CN115759498A - Unmanned aerial vehicle flight path real-time prediction method based on bidirectional long-term and short-term memory network - Google Patents
Unmanned aerial vehicle flight path real-time prediction method based on bidirectional long-term and short-term memory network Download PDFInfo
- Publication number
- CN115759498A CN115759498A CN202211387463.9A CN202211387463A CN115759498A CN 115759498 A CN115759498 A CN 115759498A CN 202211387463 A CN202211387463 A CN 202211387463A CN 115759498 A CN115759498 A CN 115759498A
- Authority
- CN
- China
- Prior art keywords
- model
- unmanned aerial
- prediction
- aerial vehicle
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Position Fixing By Use Of Radio Waves (AREA)
Abstract
The invention provides an unmanned aerial vehicle flight path real-time prediction method based on a bidirectional long-short term memory network, which comprises the following steps: the method comprises the following steps of S1, collecting a GPS positioning data sequence of a plurality of times of autonomous operation unmanned aerial vehicles flying along the same path, and performing data preprocessing by using Bessel-based geodetic coordinate conversion and least square fitting to generate a model training data set; s2, constructing a bidirectional long-short term memory path prediction model, and training the path prediction model by using the model training data set; and S3, predicting the unmanned aerial vehicle operation flight path in real time by using the trained path prediction model and combining with a prediction model compensator based on a PID (proportion integration differentiation) principle. The unmanned aerial vehicle flight path prediction model is superior to other traditional neural network models in performance and has higher prediction accuracy.
Description
Technical Field
The invention relates to the field of unmanned aerial vehicle path monitoring, in particular to an unmanned aerial vehicle flight path real-time prediction method based on a bidirectional long-short term memory network.
Background
With the rapid development of the autonomous control technology of the unmanned aerial vehicle, unmanned operation of the unmanned aerial vehicle in the fields of power inspection, traffic monitoring, target tracking, military striking and the like has become practical in recent years. Use unmanned aerial vehicle to carry out autonomic operation not only can practice thrift a large amount of human costs, compare with traditional artificial operation mode moreover, unmanned aerial vehicle of automatic flight can carry out the operation at bigger within range because of not receiving the restriction of remote control distance, possess higher control speed and precision simultaneously. For planned or periodically repeated work tasks such as data acquisition, crop spraying, etc., it is often necessary to design a flight path for the drone according to task requirements and work environment information. For example, planning a path that enables the drone to traverse a selected area or avoid an obstacle. An important premise for realizing autonomous control of the unmanned aerial vehicle is to acquire navigation information. For the above types of tasks, the most widely used navigation method is to acquire real-time geographic coordinates of the unmanned aerial vehicle by using a global positioning system, a Beidou satellite system, a Groness satellite system and other satellite positioning systems, and compare the real-time geographic coordinates with coordinates of a target point to determine the flight direction and speed of the unmanned aerial vehicle.
However, when the unmanned aerial vehicle actually operates, due to the influence of external disturbance factors such as wind power and electromagnetic interference, a flight control error or mistake may be generated, and a deviation exists between an actual flight path and a preset planned path. Due to lack of human intervention, the unmanned aerial vehicle cannot accurately judge potential abnormal operation conditions, so that a task fails and even serious consequences are caused due to the fact that a target point is missed because a control signal is not adjusted in time. In order to ensure the operation safety, the unmanned aerial vehicle has the capability of monitoring and correcting the flight path, and the capability plays an important guarantee role in smoothly completing the autonomous operation task of the unmanned aerial vehicle. Meanwhile, when the unmanned aerial vehicle operates in a relatively complex scene such as a city and a forest, the satellite positioning signal may have a relatively large error or cannot be normally received. How to obtain accurate positioning information in this situation is also an urgent problem to be solved.
The unmanned aerial vehicle autonomous operation monitoring means which is widely used at present and has a good effect reasonably predicts the position information of the unmanned aerial vehicle in a future period of time by taking the actual flight path of the unmanned aerial vehicle as a basis. And if the error between the predicted value and the set target point exceeds a certain threshold value, judging that the unmanned aerial vehicle works abnormally.
The neural network can extract the association existing among data through the powerful data feature learning capacity, and the characteristic enables the neural network to be widely applied to the motion trail prediction problem, and has remarkable superiority in prediction effect and performance. Most prediction models are based on classical neural networks such as multilayer sensing, recurrent neural networks, long-short term memory networks and convolutional neural networks. A constraint long-term and short-term memory network for flight trajectory prediction is provided in the prior art, constraints of different stages are provided according to the dynamic characteristics of an airplane, and a model can keep long-term dependence through dynamic physical constraints. Although the research on the problem of predicting the motion path of an object is mature, the research on the prediction of the flight path of the unmanned aerial vehicle by using the deep neural network is less limited by the motion mutability and complexity of the unmanned aerial vehicle.
Disclosure of Invention
In order to overcome the problems, the invention aims to provide a method for predicting the flight path of the unmanned aerial vehicle in real time based on a bidirectional long-short term memory network, so that the prediction accuracy is improved.
The invention is realized by adopting the following scheme: a method for predicting the flight path of an unmanned aerial vehicle in real time based on a bidirectional long-short term memory network comprises the following steps:
s1, collecting a GPS positioning data sequence of a plurality of times of autonomous operation unmanned aerial vehicles flying along the same path, and performing data preprocessing by using Bessel-based geodetic coordinate conversion and least square fitting to generate a model training data set;
s2, constructing a bidirectional long-short term memory path prediction model, and training the path prediction model by using the model training data set;
and S3, predicting the unmanned aerial vehicle operation flight path in real time by using the trained path prediction model and combining with a prediction model compensator based on a PID (proportion integration differentiation) principle.
Further, the step S1 is further specifically: step S11, forming an original flight path data set D = { D } used by a training model by using j unmanned aerial vehicle operation complete paths 1 ，D 2 ，D i ...，D j In which D is i A coordinate sampling point sequence which represents the ith path in time sequence; each GPS sample point includes data in three dimensions, namely latitude, longitude, and elevation; further, step S12, the position of any point P on the earth is represented by coordinates (B, L, H) in a geographic coordinate system, where B and L represent latitude and longitude, and H represents elevation; coordinates of the drone in the geographic coordinate system obtained when using GPS positioning (B) el ，L el ，H el ) Represents; coordinates (B) of the unmanned aerial vehicle in a geographic coordinate system by using Bessel geodetic coordinate conversion formula el ，L el ，H el ) Converting into coordinates (x) in meter under navigation coordinate system el ，y el ，z el ) To unify the dimensional data units of the three-dimensional position coordinates, a data set D ' = { D ' is obtained in terms of the navigation coordinate system coordinates ' 1 ，D′ 2 ，...，D′ j }; wherein, bessel geodetic coordinate conversion formula is:
wherein a is el 、b el Respectively the long axis and the short axis of the earth; b is el ，L el ，H el Latitude, longitude and elevation of the unmanned aerial vehicle in a geographic coordinate system;
let e 0 The specific expression of the eccentricity ratio of the earth is as follows:
then
Step S13, dividing D ' into D ' according to the set proportion ' Train ＝{D′ 1 ，D′ 2 ，...，D′ q And D' Validation ＝{D′ q+1 ，D′ q+2 ，...，D′ j Q < j) to generate a training set D 'for model training' Train And a verification set D 'for verifying model performance' Validation ；
Step S14, assuming that the path prediction model uses the latest m position coordinates of the unmanned aerial vehicle flight to estimate the next n position coordinates, namely the model input is [ P ] k-m+1 ，...，P k-1 ，P k ]Output is [ P ] k+1 ，P k+2 ，...，P k+n ]，P i (x i ，y i ，z i ) The coordinate of the frame i GPS positioning information received by the unmanned aerial vehicle is represented by coordinates obtained by Bessel conversion, and the position coordinate of the unmanned aerial vehicle corresponding to the frame k at the current moment is P k (ii) a To obtain tensor data in the same form as the model inputs and outputs, a sliding window of size 3 × (m + n) is set to traverse the sequence of coordinate points for each path, longL of sequence D' i L-m-n +1 matrices of the same size as the sliding windows are obtained, and further a model input matrix of 3 × m shape and a model output matrix of 1 × 3n shape are obtained, D' Train And D' Validation The matrices generated by the inner sequence constitute a training data set and a validation data set respectively. Further, when a GPS module carried by the unmanned aerial vehicle cannot be timely and normally received due to interference or communication delay, the model input sequence lacks data or the positioning accuracy is poor, the obtained positioning information has errors which affect the model prediction accuracy, and the least square fitting mode is adopted to supplement and correct the position coordinate observation data; taking the balance between the fitting error and the calculation time into consideration, and taking a quadratic function as a fitting target function; meanwhile, a calibration threshold value epsilon is set, a difference value between a position coordinate observation value obtained by a positioning module and a fitting objective function value is obtained, and when the difference value is larger than epsilon, an error exists in the observation value, so that the fitting objective function value is used for replacing the observation value, and an abnormal value of the data can be corrected to a certain extent;
normalizing the dimensional data according to the following formula:
wherein v is i Representing the ith data of a dimension, v max And v min For the maximum and minimum values of the data in this dimension,
to normalized data; normalization can ensure that the contribution of the characteristics of each dimension to the prediction result is the same, so that the accuracy of model prediction is improved.
Further, the step S2 is further specifically: the bidirectional long and short term memory network consists of a BilSTM layer, a Dropout layer, a full connection layer and an activation layer, wherein the BilSTM layer is used for extracting correlation characteristics among data of different time steps of a path time sequence; the Dropout layer can remove part of network units according to a certain probability to reduce the dependency among different units during each training, so that overfitting of a model to path data used for training is prevented, and the generalization capability of a flight path under different operation conditions is improved; the full connection layer integrates the extracted sequence characteristics, and the activation layer completes the nonlinear mapping from the characteristics to the prediction result; a modified linear unit is used as an activation function in an activation layer, so that the convergence speed and the calculation speed of the model are accelerated;
the training of the prediction model is realized by adopting an error back propagation algorithm, training data are input into the model according to the set batch size, a predicted value of the model is obtained through forward calculation, the predicted value is used for calculating a loss function value together with a data true value, on the basis, a network weight parameter is updated according to the gradient size obtained by the error back propagation algorithm, and the training data set is trained in a circulating mode until the set training period is completed; the training optimizer adopts an Adam optimizer, the mean square error is used as a loss function of the model, and the loss function formula is as follows:
where DL is the data set size, y i For the actual value of the data,
is a predicted value of the data. Further, the step S3 is further specifically: a prediction model compensator based on a proportional-integral-derivative error control theory is arranged according to the output characteristics of the bidirectional long-short term memory network and is used for providing a compensation value for a model prediction result; in order to obtain a prediction error, the output result of the model each time needs to be recorded and compared with the observed value received subsequently; setting a model prediction period to be the same as a GPS signal sampling period, namely performing prediction once each frame of positioning information is acquired; the model estimates the next n position coordinates using the most recently acquired m position coordinates, so that the model receives coordinates P from the drone during operation m Starting prediction from the reception of the coordinates P m+n Time switchThe predicted value and the true value can be compared to obtain the predicted error at each time step; by using
Representing coordinate points P obtained by the prediction model i An initial predicted value; when the ground station receives the coordinates P k When (k is more than or equal to m + n), calculating the predicted values of the latest m coordinates obtained by the prediction model
With its true value [ P k-n+1 ,...,P k-1 ,P k ]Difference e between k (ii) a The compensation value CV of the model prediction result k Is expressed as:
CV k ＝K P ·e k +K I ·(e k +e k-1 )·δ+K D ·(e k -e k-1 )/δ
in the formula K P 、K I And K D Proportional, integral and differential coefficients, respectively; see CV of k Consists of three parts; the first part is the proportionality coefficient K P And error e k For generating a reference compensation value; the second part is the integral coefficient K I And the product of the error accumulated value in a certain time is used for eliminating the steady-state error of the prediction model; and the last part is the differential coefficient K D And the product of the error change rate of the last two predictions, and adjusting the compensation value according to the error change to avoid too large compensation amplitude, which plays an important role in accelerating the response speed of the prediction model compensator under the condition that the error information has delay; adding the model output result and the compensation value to obtain a final coordinate prediction sequence
These coordinate prediction sequences are the unmanned aerial vehicle flight path.
The invention has the beneficial effects that: aiming at the problem of flight safety monitoring during autonomous operation of an unmanned aerial vehicle, the invention provides an error compensation BilSTM network path prediction model based on Bessel geodetic coordinate transformation, wherein a monitoring system is deployed at a ground station with higher hardware computing power; by using a Bessel geodetic coordinate conversion formula, coordinate conversion is carried out on latitude, longitude and elevation positioning information acquired by a GPS into navigation coordinate system coordinates with unified units, which is beneficial to extraction of correlation characteristics among all dimensional data by a neural network; when data is processed, a least square fitting method is adopted, and the problems of data missing and large errors are solved. On the basis, a Bessel-BilSTM network and a prediction model compensator based on a PID principle are designed, the output of the Bessel-BilSTM network is used as a reference value of a prediction result, and the compensator corrects the model prediction result according to an observed prediction error, so that a coordinate prediction sequence with higher precision is obtained. The experimental result obtained based on the data set consisting of the actual flight paths of the unmanned aerial vehicle operation shows that the prediction model provided by the invention can realize accurate prediction on the flight paths of the unmanned aerial vehicle within a certain time under the condition of only using GPS positioning information.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention.
Fig. 2 is a schematic diagram of the position relationship of the unmanned aerial vehicle in the geographic coordinate system and the navigation coordinate system.
Fig. 3 is a diagram of a framework of a drone operation monitoring system in an embodiment of the present invention.
FIG. 4 is a block diagram of a bidirectional long short term memory network according to an embodiment of the present invention.
FIG. 5 is a schematic diagram of the process of training the prediction model of the present invention using a back propagation algorithm.
Fig. 6 is a schematic flow chart of the present invention combined with a prediction model compensator based on the PID principle to predict the flight path of the unmanned aerial vehicle operation in real time.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
Referring to fig. 1, a method for predicting a flight path of an unmanned aerial vehicle in real time based on a bidirectional long-short term memory network includes the following steps:
s1, collecting a GPS positioning data sequence of a plurality of times of autonomous operation unmanned aerial vehicles flying along the same path, and performing data preprocessing by using Bessel-based geodetic coordinate conversion and least square fitting to generate a model training data set;
the step S1 is further specifically: step S11, forming an original flight path data set D = { D ] used by training models through j unmanned aerial vehicle operation complete paths 1 ，D 2 ，D i ...，D j In which D is i A coordinate sampling point sequence which represents the ith path in time sequence; each GPS sample point includes data in three dimensions, namely latitude, longitude, and elevation;
s12, expressing the position of any point P on the earth by using coordinates (B, L and H) in a geographic coordinate system, wherein B and L represent latitude and longitude, and H represents elevation; coordinates of the drone in the geographic coordinate system obtained when using GPS positioning (B) el ，L el ，H el ) Represents; coordinates (B) of the unmanned aerial vehicle in a geographic coordinate system by using Bessel geodetic coordinate conversion formula el ，L el ，H el ) Converting into coordinates (x) in meter under navigation coordinate system el ，y el ，z el ) To unify the dimensional data units of the three-dimensional position coordinates, a data set D '= { D' 1 ，D′ 2 ，...，D′ j }; wherein, bessel geodetic coordinate conversion formula is:
as shown in FIG. 2, wherein a el 、b el Respectively the long axis and the short axis of the earth; b is el ，L el， H el Latitude, longitude and elevation of the unmanned aerial vehicle in a geographic coordinate system;
let e 0 The specific expression of the eccentricity ratio of the earth is as follows:
then
Step S13, dividing D ' into D ' according to the set proportion ' Train ＝{D′ 1 ，D′ 2 ，...，D′ q And D' Validation ＝{D′ q+1 ，D′ q+2 ，...，D′ j Q < j) to generate a training set D 'for model training' Train And a verification set D 'for verifying model performance' Validation (ii) a When the model is used for coordinate prediction, the ground station sets a window with a corresponding size to select the positioning information of the latest sampling moments, and tensor data subjected to data fitting supplement and correction, bessel conversion and normalization processing is used as prediction model input. It is worth noting that the ground station needs to update the maximum value and the minimum value of each dimension of data in real time according to newly acquired position observation information, and the data normalization result can be ensured to be located in the interval (0, 1).
Step S14, assuming that the path prediction model uses the latest m position coordinates of the unmanned aerial vehicle flight to estimate the next n position coordinates, namely the model input is [ P ] k-m+1 ，...，P k-1 ，P k ]Output is [ P ] k+1 ，P k+2 ，...，P k+n ]，P i (x i ，y i ，z i ) The coordinate of the frame i GPS positioning information received by the unmanned aerial vehicle is represented by the coordinate obtained by Bessel conversion, and the position coordinate of the unmanned aerial vehicle corresponding to the frame k at the current moment is P k (ii) a To obtain tensor data of the same form as the model inputs and outputs, a sliding window of size 3 x (m + n) is set to traverse the sequence of coordinate points of each path, the sequence of length l D' i L-m-n +1 matrices of the same size as the sliding windows are obtained, and further a model input matrix of 3 × m shape and a model output matrix of 1 × 3n shape are obtained, D' Train And D' Validation The matrices generated by the inner sequence constitute a training data set and a validation data set respectively.
Due to the fact that the operation capability of the unmanned aerial vehicle on-board processor is limited, the calculation capability requirement of the deep learning task cannot be met, and path prediction of the unmanned aerial vehicle is completed at the ground station. The ground station deploys a database storing planned paths and flight data of the drones, and a path prediction model. The unmanned aerial vehicle operation monitoring system frame of this patent design is shown in figure 3. The process of drone path prediction may be described as: the unmanned aerial vehicle sends the received positioning information to the ground station through the data transmission module, the ground station sequences the coordinate points according to the time stamps of the data frames to obtain an unmanned aerial vehicle flight path observed value, and the sequence is used as model input data to obtain an unmanned aerial vehicle position predicted value.
Further, when a GPS module carried by the unmanned aerial vehicle cannot be normally received in time due to interference or communication delay, the model input sequence lacks data or the positioning accuracy is poor, and the obtained positioning information has errors which influence the prediction accuracy of the model, and the position coordinate observation data is supplemented and corrected by adopting a least square fitting mode; taking the balance between the fitting error and the calculation time into consideration, and taking a quadratic function as a fitting target function; meanwhile, a calibration threshold value epsilon is set, a difference value between a position coordinate observation value obtained by an unmanned aerial vehicle positioning module and a fitting objective function value is obtained, and when the difference value is larger than the epsilon, an error exists in the observation value, so that the fitting objective function value is used for replacing the observation value, and an abnormal value of the data can be corrected to a certain extent;
normalizing the dimensional data according to the following formula:
wherein v is i Representing the ith data of a dimension, v max And v min For the maximum and minimum values of the data in this dimension,
to normalized data; the normalization can ensure that the characteristics of each dimension make the same contribution to the prediction result so as to improve the accuracy of model prediction.
S2, constructing a bidirectional long-short term memory path prediction model, and training the path prediction model by using the model training data set;
as shown in fig. 4, the step S2 further specifically includes: the bidirectional long-short term memory network consists of a BilSTM layer, a Dropout layer, a full connection layer and an activation layer, wherein the BilSTM layer is used for extracting correlation characteristics among data at different time steps of a path time sequence; the Dropout layer can remove part of network units according to a certain probability to reduce the dependency among different units during each training, so that overfitting of a model to path data used for training is prevented, and the generalization capability of a flight path under different operation conditions is improved; the full connection layer integrates the extracted sequence characteristics, and the activation layer completes the nonlinear mapping from the characteristics to the prediction result; the modified Linear Unit is used as an activation function (ReLU) in the activation layer, so that the convergence speed and the calculation speed of the model can be accelerated;
the ReLU function expression is:
f(x)＝max(0,x)
the training of the prediction model is realized by adopting an error back propagation algorithm, training data are input into the model according to the set batch size, a predicted value of the model is obtained through forward calculation, the predicted value is used for calculating a loss function value together with a real data value, on the basis, a network weight parameter is updated according to the gradient size obtained through the error back propagation algorithm, as shown in fig. 5, the training data set is trained in a circulating mode (the mode is that the training data are input into the model according to the set batch size, the predicted value of the model is obtained through forward calculation, the predicted value is used for calculating the loss function value together with the real data value, and on the basis, the network weight parameter is updated according to the gradient size obtained through the error back propagation algorithm) until the set training period is completed; the training optimizer adopts an Adam optimizer, the mean square error is used as a loss function of the model, and the loss function formula is as follows:
where DL is the data set size, y i For the true value of the data,
is a predicted value of the data.
And S3, predicting the unmanned aerial vehicle operation flight path in real time by using the trained path prediction model and combining with a prediction model compensator based on a PID (proportion integration differentiation) principle.
As shown in fig. 6, the step S3 further specifically includes: a prediction model compensator based on a proportional-integral-derivative error control theory is arranged according to the output characteristics of the bidirectional long-short term memory network and is used for providing a compensation value for a model prediction result; in order to obtain a prediction error, the output result of the model each time needs to be recorded and compared with the observed value received subsequently; setting a model prediction period to be the same as a GPS signal sampling period, namely performing prediction once each frame of positioning information is obtained; the model uses the most recently acquired m position coordinates to estimate the next n position coordinates, so that the model receives the coordinates P from the receiver during operation of the drone m Starting prediction from the reception of the coordinates P m+n The predicted value and the real value can be compared to obtain the predicted error at each time step; by using
Representing coordinate points P obtained by the prediction model i An initial predicted value; when the ground station receives the coordinates P k When k is more than or equal to m + n, calculating the predicted values of the latest m coordinates obtained by the prediction model
With its true value [ P k-n+1 ,...,P k-1 ,P k ]Difference e between k (ii) a The compensation value CV of the model prediction result k Is expressed as:
CV k ＝K P ·e k +K I ·(e k +e k-1 )·δ+K D ·(e k -e k-1 )/δ
in the formula K P 、K I And K D Proportional, integral and differential coefficients, respectively; see CV of k Consists of three parts; the first part is the proportionality coefficient K P And error e k For generating a reference compensation value; the second part is the integral coefficient K I And the product of the error accumulated value in a certain time is used for eliminating the steady-state error of the prediction model; and the last part is the differential coefficient K D And the product of the error change rate of the last two predictions, and adjusting the compensation value according to the error change to avoid too large compensation amplitude, which plays an important role in accelerating the response speed of the prediction model compensator under the condition that the error information has delay; adding the model output result and the compensation value to obtain a final coordinate prediction sequence
The coordinate prediction sequences are the flight paths of the unmanned aerial vehicles. During operation, the unmanned aerial vehicle acquires positioning information at a sampling frequency delta.
In a word, the invention provides an error compensation BilSTM network path prediction model based on Bessel geodetic coordinate transformation aiming at the problem of flight safety monitoring during autonomous operation of an unmanned aerial vehicle, and a monitoring system is deployed at a ground station with higher hardware computing power; the latitude, longitude and elevation positioning information acquired by the GPS is subjected to coordinate conversion into a navigation coordinate system coordinate with a unified unit by using a Bessel geodetic coordinate conversion formula, so that the extraction of the correlation characteristics among all the dimensional data by a neural network is facilitated; when data is processed, a least square fitting method is adopted, and the problems of data missing and large errors are solved. On the basis, a Bessel-BilSTM network and a prediction model compensator based on a PID principle are designed, the output of the Bessel-BilSTM network is used as a reference value of a prediction result, and the compensator corrects the model prediction result according to an observed prediction error, so that a coordinate prediction sequence with higher precision is obtained. The experimental result obtained based on the data set consisting of the actual flight paths of the unmanned aerial vehicle shows that the prediction model provided by the invention can realize accurate prediction of the flight paths of the unmanned aerial vehicle within a certain time under the condition of only using GPS positioning information.
The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.
Claims (5)
1. An unmanned aerial vehicle flight path real-time prediction method based on a bidirectional long-short term memory network is characterized by comprising the following steps: the method comprises the following steps:
s1, collecting a GPS positioning data sequence of a plurality of times of autonomous operation unmanned aerial vehicles flying along the same path, and performing data preprocessing by using Bessel-based geodetic coordinate conversion and least square fitting to generate a model training data set;
s2, constructing a bidirectional long-short term memory path prediction model, and training the path prediction model by using the model training data set;
and S3, predicting the unmanned aerial vehicle operation flight path in real time by using the trained path prediction model and combining with a prediction model compensator based on a PID (proportion integration differentiation) principle.
2. The method for predicting the flight path of the unmanned aerial vehicle in real time based on the bidirectional long-short term memory network as claimed in claim 1, wherein: the step S1 is further specifically: step S11, forming an original flight path data set D = { D ] used by training models through j unmanned aerial vehicle operation complete paths 1 ,D 2 ,D i ...,D j In which D is i A coordinate sampling point sequence which represents the ith path in time sequence; each GPS sample point includes data in three dimensions, namely latitude, longitude, and elevation;
s12, expressing the position of any point P on the earth by using coordinates (B, L and H) in a geographic coordinate system, wherein B and L represent latitude and longitude, and H represents elevation; coordinates of the drone in the geographic coordinate system obtained when using GPS positioning (B) el ,L el ,H el ) Represents; coordinates (B) of the unmanned aerial vehicle in a geographic coordinate system by using Bessel geodetic coordinate conversion formula el ,L el ,H el ) Conversion to navigational coordinate systemCoordinates in meters (x) el ,y el ,z el ) To unify the dimensional data units of the three-dimensional position coordinates, a data set D ' = { D ' is obtained in terms of the navigation coordinate system coordinates ' 1 ,D′ 2 ,...,D′ j }; wherein, bessel geodetic coordinate conversion formula is:
wherein a is e1 、b e1 The long axis and the short axis of the earth are respectively long; b el ,L el ,H el Latitude, longitude and elevation of the unmanned aerial vehicle under a geographic coordinate system;
let e 0 The specific expression is as follows:
then
Step S13, dividing D ' into D ' according to the set proportion ' Train ＝{D′ 1 ,D′ 2 ,...,D′ q And D' Validation ＝{D′ q+1 ,D′ q+2 ,...,D′ j Q < j) to generate a training set D 'for model training' Train And a verification set D 'for verifying model performance' Validation ；
Step S14, assuming that the path prediction model uses the latest m position coordinates of the unmanned aerial vehicle flight to estimate the next n position coordinates, namely the model input is [ P k-m+1 ,...,P k-1 ,P k ]Output is [ P ] k+1 ,P k+2 ,...,P k+n ]，P i (x i ,y i ,z i ) The coordinate obtained by converting the GPS positioning information of the frame i received by the unmanned aerial vehicle through Bessel is represented, and the current moment is opposite to the coordinateThe position coordinate of the unmanned aerial vehicle in the kth frame is P k (ii) a To obtain tensor data of the same form as the model inputs and outputs, a sliding window of size 3 x (m + n) is set to traverse the sequence of coordinate points of each path, the sequence of length l D' i L-m-n +1 matrices having the same size as the sliding window are obtained, and further a model input matrix having a shape of 3m and a model output matrix having a shape of 1 × 3n are obtained, from D' Train And D' Validation The matrices generated by the inner sequence constitute a training data set and a validation data set respectively.
3. The method for predicting the flight path of the unmanned aerial vehicle in real time based on the bidirectional long and short term memory network as claimed in claim 2, wherein: when a GPS module carried by an unmanned aerial vehicle is interfered or communication delay cannot be timely and normally received, the model input sequence lacks data or the positioning accuracy is poor, the obtained positioning information has errors which affect the model prediction accuracy, and the least square fitting mode is adopted to supplement and correct the position coordinate observation data; taking a quadratic function as a fitting objective function in consideration of balance between fitting error and calculation time; meanwhile, a calibration threshold value epsilon is set, a difference value between a position coordinate observation value obtained by a positioning module and a fitting objective function value is obtained, and when the difference value is larger than epsilon, an error exists in the observation value, so that the fitting objective function value is used for replacing the observation value, and an abnormal value of the data can be corrected to a certain extent;
normalizing the dimensional data according to the following formula:
wherein v is i Representing the ith data of a dimension, v max And v min For the maximum and minimum values of the data in this dimension,
to normalized data; normalization can guarantee each dimensionThe contribution of the characteristics of (2) to the prediction result is the same, so that the accuracy of model prediction is improved.
4. The method for predicting the flight path of the unmanned aerial vehicle in real time based on the bidirectional long and short term memory network as claimed in claim 1, wherein: the step S2 is further specifically: the bidirectional long and short term memory network consists of a BilSTM layer, a Dropout layer, a full connection layer and an activation layer, wherein the BilSTM layer is used for extracting correlation characteristics among data of different time steps of a path time sequence; the Dropout layer can remove part of network units according to a certain probability to reduce the dependency among different units during each training, so that overfitting of a model to path data used for training is prevented, and the generalization capability of a flight path under different operation conditions is improved; the full connection layer integrates the extracted sequence characteristics, and the activation layer completes the nonlinear mapping from the characteristics to the prediction result; a modified linear unit is used as an activation function in an activation layer, so that the convergence speed and the calculation speed of the model are accelerated;
the training of the prediction model is realized by adopting an error back propagation algorithm, training data are input into the model according to the set batch size, a predicted value of the model is obtained through forward calculation, the predicted value is used for calculating a loss function value together with a data true value, on the basis, a network weight parameter is updated according to the gradient size obtained by the error back propagation algorithm, and the training data set is trained in a circulating mode until the set training period is completed; the Adam optimizer is adopted in the training optimizer, the mean square error is used as a loss function of the model, and the loss function formula is as follows:
where DL is the data set size, y i For the actual value of the data,
is a predicted value of the data.
5. The real-time prediction method for the flight path of the unmanned aerial vehicle based on the bidirectional long-short term memory network as claimed in claim 3, wherein: the step S3 is further specifically: a prediction model compensator based on a proportional-integral-derivative error control theory is arranged according to the output characteristics of the bidirectional long-short term memory network and is used for providing a compensation value for a model prediction result; in order to obtain a prediction error, the output result of the model each time needs to be recorded and compared with the observed value received subsequently; setting a model prediction period to be the same as a GPS signal sampling period, namely performing prediction once each frame of positioning information is obtained; the model uses the most recently acquired m position coordinates to estimate the next n position coordinates, so that the model receives the coordinates P from the receiver during operation of the drone m Starting prediction from the received coordinates P m+n The predicted value and the true value can be compared to obtain the predicted error in each time step; by using
Indicating the coordinate point P obtained by the prediction model i An initial predicted value; when the ground station receives the coordinates P k When (k is more than or equal to m + n), calculating the predicted values of the latest m coordinates obtained by the prediction model
With its true value [ P k-n+1 ,...,P k-1 ,P k ]Difference e between k (ii) a The compensation value CV of the model prediction result k Is expressed as:
CV k ＝K P ·e k +K I ·(e k +e k-1 )·δ+K D ·(e k -e k-1 )/δ
in the formula K P 、K I And K D Proportional, integral and differential coefficients, respectively; see CV of k Consists of three parts; the first part is the proportionality coefficient K P And error e k For generating a reference compensation value; the second part is the integral coefficient K I And within a certain timeThe product of the error accumulated values is used for eliminating the steady-state error of the prediction model; and the last part is the differential coefficient K D And the product of the error change rate of the last two predictions, and adjusting the compensation value according to the error change to avoid too large compensation amplitude, which plays an important role in accelerating the response speed of the prediction model compensator under the condition that the error information has delay; adding the model output result and the compensation value to obtain a final coordinate prediction sequence
The coordinate prediction sequences are the flight paths of the unmanned aerial vehicles.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211387463.9A CN115759498A (en) | 2022-11-07 | 2022-11-07 | Unmanned aerial vehicle flight path real-time prediction method based on bidirectional long-term and short-term memory network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211387463.9A CN115759498A (en) | 2022-11-07 | 2022-11-07 | Unmanned aerial vehicle flight path real-time prediction method based on bidirectional long-term and short-term memory network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115759498A true CN115759498A (en) | 2023-03-07 |
Family
ID=85357246
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211387463.9A Pending CN115759498A (en) | 2022-11-07 | 2022-11-07 | Unmanned aerial vehicle flight path real-time prediction method based on bidirectional long-term and short-term memory network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115759498A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116912320A (en) * | 2023-09-13 | 2023-10-20 | 齐鲁空天信息研究院 | Positioning method and device of object elevation coordinate, electronic equipment and medium |
-
2022
- 2022-11-07 CN CN202211387463.9A patent/CN115759498A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116912320A (en) * | 2023-09-13 | 2023-10-20 | 齐鲁空天信息研究院 | Positioning method and device of object elevation coordinate, electronic equipment and medium |
| CN116912320B (en) * | 2023-09-13 | 2023-12-22 | 齐鲁空天信息研究院 | Positioning method and device of object elevation coordinate, electronic equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109781099B (en) | Navigation method and system of self-adaptive UKF algorithm | |
| Chen et al. | A hybrid prediction method for bridging GPS outages in high-precision POS application | |
| CN103528587B (en) | Independent combined navigation system | |
| CN114048889B (en) | Aircraft trajectory prediction method based on long-term and short-term memory network | |
| CN102508493B (en) | Flight control method for small unmanned aerial vehicle | |
| Doostdar et al. | INS/GNSS integration using recurrent fuzzy wavelet neural networks | |
| Zhang et al. | A hybrid intelligent algorithm DGP-MLP for GNSS/INS integration during GNSS outages | |
| Peng et al. | UAV positioning based on multi-sensor fusion | |
| CN104635743A (en) | Minimum altitude full-course autonomous flight control system of high-speed unmanned aerial vehicle | |
| CN113190036A (en) | Unmanned aerial vehicle flight trajectory prediction method based on LSTM neural network | |
| CN112083457B (en) | Neural network optimized IMM satellite positioning navigation method | |
| CN110346821A (en) | A kind of SINS/GPS integrated attitude determination localization method solving the problems, such as GPS long-time losing lock and system | |
| CN115759498A (en) | Unmanned aerial vehicle flight path real-time prediction method based on bidirectional long-term and short-term memory network | |
| CN114216459B (en) | ELM-assisted GNSS/INS combined navigation unmanned target vehicle positioning method | |
| Chen et al. | Real-time unmanned aerial vehicle flight path prediction using a bi-directional long short-term memory network with error compensation | |
| Belhajem et al. | A robust low cost approach for real time car positioning in a smart city using Extended Kalman Filter and evolutionary machine learning | |
| CN116086445A (en) | Multi-source information time delay fusion navigation method based on factor graph optimization | |
| CN117390498B (en) | Flight capability assessment method of fixed wing cluster unmanned aerial vehicle based on Transformer model | |
| Neusypin et al. | Investigation into the nonlinear Kalman filter to correct the INS/GNSS integrated navigation system | |
| CN113126647B (en) | Collaborative guidance method based on leader and follower principle | |
| CN109799522B (en) | Multi-view-field multi-target automatic rapid polling method | |
| CN116663384A (en) | Target track prediction method under battlefield task planning background | |
| CN115639585A (en) | Multi-sensor fusion positioning method for GPS/IMU and laser radar | |
| CN114862912A (en) | Fixed wing target tracking method based on Bezier curve secondary optimization | |
| CN115014321A (en) | Bionic polarization multi-source fusion orientation method based on adaptive robust filtering |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |