realtime implementation of neural network augmented fault tolerant flight controllers for an advanced fighter aircraft on a target digital signal processor.

by:INDUSTRIAL-MAN     2019-09-20
Neural networks in recent years (NNs)
Linear and Nonlinear recognition and control have been proposed.
Linear dynamic system]5,6and 7].
For the applicability of NN to adaptive control systems, the following features are important: * applicability to non-adaptive control systemsLinear system]8]
* Parallel distributed processing and hardware implementation * Learning (either on-line or off-line)
* For multiple
In the control architecture, the attraction of neural networks lies in the fact that they are likely to learn the dynamics of the controlled process.
However, when neural networks are used for real-time control of dynamic processes, such as automatic
The landing problem, compared to the evolution time of the process being considered, is that the training time is too long.
In real-time control applications, another crucial problem is that no matter which adaptive algorithm is finally adopted to adjust the weight in the neural network, the stability of the control process must be guaranteed.
Through this process of weight change, the network \"learns\" how to control the dynamic process.
Through software simulation, the architecture of BTFC and EMRAN has been successfully verified.
However, despite the encouraging ability, the scientific community successfully validated and demonstrated NN-
Through the implementation on the appropriate hardware platform, based on the control and estimation scheme in the actual FCS.
This is due to different reasons.
One of the reasons is simple;
Most gain scheduling-
The control rules in flight control systems designed based on classical control and estimation linear theory are well demonstrated and provide satisfactory performance for many years.
This is true for most autonomous and enhanced stability systems for commercial aviation, however, these systems are not designed to provide fault tolerance.
Another reason is the lack of verification and verification tools for NN
Estimation scheme based on control.
Finally, another reason associated with the implementation issue is.
The computational workload associated with obtaining a single or multiple parallel NNs with a large number of neurons and/or complex activation functions can be very large.
This paper introduces the research results focusing on laser tissue.
The hardware implementation of classical and neural networks is given in this study.
Assist the Classic Controller to enhance the fault tolerance of high-performance fighter planes during the landing phase when subject to strong winds and failures (such as stuck control surfaces.
The control scheme adopts radial basis function network (RBFN)
Feedback error learning mechanism.
The dynamic RBFN used is only in-
Online learning, no training required.
The controller is unable to obtain information about the actuator failure for use in the reconfigure.
In particular, this effort demonstrates the feasibility of implementing a complex online learning NNflight controller scheme with two control structures with different concepts on a digital signal processor.
Aircraft mathematical model, automatic
The aircraft model used in this study is an aircraft model of a high-performance fighter [1]1].
For the purposes of this study, using cfd calculations, the pneumatic data of the lift and aileron control surfaces are split into two parts that correspond to the left and right surfaces.
The aerodynamic model also contains the ground effect model.
There are two elevators on the plane (-
25 ° to 25 ° deflection)
, Can lament together, can also lament in differential mode.
It also has a pair of wings (-
20 ~ 20 ° deflection)and a rudder (-
Deflection from 30 to 30).
Engine model (
No dynamic)
Finish six-
Degree of free simulation.
There is a hydraulic actuator on the aircraft, which drives the main control surface of the model with the first-order lag, and the atime constant is 50 milliseconds.
The rate limit of the actuator is set to 60 degrees/second. The auto-Login Problem [2]
What was studied in this study included a high-performance fighter that flew along a flight path consisting of non-light segments such as wings
Fly horizontally, coordinate turns, slide down, and finally the torch maneuver and landing.
As shown in Figure 2, the trajectory segment corresponding to these stages must fly in the presence of strong winds.
These will cause the aircraft to deviate from the specified trajectory.
However, we must make sure that all trajectory deviations are within the specified range.
The touch condition is given in tighspecs, named for easy touch
The lower medicine box as shown in Table 1.
The controller is first designed to meet the pill box specifications for all these stages under the condition that the actuator is trouble-free.
We then add the controller in order to be able to handle the sameflight segment, but with the emergence of certain fault conditions for the actuator.
The plane has three degrees of freedom, namely push-ups and Roland horizontal pendulum.
The height of the aircraft is controlled by the control surface called the elevator.
They are located at the rear of the aircraft and are placed symmetrically on either side of the aircraft axis.
Both elevators have the same control signal under normal flight conditions.
If one of the elevators fails or is stuck somewhere, the other can be used to restore the balance of the aircraft.
Therefore, the elevator is responsible for the longitudinal dynamics of the aircraft.
The rolling motion of the aircraft is controlled by the surface called the aileron.
The two wings are present on the wing and usually work under the differential mode.
If one of the wings fails, the other can be used to stabilize the aircraft.
The movement of the aircraft is controlled by the rudder located at the tail of the aircraft.
The mechanism of rudder operation is similar to that of the ship\'s rudder operation.
Therefore, the lateral dynamics of the aircraft are controlled by the wing and rudder.
In this study, we considered five types of faults as shown in Table 2, including single control surface faults and those of the control surface combination.
We ignore the situation where both elevators fail because it is often unrecoverable.
As mentioned earlier, both elevators are always directed together under normal flight conditions.
However, they can also be used to generate rolling moments when they are used differently.
The test of the input matrix shows that in the differential mode, the efficiency of the elevator in generating the rolling moment is about 60% of the aileron.
Therefore, the elevator can be used to generate pitching and rolling moments.
This is contrary to the aileron that is not valid when any pitch moment is generated.
Therefore, we can consider the case of the aileron failure. by using the elevator in differential mode, we can return the aircraft to equilibrium.
The overall scheme of the controller architecture neural controller is shown in Figure 1.
The landing task is autonomous, so a navigation function called \"tracking command generator\" is included in the block. [
Figure 1 slightly][
Figure 2:[
Figure 3 slightly]
The output of this block is referenced by the command (
Mark as \"r\" in the figure \")
, Input to the baseline track tracking controller (BTFC)
The picture is called \"classic feedback controller \".
Under normal circumstances, BTFC is designed to make the aircraft output \"y\" follow the reference vector \"r \".
Neural controller (EMRAN)
Use the reference signal and the aircraft output to generate its instruction signal.
It also uses the output of BTFC to learn the reactionary mechanics of plants (
In this case the plane)
In the feedback error learning scheme9].
Looking back at BTFC, first consider BTFC designed using classic Loop shaping SISO design technology for hardware implementation. As reportedin [3]
Let\'s say there is no feedback using the angle of attack and the side slip.
The reference command generator or the tracking controller determines the offset of the aircraft on the required ground track for each segment of the flight, and calculates the reference command consisting of height, the speed and the cross distance with the design trajectory and the angle error between the aircraft speed and the design trajectory vector.
The segment of the track is either a straight line or an arc of a circle.
Therefore, in the case of the online segment, the cross distance is only the vertical length.
In the case of a circular arc, this number is the difference between the distance to the center of the circular arc and the radius of the turn.
Similarly, the angle error can be calculated using the component of the aircraft speed in X-
The direction of the desired trajectory closest to the aircraft.
Once these quantities are known, velocity and acceleration are obtained by linear interpolation between the endpoint values at the end of each segment.
The review of EMRAN in some cases the aircraft must fly at a large angle of attack.
The control law of this maneuver is very bad. linear. This non-
Linear due to non
Linear aerodynamics and non-
Inertia Coupling in this flight state.
These control laws are based on the dynamics of motion and the inversion of the equations of motion.
In this inversion process, state variables are measured, and then these variables are used to simulate the force and moment corresponding to the unwanted aerodynamic, gravity, or inertial contribution.
Here we deal with nonlinear dynamic inverse (NDI)
Fault-tolerant controller for control surface failure [2]. The non-
The linear inverse function that appears in the NDI controller is online approached with a gradient basis function neural network.
The function approximation algorithm based on RBFNN is called the extended minimum resource allocation network.
The algorithm is based on the feedback error learning strategy proposed by Gomi and Kawato [9].
The first neural architecture considered by the hardware implementation is EMRAN.
This is a quick implementation in [reported MRAN asreported]2].
Unlike MRAN, the EMRAN controller presented here uses previous estimates from the control strategy to improve the fault tolerance of btfc design.
Only the parameters of the nearest neuron are updated.
In all other ways, the EMRAN update algorithm is similar to MRAN.
Section 3 discusses the hardware implementation of the Matlab/Simulink model of the neural network-assisted flight controller. 1 and 3.
2 Real-time implementation by porting these models to the target processor using the real-time workshop.
Also co-decomposition studio (CCS)
Used to assemble the code and load it on the target processor TMS 6713.
Simulink is a software package for dynamic system modeling, simulation and analysis.
It supports linear and nonlinear systems modeled in a continuous time, sampling time, or a mixture of both.
The system can also be multi-rate. e.
, There are different parts that have been adjusted or Updated at different prices. Real-
Real time Workshop-
Time Workshop builds applications from the Simulink chart for prototyping, testing, and deployment in real Time
Time Systems on various target computing platforms. Users of Real-
Time Workshop can direct it to generate source code to accommodate compilers, input and output devices, memory models, communication modes, and other features that their applications may need. Real-
Time Workshop is an extension of the functions of Simulink and matlab.
It automatically generates the package and compiles the source code from the Simulink model to create the real
Time software applications on various systems.
By providing a code generation environment for rapid prototyping and deployment
The basis of the production code generation capability is the time workshop.
Along with other tools and components
Time Workshop provides: * Automatic code generation tailored for various target platforms.
* Fast and direct path from system design to implementation.
* Seamless integration with MATLAB and Simulink.
* Simple graphical user interface.
* Open architecture and scalable production processes.
With the DSP platform of TI company as the embedded target, the tools of simulink and MATLAB are integrated.
The software collection allows us to design digital signal processing from concept to code development and verification.
The embedded target of the TI C6000 DSP consists of the TI C6000 target, which automatically makes rapid prototyping on our C6000 hardware target.
Target C code generated using Real-
Time workshops and development tools build executable files for our target processor. The Real-
The Time workshop build process loads the target machine code onto our motherboard and runs the executable on the digital signal processor.
Besides, a real
Time Workshop build options build a Code Composer Studio project from Real-generated C codeTimeWorkshop.
All the features provided by Code Composer Studio (CCS)
Such as editing, building, debugging, code analysis and project management tools to help us use MATLAB, Simulink, Real-
Hardware we support.
When we use this target, the build process creates a new project in Code Composer Studio and fills the project with the files that the project needs. Pre-
A necessary condition for a hardware implementation is a thorough inspection of the model files to determine if there are any blocks in it that are continuous.
If they are found they will be replaced by an equivalent discrete block.
In addition, the sampling time is set to constant value 0.
In this design is 02 s.
Using the Simulink library browser, continuous blocks are replaced by corresponding discrete blocks that retain the specified initial condition.
The next steps involved in the hardware implementation process are: * S-
Create Mexico (
MATLAB executable filefor S-
Function packaging * automatic S-
The function wrapper generates the build process and then compiles and links the model _ sf. c with model.
Another real one
Code module generated by embedded encoder in time Workshop, building a Mexico-file. The MEX-
The file name is model _ sf. mexext. (
Mexico ext is the file extension of Mexico-
The files on our platform are given by the matlab mexext command. )The MEX-
The files are stored in our working directory. Finally, Real-
Time Workshop creates and opens a header-free model with the generated S-Function block.
The following restrictions apply to ert s-
Function Summary :(i)
Continuous time is not supported when ERTS is generated-
Function packaging.
Support continuous time option does not apply to generation of ert s-
Function wrapper. (ii)
Impossible to create real multiple instances
Time workshop embedded encoder generating S-
Because the code uses static memory allocation, the function blocks in the model. [
Figure 4a omitted][
Figure 4b omitted]
Results and Discussion Figure 4a describes the trajectory of the aircraft obtained from Type I fault simulation in the XY plane (
2 degrees left elevator stuck).
The trajectory stays in (0, 0)
Landed successfully on the XY platform.
Figure 4b drawn with values of X and Y co-is the same
Coordinates obtained from hardware for the same fault (
2 degrees left elevator stuck).
It can be seen that the two XY diagrams are the same.
Therefore, for this iteration of elevator failure, the results of simulation and hardware are consistent. [
Figure 5a omitted][
Figure 5b omitted]
Contrast of left elevator fault tolerance package line-
Type I fault of BTFC/EMRAN [
Figure 6a omitted][
Figure 6b omitted]
Contrast of left elevator fault tolerance package line-
Type ii failure of BTFC/EMRAN [
Figure 7a omitted][
Figure 4b
Comparison of fault tolerance of left elevator fault aileron-
Type III failure of BTFC/EMRAN [
Figure 8a omitted][
Figure 8b omitted]
Elevator contrast on the left-
Right wing failure tolerance envelope-
Type IV failure of BTFC/EMRAN [
Figure 9a][
Figure 9b slightly]
Left wing contrast-
Right wing failure tolerance envelope-
The V-type fault of the BTFC/EMRAN hardware implementation results are completely consistent with the classic and neural network enhanced classical controllers, as shown in Figure 5-9.
But due to the accuracy and accuracy of dsp tms 6713 for real-time implementation, there are some missing points on the map.
For the entire fault scenario of the two flight controllers, the same result is obtained.
The same situation appears in the comparison of fault-tolerant envelopes for BTFC and EMRAN, as shown in Figure 5-9.
The author would like to thank and thank the professor. N.
Sundararajan, School of Electrical and Electronic Engineering, Nanyang University of Science and Technology, Singapore, allows the use of aircraft models for this work. References [1]L. T. Nguyen, M. E. Ogburn, W. P. Gilbert, K. S. Kibler, P. W. Brown, P. L.
Research on Booth/post-transaction and simulator
The flameout characteristic of a fighter with relaxation longitudinal static stability.
NASATechnical paper 1538, December. 1979. [2]Y. Li, N.
Sundararajan and P.
Saratchandran, \'robustneuro-
Design of H8 controller for automatic aircraft driving
Fixed \', ieee Transactions in aerospace and electrical subsystems40, No. 1, January2004. [3]A. A. Pashilkar, N.
Sundararajan and P.
Saratchandran \"AFault-
Fault-tolerant neural auxiliary controller for automatic aircraft driving
Journal of Aerospace Science and Technology (Elsevier), Vol.
10, Issue1, January 2006, page. 49-61. [4]A. A. Pashilkar, N.
Sundararajan and P.
Saratchandran \"AFault-
Fault-tolerant neural auxiliary controller for automatic aircraft driving
Journal of Aerospace Science and Technology (Elsevier), Vol.
10, Issue1, January 2006, page. 49-61. [5]A. A.
Pashilkar, vijia. R, et.
Al, improved fault tolerance for cars
Landing after using adaptive
Stepping neural controllers at 2007 IEEE International Conference on multi-system and control in Singapore. [6]Hunt, K. J. , Sbardato, D.
And the Gawthrop page. J.
Neural Networks for control systems
An investigation, automica, Volume 28, No. 6, p. 1083, 1992 [7]Narendra, K. S.
K.
, \"Identifying and controlling dynamic systems using neural networks\", ieee Transactions on Neural Networks, Volume 11, No.
Economic and Social Council March 1990 [8]Cybenko, G.
, \"Superimpose approximation by SigmoidalFunctions\", Mathematics of Control Signals and Systems, Volume 2, Volume 4, page303-309, 1989 [9]H. Gomi and M.
Kawato, \"neural network control of communication
Cycle system using feedbackerror-
Learning, neural network, Volume 1. 6, No. 7, 1993, pp. 933-946. Dr.
Rao (Nagaraj)1), Dr.
Abei A Pashilkar (2), Dr. T. V. RamaMurthy (3)(1)
Director, Cognitive Technology Center, Department of Electronic and Communication EngineeringV.
Bangalore Institute of Engineering, India (2)
Scientist E2, head of flight simulation Department, National Aerospace Laboratory, Bangalore, India (3)
Professor and holder of Distance Education Department
Communication Engineering, University of Engineering, Bangalore, India
Email: rnrvcct @ gmail
Custom message
Chat Online 编辑模式下无法使用
Chat Online inputting...