阅读说明：本技术 一种基于机器学习的多任务模式近地告警方法 (Multi-task mode ground proximity warning method based on machine learning ) 是由 张恒 祁鸣东 彭旭飞 雷雨 曹植 祖肇梓 于 2021-09-10 设计创作，主要内容包括：本申请提供一种基于机器学习的多任务模式近地告警方法,所述方法包括：构建数字地形图,所述数字地形图包括海拔高度；根据基本飞行参数构建飞机反应模型,根据所述飞机反应模型获得拉起过程中的下降高度ΔH；根据飞机初始高度和下降高度ΔH,计算飞机拉起后最低高度；比较海拔高度和飞机拉起后最低高度,判断飞机是否触地,并生成训练数据；构建机器学习模型,并对机器学习模型进行训练；利用机器学习模型,输出近地告警信号。(The application provides a multitask mode ground proximity warning method based on machine learning, which comprises the following steps: constructing a digital terrain map, the digital terrain map comprising an altitude; constructing an airplane reaction model according to basic flight parameters, and obtaining a descending height delta H in the pulling-up process according to the airplane reaction model; calculating the lowest height of the pulled airplane according to the initial height and the descending height delta H of the airplane; comparing the altitude with the lowest altitude of the plane after being pulled up, judging whether the plane touches the ground or not, and generating training data; constructing a machine learning model, and training the machine learning model; and outputting a ground proximity warning signal by using a machine learning model.)

1. A multitask mode ground proximity warning method based on machine learning, which is characterized by comprising the following steps:

constructing a digital terrain map, the digital terrain map comprising an altitude;

constructing an airplane reaction model according to basic flight parameters, and obtaining a descending height delta H in the pulling-up process according to the airplane reaction model;

calculating the lowest height of the pulled airplane according to the initial height and the descending height delta H of the airplane; comparing the altitude with the lowest altitude of the plane after being pulled up, judging whether the plane touches the ground or not, and generating training data;

constructing a machine learning model, and training the machine learning model;

and outputting a ground proximity warning signal by using a machine learning model.

2. The method according to claim 1, wherein an aircraft reaction model is constructed according to basic flight parameters, and the obtaining of the descent height Δ H during the pulling-up process according to the aircraft reaction model specifically comprises:

constructing an airplane reaction model according to the alarm mode;

performing kinematic analysis on a pulling-up stage of the airplane in the flying process to obtain a descending height delta H in the pulling-up process;

basic flight parameters are set.

3. The method of claim 2, wherein the basic flight parameters include pilot's operational response time tdelay, aircraft track inclination μ, normal overload n, initial altitude L₀Initial rate of decrease V_down0Initial track angle mu₀。

4. The method according to claim 3, wherein the kinematic analysis of the aircraft during the phase of lift-off during flight to obtain the height of descent Δ H during lift-off comprises:

calculating thrust F_NResistance D and normal aircraft overload n;

according to the resistance D and the thrust F_NAnd normal overload n, using formulaCalculating the speed V and the track inclination angle mu of the airplane;

wherein g is the acceleration of gravity and m is the total weight of the aircraft;

according to the speed V and the track inclination angle mu of the airplane, using a formulaCalculating the height of descent Δ H during the pull-up process, wherein t₀To time an alarm occurs, Δ t is the pull-up phase time.

5. The method according to claim 1, characterized in that the minimum height after the aircraft is pulled up is calculated from the initial altitude and the descent altitude Δ H of the aircraft; comparing the altitude with the lowest altitude of the plane after being pulled up, judging whether the plane touches the ground or not, and generating training data, wherein the method specifically comprises the following steps:

carrying out multiple flight simulation on the airplane by using the airplane reaction model, and obtaining simulation data of whether the airplane touches the ground or not;

generating a corresponding label according to the simulation data, wherein the label comprises an alarm and a non-alarm;

and sampling for multiple times according to the critical point data of alarming and non-alarming so as to obtain balanced training data, and obtaining a training set and a test set according to the training data.

6. The method of claim 1, wherein constructing and training the machine learning model specifically comprises:

constructing a machine learning model, wherein the input of the machine learning model is a descent rate, a horizontal rate, a radio altitude, a horizontal speed, a terrain approach rate, a landing gear state and an ILS state, and the output of the machine learning model comprises a non-alarm mode and an alarm mode;

training the machine learning model by utilizing the training set;

and testing the machine learning model by using the test set.

7. The method of claim 1, wherein the machine learning models include DNN models, support vector machines, bayesian models, and random forests.

8. The method of claim 3,

the value range of the operation response time is gamma distribution G (4, 1);

the numeric area of the change rate of the flight path angle is gamma distribution G (7, 0.3);

the value range of the touchdown time is uniformly distributed U (20, 35);

the value range of the landing gear state is binomial distribution (0, 1);

the value range of the ILS state is a binomial distribution (0, 1, 2).

Technical Field

The invention relates to a ground proximity warning system, in particular to a multitask mode ground proximity warning method based on machine learning.

Background

In a ground proximity warning system, the design of an envelope is often influenced by a plurality of factors, and in a specific terrain condition, the traditional method is to modify the threshold value of the envelope manually and has low efficiency, so how to improve the adaptability of different flight conditions under the condition of ensuring the safety of an airplane. The method combines the real performance state, the real topographic map and the simulated topographic map of the airplane to simulate the flight parameters of the airplane, and trains a general near-ground warning machine learning warning model by using simulated data. The purpose of outputting different ground proximity warning signals according to environment information such as different terrains, normal overload and the like and airplane information is achieved. On the basis of ensuring the accuracy by accurate modeling, different environment information and airplane information parameters can be designed for various task modes, and the generation efficiency of the ground proximity alarm scheme is improved by training the model by a machine learning technology.

Disclosure of Invention

The application provides a multitask mode ground proximity warning method based on machine learning, and the efficiency and the accuracy of warning can be improved.

In order to solve the technical problem, the present application provides a multitask mode ground proximity warning method based on machine learning, the method including:

constructing a digital terrain map, the digital terrain map comprising an altitude;

constructing an airplane reaction model according to basic flight parameters, and obtaining a descending height delta H in the pulling-up process according to the airplane reaction model;

calculating the lowest height of the pulled airplane according to the initial height and the descending height delta H of the airplane; comparing the altitude with the lowest altitude of the plane after being pulled up, judging whether the plane touches the ground or not, and generating training data;

constructing a machine learning model, and training the machine learning model;

and outputting a ground proximity warning signal by using a machine learning model.

Preferably, an aircraft reaction model is constructed according to the basic flight parameters, and the descent height Δ H in the pulling-up process is obtained according to the aircraft reaction model, which specifically includes:

constructing an airplane reaction model according to the alarm mode;

performing kinematic analysis on a pulling-up stage of the airplane in the flying process to obtain a descending height delta H in the pulling-up process;

basic flight parameters are set.

Preferably, the basic flight parameters include pilot's operational response time tdelay, aircraft track inclination μ, normal overload n, initial altitude L₀Initial rate of decrease V_down0Initial track angle mu₀。

Preferably, the kinematic analysis is performed on the aircraft in the pulling-up stage in the flight process to obtain the descent height Δ H in the pulling-up process, and specifically includes:

calculating thrust F_NResistance D and normal aircraft overload n;

according to the resistance D and the thrust F_NAnd normal overload n, using formulaCalculating the speed V and the track inclination angle mu of the airplane;

wherein g is the acceleration of gravity and m is the mass of the airplane;

according to the speed V and the track inclination angle mu of the airplane, using a formulaCalculating the height of descent Δ H during the pull-up process, wherein t₀To time an alarm occurs, Δ t is the pull-up phase time.

Preferably, the lowest height of the airplane after being pulled up is calculated according to the initial height and the descending height delta H of the airplane; comparing the altitude with the lowest altitude of the plane after being pulled up, judging whether the plane touches the ground or not, and generating training data, wherein the method specifically comprises the following steps:

carrying out multiple flight simulation on the airplane by using the airplane reaction model, and obtaining simulation data of whether the airplane touches the ground or not;

generating a corresponding label according to the simulation data, wherein the label comprises an alarm and a non-alarm;

and sampling for multiple times according to the critical point data of alarming and non-alarming so as to obtain balanced training data, and obtaining a training set and a test set according to the training data.

Preferably, the constructing and training of the machine learning model specifically includes:

constructing a machine learning model, wherein the input of the machine learning model is a descent rate, a horizontal rate, a radio altitude, a horizontal speed, a terrain approach rate, a landing gear state and an ILS state, and the output of the machine learning model comprises a non-alarm mode and an alarm mode;

training the machine learning model by utilizing the training set;

and testing the machine learning model by using the test set.

Preferably, the machine learning model includes Deep Neural Network (DNN) model, support vector machine, bayesian model and random forest.

Preferably, the value range of the operation response time is gamma distribution G (4, 1);

the numeric area of the change rate of the flight path angle is gamma distribution G (7, 0.3);

the value range of the touchdown time is uniformly distributed U (20, 35);

the value range of the landing gear state is binomial distribution (0, 1);

the value range of the ILS state is a binomial distribution (0, 1, 2).

In summary, the application provides a multi-task mode ground proximity warning method based on machine learning, which simulates flight parameters of an airplane by combining the real performance state of the airplane, a real topographic map and a simulated topographic map, and trains a general ground proximity warning machine learning warning model by using simulated data. The purpose of outputting different ground proximity alarm signals according to different terrains and normal overload is achieved, specific alarm signals under corresponding working conditions can be automatically output according to different parameter settings, and the method has higher flexibility. Compared with the traditional parameterized simulation method, the method solves the problem of low efficiency caused by manually fitting the envelope and adjusting the alarm threshold.

Drawings

FIG. 1 is a schematic structural diagram of a ground proximity warning system provided in the present application;

FIG. 2 is a schematic diagram illustrating a movement phase of an aircraft after an alert signal provided by the present application has occurred;

fig. 3 is a schematic diagram of a force analysis provided in the present application.

Fig. 4 is a thrust force graph.

Fig. 5 is a graph of lift and drag levels for an aircraft.

Detailed Description

The machine learning technology can fit nonlinear data, and the alarm envelope design requirements under different terrains are met. The method generates different alarm mechanisms aiming at specific training tasks by simulating the terrain under different scenes, thereby improving the performance of the airplane, and the overall architecture is shown in figure 1.

Step 1: constructing a digital terrain map, the digital terrain map comprising an altitude;

specifically, the actual digital terrain elevation map is collected, and various digital terrain elevation map environments are established. And respectively collecting various digital terrain elevation map data of cities, plains, mountainous areas, oceans and the like, and taking the data as a test scene. And simultaneously, generating a simulated topographic map by adopting a Markov method.

Step 2: constructing an airplane reaction model according to basic flight parameters, and obtaining a descending height delta H in the pulling-up process according to the airplane reaction model;

step 2-1: constructing an airplane reaction model according to the alarm mode;

the common alarm modes include six, namely a mode 1: the descent rate is too big to report an emergency and ask for help or increased vigilance, and mode 2 reports an emergency and asks for help or increased vigilance for the too big rate of topography approach, divide into 2A and 2B mode, and mode 3 reports an emergency and asks for help or increased vigilance for takeoff or re-flying, and mode 4 is that the headroom is not enough under the non-landing form, divide into 4A, 4B, 4C according to undercarriage state and flight phase, and wherein 4A divide into topography and undercarriage according to the airspeed again and report an emergency and ask for help or increased vigilance for 4B. The mode 5 is a warning of the deviation of the lower slideway, and the mode 6 is a warning of the overlarge roll angle.

Taking mode 1 as an example, the following studies were carried out:

as shown in fig. 2, the aircraft reaction model includes three phases: reaction stage, pull-up stage, and stability maintaining stage.

Wherein, in the reaction phase, the alarm is generated, but the motion state of the airplane is not changed, and the time is defined as the reaction time. In the pull-up stage, the state of the airplane is changed, and the pull-up action is generated until the pull-up angle is stably maintained. The steady hold phase refers to flight with a constant track angle.

Step 2-2: performing kinematic analysis on a pulling-up stage of the airplane in the flying process to obtain a descending height delta H in the pulling-up process;

specifically, the height Δ H of the drop during the pull-up process includes:

step 2-2-1: calculating thrust F_NResistance D and normal aircraft overload n;

specifically, the thrust force F_NThe current mach number M, the altitude H, and the query thrust curve can be obtained, see fig. 4.

Specifically, the normal overload n can be obtained through an onboard instrument.

In particular, the current drag D on the aircraft is determined by the drag coefficient C of the aircraft_DThus obtaining the product.

D＝1/(2C_DρV²S_W)

It should be noted that, the air density ρ ═ f (h) may refer to a standard atmospheric table, V is the current airspeed of the aircraft, and Sw is the wing area.

Note that the lift coefficient C is defined by_LObtaining the current Mach number M of the airplane obtained by the airplane instrument, and obtaining the drag coefficient C of the airplane by inquiring the lift-drag level curve of the airplane_D. See fig. 5.

Note that the lift coefficient C_LThe calculation method comprises the following steps:

coefficient of lift C_L＝(2L)/(ρV²S_W)

The formula air density ρ ═ f (h) can be referred to a standard atmospheric table, V is the current airspeed of the aircraft, and Sw is the wing area. L is aircraft lift.

The aircraft lift force L is calculated by the following method: calculating to obtain an aircraft lift force L by using a formula n which is L/G according to the gravity G and the aircraft normal overload n; wherein, the gravity G is mg, m is the mass of the airplane, and G is the gravity acceleration of the position of the airplane. And n is normal overload and can be obtained through an onboard instrument.

Step 2-2-2: according to the resistance D and the thrust F_NAnd normal overload n, using formulaCalculating the change rate of the speed V of the airplane and the change rate of the track inclination angle mu;

wherein g is the gravitational acceleration and m is the aircraft mass.

Step 2-2-3: according to the speed V and the track inclination angle mu of the airplane, using a formulaCalculating the height of descent Δ H during the pull-up process, wherein t₀To time an alarm occurs, Δ t is the pull-up phase time.

And 2-3, setting basic flight parameters.

Wherein the basic flight parameters include pilot's operational response time t_delayInclination angle mu of flight path of airplane, normal overload n and initial height L₀Initial rate of decrease V_down0Initial track angle mu₀Normal overload n₀The parameters are all random numbers. The range of the basic flight parameters is as follows:

meaning of variable	Variable sign	Distribution state	Model parameters
				Cruising speed	Vt	Is uniformly distributed	Vmin-Vmax
Operational response time	tdelay	Gamma distribution	G(4，1)
				Rate of change of track angle	μ	Gamma distribution	G(7，0.3)
Initial rate of decline	Vdown0	Is uniformly distributed	Vdownmin,Vdownmax
				Time to ground	T	Is uniformly distributed	(20，35)
Landing gear state	gear	Distribution of two terms	0，1
				ILS State	ILS	Distribution of two terms	0，1，2
Radio altitude	AGL	Is uniformly distributed	AGLmax-AGLmin

And step 3: calculating the lowest height of the pulled airplane according to the initial height and the descending height delta H of the airplane; comparing the altitude with the lowest altitude of the plane after being pulled up, judging whether the plane touches the ground or not, and generating training data;

3-1, performing multiple flight simulation on the airplane by using an airplane reaction model, and acquiring simulation data of whether the airplane touches the ground or not;

in practical application, random numbers are generated according to the space range specified by data, and then the time from alarm to ground collision of the aircraft under different thresholds is calculated.

And 3-1-1, calculating the motion track and the touchdown time in the non-alarm mode.

Randomly generating radio altitude, randomly generating descent rate, randomly generating terrain, randomly generating touchdown time T, calculating a normal track in a no-alarm mode, and recording whether touchdown occurs within T seconds according to whether the terrain altitude is in contact with the pull-up falling altitude and the terrain altitude.

And 3-1-2, calculating a track in an alarm mode, and recording whether the contact is made within T seconds.

Randomly generating operation corresponding time t_delayAnd randomly generating normal overload n, inquiring corresponding airplane model parameter information, and calculating delta H according to the step 2-2.

And 3-1-3, circulating for multiple times, and respectively counting the grounding states when the alarm is given and the grounding states when the alarm is not given.

3-2, generating corresponding labels according to the simulation data, wherein the labels comprise alarming and non-alarming;

	case 1	Case 2	Case 3	Case 4
					Without alarm	Ground contact	Ground contact	Without touching ground	Without touching ground
Alarm system	Ground contact	Without touching ground	Ground contact	Without touching ground
					Label (R)	Alarm system	Alarm system	Do not give an alarm	Do not give an alarm

Similarly, the simulation data of other modes are simulated and the label is generated.

And 3-3, sampling for multiple times according to the critical point data of alarming and non-alarming so as to obtain balanced training data, and obtaining a training set and a test set according to the training data.

And randomly sampling the alarm samples and the non-alarm samples to obtain the scale samples with the same number as the positive and negative samples. Meanwhile, according to the training set: the test set was 7: and 3, extracting the samples according to the proportion to obtain a training set and a test set.

Step 4, constructing a machine learning model, and training the machine learning model;

the machine learning model includes a Deep Neural Networks (DNN) model.

Step 4-1, constructing a machine learning model, wherein the input of the machine learning model is a descending rate, a horizontal rate, a radio altitude, a horizontal speed, a terrain approaching rate, an undercarriage state and an ILS state, and the output of the machine learning model comprises a non-alarm mode and an alarm mode;

the input and output of the machine learning model are set as follows:

in practical application, the alarm mode 1 is an alarm for over-large descent rate, the alarm mode 2 is an alarm for over-large terrain approach rate and is divided into 2A and 2B modes, the alarm mode 3 is an alarm for takeoff or missed approach, the alarm mode 4 is an alarm for insufficient clearance height in a non-landing state, the alarm mode is divided into 4A, 4B and 4C according to the landing gear state and the flight stage, wherein the alarm mode 4A is divided into terrain and landing gear alarms according to airspeed, and the alarm mode 4B is divided into terrain and flap alarms. The mode 5 is a warning of the deviation of the lower slideway, and the mode 6 is a warning of the overlarge roll angle.

Taking the DNN model as an example, the machine learning model is:

4-2, training the machine learning model by using the training set;

according to the network architecture, an adam optimizer is adopted, the learning rate is 0.01, the droupout rate is set to be 0.3, and training is carried out. And (5) adopting an early-stopping strategy, and stopping training when the test error is increased.

And 4-3, testing the machine learning model by using the test set.

Inputting test data of the trained model for an experiment, carrying out model hyper-parameter tuning according to a test result, and skipping to 4-1 until the performance meets the requirement.

And 5, outputting a ground proximity warning signal by using the machine learning model.

And 6, updating the strategy of the sample.

And when a new terrain data model is input or the value range of input data changes, randomly selecting related parameters, calculating to obtain a new sample, adding the new sample into a sample library, and executing the step 3 and the step 4.