CN110309909A - A kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed - Google Patents
A kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed Download PDFInfo
- Publication number
- CN110309909A CN110309909A CN201910562518.7A CN201910562518A CN110309909A CN 110309909 A CN110309909 A CN 110309909A CN 201910562518 A CN201910562518 A CN 201910562518A CN 110309909 A CN110309909 A CN 110309909A
- Authority
- CN
- China
- Prior art keywords
- neuron
- output
- target
- high speed
- wide range
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
A kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed, first proposed learning sample method for building up;Then the study of the target characteristics of motion and training mechanism based on improved BP are constructed;Finally by Single-step Prediction and rolling forecast method, empty day moving-target high speed a wide range of the intelligent, quick of motor-driven track, Accurate Prediction are realized;The present invention only needs to know the history and the position data at current time of empty day moving-target, motion model without target, the convergence rate of traditional BP neural network is improved by design factor of momentum and using variable step iterative strategy simultaneously, reduces the oscillation in convergence process, greatly improve the precision of trajectory predictions, it may be directly applied to all kinds of high speeds, the trajectory predictions problem of highly maneuvering target, with stronger applicability, theoretical basis and technological reserve are provided for tasks such as the monitoring of the hypersonic aircrafts such as X-37B, tracking, interceptions to be subsequent.
Description
Technical field
The present invention relates to a kind of intelligent real-time predicting methods of a wide range of maneuvering target track of high speed, belong to Spacecraft Control
Field.
Background technique
The prediction of high speed grand movement target trajectory is to realize that empty day moving-target intercepts, tracks and persistently observe important
It ensures.With the development of space technology, the movement velocity of spacecraft and acceleration maneuverability are gradually promoted, and are that moving-target is pre-
Survey brings great challenge.On the one hand, the complicated multiplicity of target forms of motion and unknown, prior information deficiency, it is difficult to pre- in advance
It surveys, need to predict the motor behavior of target by proposing the intelligent algorithm with strong learning ability;On the other hand, target maneuver energy
Power is strong, the resources critical constraints such as calculating, storage on star, need to be by the trajectory predictions method of proposition low complex degree, to moving target
Path implementation is quick and precisely predicted.
Since the motor pattern of a wide range of maneuvering target of high speed is complicated, its dynamics and kinematics model is made to be difficult to accurately obtain
It takes, the prediction techniques based on model such as traditional Kalman filtering, regression forecasting, strong tracking algorithm is caused to be difficult to play a role.
For this problem, in recent years, it is concerned in theoretical and engineering field, interactive multi-model prediction algorithm.By reasonably
The dynamics and kinematics model library for nonstandard of maneuvering target are constructed, so that multiple model concurrent workings, by information fusion to movement mesh
The prediction of target path implementation.But the method needs to include maneuvering target model as comprehensive as possible, leads to its computationally intensive, number
It is complicated according to processing, it is difficult to which that a wide range of maneuvering target track of high speed is predicted in real time.In order to reduce the complexity of calculating, usually
By the way of least square fitting, using the historical data of maneuvering target, by constructing using time and position as input and output
The multinomial of variable predicts maneuvering target in the position of future time instance.But this method limitation is very big, if target trajectory
It is unsatisfactory for the form (for example, movement of being at the uniform velocity diversion in the planes) of time polynomial, then the error of its fitting and prediction is larger.
In consideration of it, in order to intercepted to empty day moving-target (specifically including that aircraft, guided missile, hypersonic aircraft etc.) and with
Track provides reliable, complete prior information, and makes up deficiency existing for existing theoretical method, and the invention patent is effective for lacking
The a wide range of maneuvering target track of high speed quickly, the status of Accurate Prediction method, propose a kind of intelligent real-time track prediction technique.
Using target histories time-position data as learning sample, training sample model is established, gives BP neural network construction
With training mechanism, empty day moving-target trajectory predictions problem is made to be converted into the study forecasting problem of neural network.By introducing momentum
The factor reduces the oscillation trend in learning process, improves constringency performance, solves convergence rate and stability using step length changing method
Between contradiction, and give prediction error correcting method.
Summary of the invention
Technology of the invention solves the problems, such as: overcome the deficiencies in the prior art, proposes a kind of intelligent real-time track prediction
Method, it is only necessary to the history and the position data at current time of empty day moving-target be instructed to pass through improvement without the motion model of target
BP neural network carries out the dynamic optimization of gradient decline mode to the error surface of cost function, moment product before fully considering
Tired experience introduces factor of momentum and error revising strategies, effectively increases the intelligent level and precision of trajectory predictions, expands
The application range and application value of trajectory predictions algorithms, for it is subsequent for the monitoring of the hypersonic aircrafts such as X-37B, tracking,
The tasks such as interception provide theoretical basis and technological reserve.
The technical solution of the invention is as follows: a kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed,
It comprises the following steps that
Step 1: building BP neural network model;
Building includes three layers of BP neural network model of input layer, hidden layer and output layer;There are two classes in the model
Signal: function signal and error signal;Wherein, function signal is carried out just from input layer to output layer in the form of output function
To propagation, one layer of neuron under the influence of each layer of neuron;Error signal be in the form of error function from output layer to
Input layer carries out backpropagation;
Step 2: building learning sample;
The method of learning sample building is as follows:
It will past t0~tmThe sampled data of moment corresponding m+1 group moving-target position is enabled as initial learning sampleWithRespectively as input learning sample and output
Learning sample, and input learning sample and output learning sample are normalized;
Wherein, tm=t0+ mT, T represent the spaceborne computer sampling period;M is positive integer, indicates to adopt for m-th that be counted
The sample period;tξIndicate sampling instant, x (tξ) indicate tξMoment corresponding moving-target position;ξ=1,2,3 ..., m;t0Indicate initial
Sampling instant, x (t0) indicate t0Moment corresponding moving-target position;
Step 3: training BP neural network model;Specific step is as follows:
(1) cost function ε (n) is established:
ek(n)=dk(n)-yk(n),
Wherein: l is the number of output layer neuron;K=1,2,3 ..., l indicates k-th of neuron of output layer;ek
It (n) is the reality output y of k-th of output neuronk(n) with desired output dk(n) error function;N indicates the n-th step iteration, is
Positive integer;
(2) the weighed value adjusting strategy Δ w of hidden layer neuron j to output layer neuron k is constructedkj(n) are as follows:
In formula: η (n) is learning rate, and its calculation formula is η (n)=2γη (n-1), the calculation formula of exponent gamma be γ=
sgn[ε(n-1)ε(n)];δkIt (n) is partial gradient, its calculation formula is
Indicate output layer neuron k activation primitiveDerivative;vk
It (n) is induction local field, its calculation formula iswkjIt (n) is hidden layer neuron j to output layer
The weight of neuron k;yj(n) reality output of hidden layer neuron j is indicated;αkjFor hidden layer neuron j to output layer nerve
The factor of momentum of first k;J is positive integer;
(3) the weighed value adjusting strategy Δ w of input layer i to hidden layer neuron j is constructedji(n) are as follows:
In formula:Indicate hidden layer neuron j activation primitiveDerivative, yi(n) input layer is indicated
The reality output of neuron i, αjiFor input layer i to the factor of momentum of hidden layer neuron j, i is positive integer;
Step 4: by the time-position coordinates (t of a sampling instant in targetm,x(tm)) input as neural network,
Predict the lower sampling instant t of targetm+1Position x (tm+1), and be normalized, complete Single-step Prediction;
Step 5: renewal learning sample is
And step 3 is repeated, complete rolling forecast.
Compared with the prior art, the invention has the advantages that:
(1) the invention proposes a kind of intelligent Forecastings for quick moving-target track complicated and changeable, with existing skill
Art is compared, and the method for the present invention is no longer required for target movement model it is known that without estimating or predicting target movement model, it is only necessary to know
The position data at road target histories and current time, this greatly reduces the data quick-processing pressure of spaceborne computer, is avoided
Influence of the model error to prediction result, greatly improved the precision of existing trajectory predictions method, extends existing method
Use scope.
(2) the invention proposes the learning sample construction method of BP neural network fast convergence and roll more new strategy, with
Traditional BP neural network is compared, and the method for the present invention has comprehensively considered precision of prediction and predicted the influence of rapidity, and by reasonable
It constructs learning sample and establishes training mechanism and improve the instruction of network under the premise of guaranteeing precision of prediction, promoting predetermined speed
Practice speed, provides theoretical foundation for high speed, the trajectory predictions of highly maneuvering target.
Detailed description of the invention
Fig. 1 is the three-dimensional physical location and predicted position figure of empty day moving-target;
Fig. 2 is the X-axis physical location and predicted position figure of empty day moving-target;
Fig. 3 is the Y-axis physical location and predicted position figure of empty day moving-target;
Fig. 4 is the Z axis physical location and predicted position figure of empty day moving-target;
Fig. 5 is the trajectory predictions deviation map of empty day moving-target;
Fig. 6 is the schematic diagram of BP neural network model;
Fig. 7 is that the neuron k of output layer is connected to the signal flow diagram of hidden layer neuron j.
Specific embodiment
A specific embodiment of the invention is further described in detail with reference to the accompanying drawing.
A kind of the step of intelligent real-time predicting method of a wide range of maneuvering target track of high speed of the present invention, is as follows:
Step 1: building BP neural network model
Building includes three layers of BP neural network model of input layer, hidden layer and output layer as shown in Figure 6,7.In the model
It is middle that there are two class signals, i.e. function signal and error signal;Wherein, function signal be in the form of output function from input layer to
Output layer carries out forward-propagating, one layer of neuron under the influence of each layer of neuron;Error signal is the shape with error function
Formula carries out backpropagation from output layer to input layer.
Step 2: building learning sample;
The method of learning sample building is as follows:
It will past t0~tmThe sampled data of moment corresponding m+1 group moving-target position is enabled as initial learning sampleWithRespectively as input learning sample and defeated
Learning sample out, and input learning sample and output learning sample are normalized;
Wherein, tm=t0+ mT, T represent the spaceborne computer sampling period;M is positive integer, indicates to adopt for m-th that be counted
The sample period;tξIndicate sampling instant, x (tξ) indicate tξMoment corresponding moving-target position;ξ=1,2,3 ..., m;t0Indicate initial
Sampling instant, x (t0) indicate t0Moment corresponding moving-target position;
Step 3: training BP neural network model;Specific step is as follows:
(1) cost function ε (n) is established:
ek(n)=dk(n)-yk(n),
Wherein: l is the number of output layer neuron;K=1,2,3 ..., l indicates k-th of neuron of output layer;ek
It (n) is the reality output y of k-th of output neuronk(n) with desired output dk(n) error function;N indicates the n-th step iteration, is
Positive integer;
(2) the weighed value adjusting strategy Δ w of hidden layer neuron j to output layer neuron k is constructedkj(n) are as follows:
In formula: η (n) is learning rate, and its calculation formula is η (n)=2γη (n-1), the calculation formula of exponent gamma be γ=
sgn[ε(n-1)ε(n)];δkIt (n) is partial gradient, its calculation formula is
Indicate output layer neuron k activation primitiveDerivative;vk
It (n) is induction local field, its calculation formula iswkjIt (n) is hidden layer neuron j to output layer
The weight of neuron k;yj(n) reality output of hidden layer neuron j is indicated;αkjFor hidden layer neuron j to output layer nerve
The factor of momentum of first k;J is positive integer;
(3) the weighed value adjusting strategy Δ w of input layer i to hidden layer neuron j is constructedji(n) are as follows:
In formula:Indicate hidden layer neuron j activation primitiveDerivative, yi(n) input layer is indicated
The reality output of neuron i, αjiFor input layer i to the factor of momentum of hidden layer neuron j, i is positive integer;
Step 4: by the time-position coordinates (t of a sampling instant in targetm,x(tm)) input as neural network,
Predict the lower sampling instant t of targetm+1Position x (tm+1), and be normalized, complete Single-step Prediction;
Step 5: renewal learning sample is
And step 3 is repeated, complete rolling forecast.
Embodiment:
Fig. 1 is three-dimensional ballistic missile (a wide range of maneuvering target of high speed) flight path using STK Software Create;Wherein,
The sampling time is 1s in emulation, is carried out curve fitting using the historical data of preceding 100s, without prediction.
Using the movement objective orbit intelligence real-time predicting method proposed by the present invention based on improved BP, obtain
X-axis physical location and predicted position, Y-axis physical location and predicted position, the Z axis actual bit of the moving-target of sky day shown in Fig. 2~Fig. 5
It sets and predicted position and trajectory predictions deviation.It can be seen that the obtained prediction locus of the method for the present invention from Fig. 2~Fig. 5
It coincide substantially with the real trace of a wide range of maneuvering target of high speed, the prediction deviation maximum in X-axis is no more than 50km, in Y-axis
On prediction deviation maximum be no more than 100km, prediction deviation maximum on Z axis is no more than 60km.
Specific embodiment can be seen that proposed by the invention a kind of based on improved BP through the invention
Empty day moving-target track intelligence method for quick predicting, overcome existing method real-time is poor, precision of prediction is low, do not consider target spy
The defect of different maneuver model, it is theoretical based on intelligent Neural Network, give the learning strategy and rail of empty day moving-target motor pattern
Mark prediction technique effectively extends the scope of application of existing trajectory predictions method, and having broad application prospects makes with high
With value, all kinds of high speeds, the trajectory predictions problem of highly maneuvering target, real-time rolling forecast sky day moving-target may be directly applied to
The position of future time instance, and prediction error is modified, there is stronger applicability, be directed to the contour ultrasound of X-37B to be subsequent
The tasks such as fast aircraft monitors, tracking, interception provide theoretical basis and technological reserve.
The content that description in the present invention is not described in detail belongs to the well-known technique of professional and technical personnel in the field.
Claims (7)
1. a kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed, which is characterized in that comprise the following steps that
Step 1: building BP neural network model;
Step 2: building learning sample;The method of learning sample building is as follows:
It will past t0~tmThe sampled data of moment corresponding m+1 group moving-target position is enabled as initial learning sampleWithRespectively as input learning sample and defeated
Learning sample out, and input learning sample and output learning sample are normalized;
Wherein, tm=t0+ mT, T represent the spaceborne computer sampling period;M is positive integer, indicates m-th of the sampling to be counted week
Phase;tξIndicate sampling instant, x (tξ) indicate tξMoment corresponding moving-target position;ξ=1,2,3 ..., m;t0Indicate initial samples
Moment, x (t0) indicate t0Moment corresponding moving-target position.
Step 3: training BP neural network model;
Step 4: by the time-position coordinates (t of a sampling instant in targetm,x(tm)) input as neural network, prediction
The lower sampling instant t of targetm+1Position x (tm+1), and be normalized, complete Single-step Prediction;
Step 5: renewal learning sample is
And step 3 is repeated, complete rolling forecast.
2. a kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed according to claim 1, feature
It is, the specific method is as follows for step 1:
Building includes three layers of BP neural network model of input layer, hidden layer and output layer;There are two class signals in the model:
Function signal and error signal;Wherein, function signal is to carry out positive biography from input layer to output layer in the form of output function
It broadcasts, one layer of neuron under the influence of each layer of neuron;Error signal be in the form of error function from output layer to input
Layer carries out backpropagation.
3. a kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed according to claim 1 or 2, special
Sign is that specific step is as follows for step 3:
Step 3.1 establishes cost function ε (n);
Wherein, n indicates the n-th step iteration, is positive integer;
Step 3.2, the weighed value adjusting strategy Δ w for constructing hidden layer neuron j to output layer neuron kkj(n);K=1,2,
3 ..., l, l are the number of output layer neuron;J is positive integer;
Step 3.3, the weighed value adjusting strategy Δ w for constructing input layer i to hidden layer neuron jji(n)。
4. a kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed according to claim 3, feature
It is, in step 3.1:
ek(n)=dk(n)-yk(n),
Wherein: k-th of neuron of k expression output layer;ekIt (n) is the reality output y of k-th of output neuronk(n) and it is expected
Export dk(n) error function.
5. a kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed according to claim 4, feature
It is, in step 3.2, the weighed value adjusting strategy Δ w of hidden layer neuron j to output layer neuron kkj(n) are as follows:
In formula: η (n) is learning rate, and its calculation formula is η (n)=2γη (n-1), the calculation formula of exponent gamma are γ=sgn [ε
(n-1)ε(n)];δkIt (n) is partial gradient, its calculation formula is
Indicate output layer neuron k activation primitiveDerivative;vkIt (n) is induction
Local field, its calculation formula iswkjIt (n) is hidden layer neuron j to output layer neuron k's
Weight;yj(n) reality output of hidden layer neuron j is indicated;αkjFor hidden layer neuron j to the momentum of output layer neuron k
The factor.
6. a kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed according to claim 5, feature
It is, in step 3.3, the weighed value adjusting strategy Δ w of input layer i to hidden layer neuron jji(n) are as follows:
In formula:Indicate hidden layer neuron j activation primitiveDerivative, yi(n) input layer is indicated
The reality output of i, αjiFor input layer i to the factor of momentum of hidden layer neuron j, i is positive integer.
7. a kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed according to claim 6, feature
It is, in step 3.3,
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910562518.7A CN110309909A (en) | 2019-06-26 | 2019-06-26 | A kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910562518.7A CN110309909A (en) | 2019-06-26 | 2019-06-26 | A kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110309909A true CN110309909A (en) | 2019-10-08 |
Family
ID=68076300
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910562518.7A Pending CN110309909A (en) | 2019-06-26 | 2019-06-26 | A kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110309909A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111399534A (en) * | 2020-02-25 | 2020-07-10 | 清华大学 | Method and system for capturing aerial medium-high speed moving targets by multiple unmanned aerial vehicles |
CN111931287A (en) * | 2020-07-06 | 2020-11-13 | 北京电子工程总体研究所 | Near space hypersonic target trajectory prediction method |
CN112348223A (en) * | 2020-08-21 | 2021-02-09 | 哈尔滨工业大学 | Missile flight trajectory prediction method based on deep learning |
CN112612001A (en) * | 2020-11-30 | 2021-04-06 | 天津光电通信技术有限公司 | Track prediction method and device based on BP neural network algorithm |
-
2019
- 2019-06-26 CN CN201910562518.7A patent/CN110309909A/en active Pending
Non-Patent Citations (2)
Title |
---|
杨晓晓: "高速运动物体轨迹预测的研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
钱夔等: "基于BP神经网络的空中目标航迹预测模型", 《指挥信息系统与技术》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111399534A (en) * | 2020-02-25 | 2020-07-10 | 清华大学 | Method and system for capturing aerial medium-high speed moving targets by multiple unmanned aerial vehicles |
CN111399534B (en) * | 2020-02-25 | 2021-09-17 | 清华大学 | Method and system for capturing aerial medium-high speed moving targets by multiple unmanned aerial vehicles |
CN111931287A (en) * | 2020-07-06 | 2020-11-13 | 北京电子工程总体研究所 | Near space hypersonic target trajectory prediction method |
CN111931287B (en) * | 2020-07-06 | 2023-02-24 | 北京电子工程总体研究所 | Near space hypersonic target trajectory prediction method |
CN112348223A (en) * | 2020-08-21 | 2021-02-09 | 哈尔滨工业大学 | Missile flight trajectory prediction method based on deep learning |
CN112612001A (en) * | 2020-11-30 | 2021-04-06 | 天津光电通信技术有限公司 | Track prediction method and device based on BP neural network algorithm |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110309909A (en) | A kind of intelligent real-time predicting method of a wide range of maneuvering target track of high speed | |
CN111351488B (en) | Intelligent trajectory reconstruction reentry guidance method for aircraft | |
Zeng et al. | A deep learning approach for aircraft trajectory prediction in terminal airspace | |
CN109343341A (en) | It is a kind of based on deeply study carrier rocket vertically recycle intelligent control method | |
CN112114521B (en) | Intelligent prediction control entry guidance method for spacecraft | |
CN114048889A (en) | Aircraft trajectory prediction method based on long-term and short-term memory network | |
CN111176263B (en) | Online identification method for thrust fault of aircraft based on BP neural network | |
CN111258302B (en) | Aircraft thrust fault online identification method based on LSTM neural network | |
CN112947592B (en) | Reentry vehicle trajectory planning method based on reinforcement learning | |
CN111813146B (en) | Reentry prediction-correction guidance method based on BP neural network prediction voyage | |
Ma et al. | Deep reinforcement learning of UAV tracking control under wind disturbances environments | |
Dong et al. | Deep learning in aircraft design, dynamics, and control: Review and prospects | |
CN113126643A (en) | Intelligent robust reentry guidance method and system for hypersonic aircraft | |
Yang et al. | LSTM-based deep learning model for civil aircraft position and attitude prediction approach | |
Xie et al. | Dual-channel and bidirectional neural network for hypersonic glide vehicle trajectory prediction | |
CN112348223A (en) | Missile flight trajectory prediction method based on deep learning | |
CN114370793A (en) | Rocket sublevel return and vertical landing guidance method | |
CN114037050B (en) | Robot degradation environment obstacle avoidance method based on internal plasticity of pulse neural network | |
CN106569503A (en) | Triggering-type predictive control-based aircraft control method | |
CN114415730B (en) | Intelligent planning method for escape trajectory of spacecraft | |
CN117289709A (en) | High-ultrasonic-speed appearance-changing aircraft attitude control method based on deep reinforcement learning | |
Wang et al. | Flight trajectory prediction of general aviation aircraft based on lstm model | |
Fan et al. | Aircraft trajectory prediction based on residual recurrent neural networks | |
Xie et al. | A novel migrant PSO algorithm for vehicle reentry trajectory optimization | |
CN113406957A (en) | Mobile robot autonomous navigation method based on immune deep reinforcement learning |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191008 |
|
RJ01 | Rejection of invention patent application after publication |