynamic impact of strong airflow on the aircraft attitude and ensure the stable flight of the aircraft in a strong airflow environment.

[0028] When the equipment fault monitoring sensor detects that the engine temperature , hydraulic system pressure , Electrical system current Parameters outside the normal range When the equipment failure occurs, the backup neuron pathway is activated according to the preset fault type and the corresponding relationship between the backup neuron. The input of the backup neuron is other normal equipment operation data and aircraft flight status history information. This information is connected through the pre-trained weights. The system is organically integrated and the output is the attitude control compensation in the event of equipment failure. , thereby effectively compensating for the functional loss corresponding to the faulty equipment, maintaining the continuity and stability of the aircraft attitude control, and avoiding the loss of aircraft attitude control due to equipment failure.

[0029] Once the bird strike sensor detects a change in pressure on the aircraft surface , vibration frequency Waiting for information to satisfy and ( and The bird strike event is determined based on the aircraft structural strength and the threshold value determined by the bird strike damage model. ( , , are the three-dimensional coordinate position in the aircraft body coordinate system) and the force Determine the change in network structure parameters and the adjustment of cycle step length The relationship between the bird strike position and force is , which is determined based on the change of the aircraft's aerodynamic shape and the principle of flight mechanics; the increase or decrease in the number of neurons in the hidden layer The relationship with bird strike is In this way, the network's flight status feature extraction method and processing logic after the aircraft is partially damaged are changed, so that the network can quickly adapt to the changes in the aircraft's structure and aerodynamic performance caused by bird strikes, ensuring that the aircraft can still maintain a relatively stable flight attitude after a bird strike.

[0030] Effects brought by this embodiment: During the entire flight process, whether it is a normal flight state or encountering emergencies such as strong airflow, bird strikes or equipment failures, the aircraft attitude control method based on deep learning shows excellent performance. Through the deep mining and learning of the time series characteristics of flight data by the recurrent neural network, and the efficient coordinated operation of the real-time monitoring and adaptive adjustment mechanism, the accuracy, timeliness and stability of the aircraft attitude control are greatly improved. When facing strong airflow, the aircraft attitude can be quickly adjusted to reduce the risk of turbulence and deviation from the route; after encountering a bird strike, it can quickly adapt to changes in the aircraft structure to avoid attitude loss; when equipment failure occurs, it effectively compensates for the lack of function and maintains a stable flight attitude, which significantly reduces the probability of flight accidents, improves flight safety, reduces flight delays and diversions caused by emergencies, ensures the safety and comfort of passengers, and also improves the operating efficiency and economic benefits of airlines, enhances the adaptability and reliability of aircraft in complex and changeable actual flight environments, and provides strong technical support for the steady development of the modern aviation transportation industry.

[0031] Embodiment 2 Taking a drone performing a military reconnaissance mission as an example, this paper explains the application of this deep learning-based aircraft attitude control method in special scenarios.

[0032] During the development and debugging phase of the drone, an aircraft attitude control network based on a recurrent neural network is constructed. The input layer receives the flight attitude angle data of the drone (such as pitch angle , yaw angle , Roll Angle ), flight speed data , location coordinate data , as well as special sensor data deployed specifically for military mission environments, such as electromagnetic interference monitoring sensor data (used to detect changes in the strength and frequency of the surrounding electromagnetic environment), ground threat detection sensor data (such as detecting radiation signals from ground air defense firepower or radar reflection signals from moving targets), airframe structural stress sensor data (monitoring the stress of the aircraft during high-speed maneuvers or when attacked) and integrating them into an input data set , in the recurrent hidden layer, Moment Neuron input , activation function ,in , the time series characteristics of flight data are learned through the loop structure, and the output layer outputs the attitude control instructions of the UAV, such as the adjustment amount of the wing rudder , Tail rudder adjustment amount and engine thrust adjustment wait.

[0033] During the network training process, the data of the UAV in simulated military combat scenarios are collected, including flight status data and corresponding attitude control data when flying in different electromagnetic interference intensity areas, avoiding simulated ground air defense fire attacks, and performing high-speed mobile reconnaissance actions. The data is preprocessed, and the noise is removed by using a filtering method based on wavelet transform. The outlier processing is based on the data mean. and standard deviation , through the formula Identify and correct outliers, and normalize data using , and then divide the data into training set, validation set and test set, using a custom gradient descent algorithm Train the network, where the loss function , and uses early stopping and model integration technology to improve network performance.

[0034] When assembling the UAV, electromagnetic interference monitoring sensors are installed at the wing tips and belly of the fuselage and other areas that are sensitive to electromagnetic signals. Ground threat detection sensors are installed under the nose and on both sides of the fuselage. Body structural stress sensors are distributed at the wing roots and key connection parts 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.

[0035] When performing military reconnaissance missions, sensors collect data in real time and transmit it to the attitude control network. The anomaly detection module first sets a threshold for each sensor data. , When sensor data satisfy The labeled data is then input into an anomaly detection algorithm based on clustering and decision trees. The algorithm first clusters the normal flight data to obtain the cluster centers. and cluster radius , calculate the distance between data and cluster center , it is considered abnormal if it is greater than the radius. Then the decision tree classification determines the abnormal type and triggers the adaptive adjustment module. Exceeding the preset threshold When the electromagnetic interference is abnormal, the electromagnetic interference frequency and direction information to determine the weight adjustment coefficient of key sensor data , the key sensor data weight becomes , and adjust the connection weights of the neurons in the cyclic hidden layer at the same time, and the connection weight adjustment amount and EMI data through custom mapping functions association, making the network adaptable to the influence of electromagnetic interference.

[0036] When the ground threat detection sensor detects the ground threat signal strength Exceeding the threshold When the threat direction D and distance L are used, the network structure parameters are adjusted and the cycle step length is adjusted. The relationship with threat information is , the increase or decrease in the number of neurons in the hidden layer The relationship with the threat situation is , so that the drone can quickly make evasive posture adjustments. When the body structure stress sensor detects the stress value Exceeding the threshold When the aircraft is judged to be abnormally stressed, the spare neuron pathway is activated. The spare neurons are connected by pre-trained weights based on other flight data and historical information. Output attitude control compensation , maintain the stability of the UAV's attitude and ensure that the reconnaissance mission can continue.

[0037] Effects brought about by this embodiment: In the military reconnaissance mission scenario, this aircraft attitude control method enables the UAV to have a strong ability to cope with complex environments and sudden threats. Through accurate anomaly detection and rapid adaptive adjustment, it can fly stably and maintain reconnaissance effectiveness in an electromagnetic interference environment; it can quickly make evasive actions when facing ground threats to improve survivability; it can still maintain a stable attitude when the body is subjected to abnormal forces to ensure the continuity of the mission, which effectively improves the reliability, adaptability and mission success rate of the UAV in military operations, provides a solid technical guarantee for military reconnaissance missions, enhances the intelligence collection capability and battlefield situation awareness capability of the combat system, and helps to gain advantages in complex military confrontation environments.

[0038] It will be apparent to those skilled in the art that the 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 the spirit or essential features of the invention. 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, and it is intended that all variations falling within the meaning and scope of the same elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for aircraft attitude control based on deep learning, characterized in that: The method comprises the following steps: Construct an aircraft attitude control network based on a recurrent neural network. The input layer receives various aircraft status data and real-time monitoring sensor data, the recurrent hidden layer processes timing information, and the output layer outputs the aircraft attitude control value. Train the aircraft attitude control network, using the aircraft's state and attitude control data under different flight conditions, divide the data set, train the network and verify its generalization ability; Strong airflow sensors, bird strike sensors and equipment fault monitoring sensors are deployed on the aircraft. Strong airflow sensors are installed on the leading edge of the aircraft wing and the top of the fuselage, which are susceptible to airflow. Bird strike sensors are installed on the aircraft head, wings and tail. Equipment fault monitoring sensors are connected to the aircraft's engine, hydraulic system and electrical system. During the flight, real-time monitoring sensors collect data and transmit it 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 uses an anomaly detection method based on a combination of thresholds and custom machine learning algorithms. The adaptive adjustment module optimizes the network in real time according to the preset response strategy, and adjusts the weight of key sensor data, activates spare neurons and changes the network structure parameters to deal with emergencies. The key sensor data weight adjustment coefficient is defined as , when a specific anomaly is detected, the corresponding key sensor data weight becomes , the spare neurons are activated according to the preset activation conditions , when satisfied Activate spare neurons, network structure parameter changes include adjusting the cycle step length ,in is the adjustment amount and number of hidden layer neurons determined according to the abnormal situation ,in The increase or decrease is determined based on abnormal circumstances.

2. The aircraft attitude control method based on deep learning according to claim 1, characterized in that: When constructing the aircraft attitude control network based on the recurrent neural network, the aircraft state data is defined as , real-time monitoring sensor data is , the input layer data is integrated into , define the recurrent hidden layer Moment The input of a neuron is ,in is the connection weight, For the previous moment The output of a neuron, is the bias term, and the activation function of the neuron is , For a custom nonlinear activation function ,in and is a trainable parameter. The neurons in the recurrent hidden layer of the recurrent neural network are connected through time steps to learn the historical information of the aircraft's flight status. The connection weight update formula between the neurons in the recurrent hidden layer is: ,in is the weight update step size, The weight update calculated by the back-propagation algorithm. The calculation of the error term in the back-propagation algorithm involves a custom error propagation function , represents the multiplication of corresponding elements, is the activation function The derivative of the flight state is used to calculate the error term at each moment through time step back propagation, and then the connection weights are updated to better learn the complex relationship between the flight state history information and the current attitude control amount.

3. The aircraft attitude control method based on deep learning according to claim 1, characterized in that: The training process of the training aircraft attitude control network adopts a custom gradient descent algorithm ,in are network parameters, is the learning rate, is the loss function about The gradient of the loss function , where N is the number of samples, is the real attitude control quantity, Predict the posture control amount for the network, is the regularization parameter, For the The connection weights of the layer network.

4. The aircraft attitude control method based on deep learning according to claim 1, characterized in that: The anomaly detection module adopts an anomaly detection method based on a combination of threshold and machine learning algorithm. First, a threshold is set for each sensor data. , when the sensor data satisfy The marked data is then input into a custom machine learning algorithm for comprehensive judgment. The threshold setting is determined according to the aircraft model, flight environment and sensor characteristics. The machine learning algorithm uses a clustering-based anomaly detection algorithm to cluster the sensor data under normal flight conditions and obtain the cluster center. and cluster radius , when the newly collected sensor data Distance from cluster center Larger than the cluster radius When an abnormal data is identified, all abnormal data are marked and input into the decision tree-based classifier for further judgment. The decision tree is constructed based on a large amount of historical abnormal data and the corresponding real abnormal type. The optimal feature is selected through the information gain index to split the node and finally determine the abnormal type.

5. 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 the neurons in the cyclic hidden layer when the strong airflow is abnormal. The strong airflow abnormality is determined based on the airflow intensity collected by the strong airflow sensor. , Direction change rate data, when and It is judged as strong airflow abnormality when and is the preset threshold. At this time, the weight adjustment coefficient of key sensor data is determined according to the geometric relationship between the strong airflow direction and the aircraft attitude. , the strong airflow comes from the left side of the aircraft and mainly affects the aircraft's roll attitude, then increase the data weight of the angle of attack sensor and pitot sensor on the left wing, and adjust the coefficient According to the air flow strength and the rate of change of direction Functional relationship Determine, and adjust the connection weights of the neurons in the cyclic hidden layer at the same time, the connection weight adjustment amount Data related to strong airflow is mapped using a custom mapping function Associated.

6. The aircraft attitude control method based on deep learning according to claim 1, characterized in that: The adaptive adjustment module activates the backup neuron pathway when equipment fails, and the equipment failure monitoring sensor monitors the operating parameters of key aircraft equipment, including engine temperature. , hydraulic system pressure , Electrical system current When these parameters are outside the normal range When the equipment fails, the spare neuron pathway is activated according to the preset correspondence between the fault type and the spare neuron. When the engine temperature is too high, the spare neuron pathway specially designed for engine failure compensation is activated. The spare neuron input in this pathway is other normal equipment operation data and aircraft flight status history information. Through the pre-trained connection weights Integrate information and output the attitude control compensation in case of engine failure ,in To input information.

7. The aircraft attitude control method based on deep learning according to claim 1, characterized in that: The adaptive adjustment module changes the recurrent neural network structure parameters when a bird strikes. The bird strike sensor detects changes in the pressure on the aircraft surface. and vibration frequency Information to determine bird strike events, when and A bird strike is determined when and is the preset threshold value, based on the bird strike location ( , , are the three-dimensional coordinate position) and the force Determine the amount of change in network structure parameters. The bird strike position is close to the wing tip and the force is large, which may cause a large change in the wing aerodynamic performance, so adjust the cycle step length and the number of neurons in the hidden layer , cycle step adjustment amount The relationship between the bird strike position and force is , the increase or decrease in the number of neurons in the hidden layer The relationship with bird strike is , changing the network’s flight status feature extraction and processing logic after partial damage to the aircraft.

8. The aircraft attitude control method based on deep learning according to claim 1, characterized in that: The data collected by the real-time monitoring sensor must be preprocessed before being transmitted to the aircraft attitude control network. Data preprocessing includes noise removal, outlier processing and data normalization. Noise removal uses a filtering method based on wavelet transform to decompose the sensor data into wavelet coefficients of different frequencies, and the corresponding wavelet coefficients are removed according to the noise frequency characteristics. Outlier processing uses a statistical method to calculate the mean of the data. and standard deviation , when the data satisfy When it is determined as an abnormal value and corrected, The data is normalized using the linear normalization formula. ,in and The minimum and maximum values ​​of the data are set so that the data is between 0 and 1.

9. The aircraft attitude control method based on deep learning according to claim 1, characterized in that: When the input layer data of the aircraft attitude control network is integrated, different types of aircraft status data and real-time monitoring sensor data are encoded and processed. For discrete data such as aircraft flap position status One-hot encoding is used, where Represents different flap positions, converted into a binary vector , for continuous data flight speed Using piecewise linear encoding, according to the speed range Convert the speed value to the corresponding encoding value , the encoded data is more conducive to network processing.

10. The aircraft attitude control method based on deep learning according to claim 1, characterized in that: When training the aircraft attitude control network, the data set is divided by a stratified sampling method, and the stratification is performed according to different categories of aircraft flight status, including take-off, cruising and landing, so that the proportion of each category in the training set, the validation set and the test set is the same. The early stopping method is used to prevent overfitting during the training process, and the loss function value of the validation set is defined to be continuous. The training is stopped when the accuracy has not decreased within epochs. At the same time, the model integration technology is used to train multiple recurrent neural network models with different structures and parameter initializations. When predicting the aircraft attitude control value, the output of multiple models is integrated by weighted average.

Images