[IEEE 2007 Mediterranean Conference on Control & Automation - Athens, Greece...

Abstract— For the autonomous flight, navigation, guidance and control of the EFIGENIA Unmanned Aerial Vehicle it is essential to have high performance 6-DOF attitude and heading reference system measurements. This paper presents the design and development of a real-time intelligent attitude and heading reference system I-AHRS, as in the hardware, as in the intelligent digital neural network software scheme, analysis, design and construction for the orientation calculation for the EFIGENIA EJ-1B MOZART and the EFIGENIA EJ-2B MARIA Autonomous Unmanned Aerial Vehicles UAVs. The EFIGENIA I-AHRS consists of three MEMS accelerometers, three MEMS rate-gyros and three magneto-resistive transducers that send its outputs to a digital neural network in which is possible to develop a strategy for attitude estimation. Additionally it is well known that the Kalman Filter is an option as multi-sensor data fusion and integration, however it has some adaptability limitations. In this paper, FPGA reconfigurable hardware digital neural network architecture is presented and utilized to replace the Kalman Filter in the integration of MEMS IMU inertial sensors signals and the Magneto resistive sensors.

I. INTRODUCTION he EFIGENIA EJ-1B MOZART and the EFIGENIA EJ-2B MARIA are a couple of very own designed autonomous unmanned aerial vehicles UAV, each one

developed for perform different mission tasks. EFIGENIA EJ-1B MOZART is a Short or Vertical take-Off and Landing (S/VTOL) UAV Unmanned airplane, which uses rotating blades in order to create forces necessary to Fly. This UAV is built of robust, lightweight, and high-strength materials. Its aerospace design introduces a Hybrid S/VTOL Rotor and Tailless Forward Swept Wing UAV Unmanned Airplane Concept with the purpose of allowing to the air vehicle an excellent aeromechanical behavior. This EFIGENIA EJ-1B uses a forward swept wing with a planform area of 254.143.30 sq mm. The flap geometry in this UAV airplane offer a flaperon characteristics allowing to the aerial vehicle to meet with the aerodynamic

Mario Andres Cordoba is Electronics Engineer and is Aeronautical Engineering Researcher with the EFIGENIA Aerospace Robotics UAV, Unmanned Aerial Vehicles Research, www.efigenia-aerospace.com, e-mail: [email protected]

characteristics and without producing excessive download in hover. The EJ-1B is powered by two 2-stroke piston engines located one on the tail fuselage and one more engine inside the aerial vehicle body. In contrast, the tail piston engine has been adapted to conforming a thrust vectoring unit to aimed high performance flight control system, maneuverability and agility at low speeds, Figure 1.

Figure 1. The EFIGENIA EJ-1B MOZART S/VTOL UAV autonomous Unmanned Aircraft.

The EFIGENIA EJ-2B MARIA is an Autonomous High Altitude Long Endurance Unmanned Aerial Vehicle UAV specifically designed for environmental and oceanographic Earth studies. This vehicle has its very own designed autonomous avionics autopilot system based on artificial intelligence Fuzzy Logic control techniques combined with Neural Networks. The aerodynamics design of EFIGENIA EJ-2B is the key of this unmanned airplane because of its configuration is optimized for fly at high altitude and fuel consumption efficiency. The EFIGENIA EJ-2B MARIA was designed to demonstrate feasibility of the low cost autonomous Unmanned Aerial Vehicle UAV development for high altitudes and long endurance missions. The vehicle is powered by two 2.0 HP engines located each one in the nose (tractor propeller configuration) and the other one at the tail fuselage (pusher prop configuration). The principal stages of the design were based on flight characteristics including the mission characteristics, and then the UAV mass estimation, fuel volume and consumption for

ATTITUDE AND HEADING REFERNCE SYSTEM I-AHRS FOR THE EFIGENIA AUTONOMOUS UNMANNED AERIAL

VEHICLES UAV BASED ON MEMS SENSOR AND A NEURAL NETWORK STRATEGY FOR ATTITUDE ESTIMATION

Mario Andrés Córdoba G.

T

2007 Mediterranean Conference on Control andAutomation, July 27 - 29, 2007, Athens - Greece

1-4244-1282-X/07/$25.00 ©2007 IEEE

T02-007

the mission, the altitude and speed of flight, pressure distributions, Lift and Drag estimations profiles for selected airfoil wing sections, wing planform geometry, propulsion and engine size, etc. Its aeronautical design offers an efficient configuration as the main wing is aerodynamically efficient and the fuselage is located between two “booms” which hold the twin conventional tail surfaces, Figure 2.

Figure 2. The EFIGENIA EJ-2B MARIA high altitude long endurance (HALE) Unmanned Aerial vehicle UAV. The flight operations of each one of the EFIGENIA UAV are specified by control station which because of its small physical dimensions could be resided onboard card, ship, airplane or ground. The control station continuously maintains communication with the airborne platform and payload. This has been designed to operating under concept of “virtual Cockpit” which allows to the operator the possibility of feeling the realism of flight operations, flight conditions and its performance. The vehicle have 6 DOF for controlling the EFIGENIA UAV attitude in flight using independent embedded fuzzy logic 16-bit DSP controller multiprocessor flight control computer in Roll, Pitch, and Yaw axis Figure 3.

Figure 3. Principal axes system for the EFIGENIA EJ-2B HALE UAV

EFIGENIA UAV aircraft was modeled as a rigid body moving in space, using the variables (x, y, z) to represent the position of the UAV vehicle in body coordinates, and the variables ψθφ ,, to represent the roll, pitch, and yaw angles of the EFIGENIA UAV with respect to the body coordinates. The combination of neural network and fuzzy logic system make possible to create an effective method for implementing the EFIGENIA autonomous navigation and flight control technique. In this way, the system allows a massive parallelism; learning ability, fault tolerance, etc., capabilities. This system is divided in two important subsystems: The Guidance and Navigation Computer and the Fuzzy Logic flight control system computer, Figure 4.

II. I-AHRS DESIGN PHILOSOPHY

A. System Overview

The EFIGENIA UAV aircrafts perform its autonomous flight based on an embedded Fuzzy Logic Fight Control System architecture. The EFIGENIA fuzzy logic flight control system drive the attitude of the vehicle, this means that the attitude information of the UAV vehicle must be measured every time. For this reason was designed a low-cost electronic Intelligent Attitude and Heading Measurement System I-AHRS, Figure 4.

Figure 4. EFIGENIA designed I-AHRS block diagram The electronic I-AHRS is composed by three important parts: the sensor board, the processing main board and the neural network attitude estimation scheme. The sensor board has been designed, tested and constructed based on three orthogonally mounted miniature MEMS gyros, three orthogonally mounted MEMS accelerometers and a three axis magnetometer along the body X, Y, Z axis.

Pitch

Roll

Yaw


T02-007

The processing main board is based on a reconfigurable programmable logic device FPGA which run the digital neural network and a dspic30F4013 digital signal processor as a system coprocessor.

B. EFIGENIA UAV Attitude Representation The attitude of the EFIGENIA UAV was defined using three consecutive rotations. This angular rotation was referred as the Euler Angles ( ), ,ψ θ φ which determined the attitude of the UAV with respect to a local level reference frame. The following rotations were applied:

• First, the body frame is rotated around the Z-axis by an angleψ .

• Second, the body frame is rotated around the Y-axis by an angleθ .

• Finally, the body frame is rotated around the X-axis by an angleφ .

Thus, the Euler angles are given by:

( )

( )

tan .sin .cos

.cos .sin

sec . .sin .cos

p q r

q r

q r

φ θ φ φ

θ φ φ

ψ θ φ φ

•

•

•

= + +

= −

= +

For the EFIGENIA UAV flight dynamics and mission profile the range of the Euler angles values ( ), ,ψ θ φ were limited with the purpose to avoid singularities from the trigonometric functions at certain angles. This case was modeled in Simulink using the Aerosim aeronautical toolbox as shown in Figure 5.

Figure 5. EFIGENIA UAV attitude representation and modeling using Euler angles.

III. ELECTRONIC HARDWARE DESIGN

The intelligent attitude heading and reference system I-AHRS electronic hardware design has been based on MEMS gyros and accelerometers and magneto-resistive transducers.

The MEMS transducer output an electrical signal (voltage) proportional to the physical variable sensed in the UAV, in this case, angular velocity and acceleration. Using the MEMS technology on those sensors, the hardware cost and the physical size of the circuits are reduced representing and advantage for the use at the avionics section of the EFIGENIA UAV. The I-AHRS is composed by three ADXRS150 gyros mounted orthogonally for measure the angular velocities { , ,p q r }, and three ADXL203 accelerometer transducers with it’s respectively axes { , ,X Y ZA A A } aligned with the EFIGENIA UAV aircraft axes. Finally the system add a 3-axis magneto-resistive transducer { , ,X Y ZH H H } for complete the data set information of the unit, Figure 6.

Figure 6. Sensitive axes for the accelerometers and gyros.

A. The Accelerometer Transducer The accelerometers are used in the electronic I-AHRS to measure tilt angles with respect to gravity, and as navigation system data inputs. The transducer used for acceleration measurement is a 1.7g± surface micromachined device. This offer low noise,

wide dynamic range, reduced power consumption and improved gzero − bias drift.

Each transducer is a two sensitive axes, orthogonal ( 90 ) to each other. The device has their sensitive axes in the same plane as the silicon chip, Figure 7.

Figure 7. I-AHRS accelerometer electronic circuit Typical noise density level intrinsic to the device is

110 gHzµ

RMS,

(1) (2) (3)


T02-007

and the device typical scale factor is 1000 mvg with a

sensitivity drift due to temperature of 0.3%± . Due to the nature of accelerometer transducer device construction, the output voltage ( outV ) is a function of both the acceleration input ( a ) and the power supply voltage ( sV ) as shows in Figure 8.

Figure 8. AHRS accelerometer transfer function These accelerometers, as inclinometer, measure an angle relative to the force of gravity, and have a too low band-width (slow response) to be able to control the attitude of the EFIGENIA UAV. Table II list the accelerometer transducer performance used in the I-AHRS circuit design.

B. The Angular Rate (Gyro) Transducer The I-AHRS designed use angular rate sensors to measure the rate of rotation about each axis. The outputs of each gyro are the analog voltage in response to its angular rate physical movement. The three gyros are mounted in orthogonal way that allow to measure the angular velocities p, q, r, around the accelerometer axes system, Figure 9.

Figure 9. I-AHRS angular rate sensor electronic circuit Three angular rate transducer units are required to sense all three UAV aircraft axis (Pitch, Roll, Yaw), and each reports it axis rotational sensing in the form of difference analog outputs. The angular rate offering by these gyros are acceptable in terms of band-width and are used to find the attitude angle by integration, over the time, of the sensor output data on each axis. At this point, the big problem is that due to the numerical integration process the measurements errors are accumulated causing drift on the angular measurement, and then is necessary to apply a correction procedure. The system has one angular rate sensor per axis, giving an analog voltage output that is proportional to the rotational turn-rate about each axis. Figure 10 represents the transfer function of the angular rate transducer.

Figure 10. I-AHRS angular rate transfer function The detailed specification for the angular rate transducers used in the EFIGENIA intelligent attitude and heading reference system is listed below in the Table III. The rate gyro data is sampled at 100 Hz, given a good response about the EFIGENIA UAV autonomous aircraft attitude over short periods of time. These sensors provide an output proportional to the rotation rate contaminated by noise and with the so called drift effect.

TABLE II ACCELEROMETER TRANSDUCER SPECIFICATIONS

Parameter X-Y-AXIS Z-AXIS

Range 1.7[ ]g± 1.7[ ]g± Sensitivity

1000 mVg

1000 mVg

Frequency Response 5.5 KHz 5.5 KHz Supply Voltage 5 V 5 V

Operating Temperature

-40 to 85 deg -40 to 85 deg

Noise Performance 110 ugHz

110 ug

Hz

Non-Linearity 0.5% of FS 0.5% of FS Alignment Error 1± 1±


T02-007

For the short periods of time the drift can be approximated by a constant bias, and the orientation can be evaluated with a single integration of the angular rate signal output.

The transfer equation returns the rotation angle with two added terms. A random walk due to the noise and other term that grows with time proportional to the gyro bias. The maximum random walk error is proportional to the noise standard deviation and the square root of time. This error has to be considered in case where this value is comparable with the expected drift of the gyro. Another aspect that needs to be considered is the resolution of the analog to digital converter A/D used. The gyro output data is connected to a 12-bit A/D, returning a 0 to 4096 numerical value that corresponds to the ratio of the input voltage (0 to 5V). The frequency response is critical to the proper system operation. In this way, it has been decided to implement a 5th order low-pass filter to filter the high frequency noise on the gyro output signal. As result of the integration procedure, the value of the angular rate sensor in steady position suffer of an offset drift action respect of its initial steady position value because of the integration over long periods of time. As a one solution for this problem has been developed a Neural Network scheme for estimate the EFIGENIA UAV attitude orientation in which are fused the 3-axes gyros, accelerometers and magnetometer sensors data

C. The Magneto-Resistive Transducers The magnetometer sensor is used to provide heading information. In this case, was used a 3-axis solid state magnetic transducer sensitive to the earth’s magnetic field, figure 13.

Figure 13. I-AHRS magnetometer circuit A digital signal processor DSP (dspic30F3013) transforms the 3-axis X, Y, Z measurements to a coordinate system calculating the components as follows:

2 2H HH X Y= + (4)

And then:

arctan H

H

YX

ψ

=

(5)

Table IV shows the Magneto-resistive transducer specifications for the EFIGENIA UAV I-AHRS SYSTEM.

The measured outputs of the magnetic transducer, X and Y, are plotted showing a circle as illustrated in Figure 14. Figure 14. I-AHRS Magnetic transducer outputs X vs Y

-200 -150 -100 -50 0 50 100 150 200-200

-150

-100

-50

0

50

100

150

200

Xh Magnetic Field

Yh

Mag

netic

Fie

ld

EFIGENIA I-AHRS Magnetic Sensors

TABLE III ANGULAR RATE TRANSDUCER ESPECIFICATIONS


Range 150 secdec ± 150 secdec ±

Sensitivity deg12.5 secmV

deg12.5 sec

mV

Frequency Response 40 Hz 40 Hz Supply Voltage 5 V 5 V

Non-Linearity 0.1% of FS 0.1% of FS Axis-Misalignment 1± 1± Noise 0.05 deg/s/ Hz 0.05 deg/s/ Hz

TABLE IV MAGNETO-RESISTIVE TRANSDUCER ESPECIFICATIONS


Sensitivity 1.0mVV gauss

1.0mVV gauss

Frequency Response 5MHz 5MHz Noise Density 50 nV

Hz

50 nVHz


T02-007

IV. DIGITAL NEURAL NETWORK ATTITUDE ESTIMATION FILTER DESIGN

Artificial neural networks have been used in a wide variety of robotics applications. The idea in the I-AHRS sensor data fusion filter is to combine the outputs of the accelerometers, angular rate and magnetometers sensors to obtain a good estimate of the orientation attitude, calculating an error data between the estimated angles and the physical system angles, Figure 15.

Figure 15. Digital Neural Network Attitude estimation block diagram The idea in the EFIGENIA digital neural network I-AHRS computer was to make an ideal technique for improve the aerial vehicle attitude calculation and estimation process. During the flight, the system combines the data information from multiple sensors such as accelerometers, angular rate gyros, and magnetic sensors. The objective of this process is to provide high accuracy and low cost reliable system solution that ensures to the EFIGENIA enhanced orientation accuracy and minimized common errors from the system. For this purpose, was used a multilayer digital neural network as on-line learning estimator because of its high performance in multivariable and non-linear systems. In this way, the first step in the development of the I-AHRS computer was the design and development of a FPGA digital neural network chip. This chip is based on reconfigurable FPGA logic device, which contains important amount internal neurons (process elements). Each neuron processes a 32-bit inputs and weights. All sensors are mounted in a strap-down way. This gives more flexibility and reliability to the design, in other words; this means software mathematical transforms computation rather mechanical operations.

A. Neural Network FPGA based architecture design

The implementation of artificial neural networks is mostly done on a conventional processor computer board, in which this job consuming a long run time periods.

As this system integrate data stream from the all different sensors to build a consistent model that represent the features of the UAV aircraft attitude sensed and to provide continuous accuracy stabilization, then was developed a solution for the run time and speed processing, implementing the neural network on a parallel reconfigurable computer based on a Cyclone-II Field Programmable Gate Arrays (FPGA) combined with a dsPIC30F4013 Digital Signal Processors (DSP), because of the fast prototyping, reusability and high speed run-time data processing. The artificial neural network model used for this FPGA multichip computer architecture is the Back-Propagation Multi-Input Multi-Output scheme (MIMO). All inputs and outputs for this Back-Propagation are digitalized for allow to the artificial neuron (process element) calculates a weighted sum of its inputs, and add its threshold value. Next, the resultant value is evaluated by a non-linear function. The I-AHRS digital neural network computer employs the inertial sensors, and magnetic sensors as inputs to the neural network input layer. Then the reconfigurable neuro-computer starts the processing. The back-propagation neural network requires training information for supply the system with some sort of “experience”. In this way, the network must be trained using a learning rule: This process consist in providing external patterns which are compared with the network outputs through the training phases. This learning rule is used to estimate the neural network weights values, minimizing the sum squared error (delta) and adjusting continually the value of the weights. The system initializes the weights setting with random values. After these initialization parameters, then are applied an input vector to be read by the network:

tpNppppp XXXXXX ].....,.........,,,[ 4321= (6)

The neural network calculates the sum of the weighted inputs value ( h

pjneta ) for the hidden layer:

∑=

+=N

i

hjpi

hji

hpj XWneta

1θ (7)

Where:

p : p-th training vector j : j-th artificial neuron form the hidden layer

This sum of the weighted inputs conform the input to the transfer function.

)( hpj

hjpj netafi = (8)


T02-007

For the output layer the system calculate, again, the neta values from its inputs:

∑=

+=L

j

okpj

okj

opj iWneta

1θ (9)

and then:

)( opk

okpk netafo = (10)

The network has calculated the actual output values. But may will exist a difference between the network actual output values and the desired output values (an error). Then, for any output neuron, the errors for the output units are:

)(*)( ' opk

okpkpk

opk netafOy −=δ (11)

Where:

pky is the desired output.

pkO is the actual output.

)(' ppk

ok netaf is the gradient of a sigmoid function evaluated

at the respective neuron. For the logistic function:

1)( )1()( −−+=ojknetao

jko

k enetaf (12) The system can now calculate the error for the hidden layer units:

∑=k

okj

opj

hpj

hj

hpj wnetaf δδ *)(' (13)

Adjusting the weight values in the output layer will be:

pjopk

okj

okj itwtw ηδ+=+ )()1( (14)

Where: η is the learning rate parameter. For the hidden layer, the weight estimation and adjust values:

ihpj

hji

hji xtwtw ηδ+=+ )()1( (15)

allowing to compute the most effective attitude and obtain high accuracy outputs enhancing the I-AHRS performance.

B. Arithmetic Precision of the Hardware / Software Neural network Implementation

This FPGA Digital Neural Network is an inherently concurrent (parallel) architecture. The numerical precision is an important aspect for the system performance attitude estimation. The major point to take into account is the rule area vs. precision because the system must have a balance between the required numerical precision for an accurate Neural Network result and convergence, and the FPGA implementation capacity chip selection, avoiding errors into the system. For this case the system use a Back-Propagation Neural network scheme with a resolution of 32-bit with its input and output data normalized into [0 , 1] range and a sigmoid and linear activation functions :

1( )1 inetaf s

e−=+

(16)

( )f s s= (17)

In this concern, the Neural Network estimation system was implemented over a Cyclone-II FPGA because it’s high flexibility and capacity. For the implementation, the neural network architecture concept was simulated and validated into Matlab® and the Altera DSP Builder® toolbox. The first step was to simulate the basic process element for the neural network. Once the process element has been implemented and tested, the completed Multilayer Neural Network was constructed and simulated for finally be downloaded on the Cyclone-II FPGA electronic circuit; Figure 18.

Figure 19. FPGA Multilayer Neural Network implementation and simulation.


T02-007

V. EXPERIMENTAL RESULT To verify the results, some experiments were performed as in the ground as in the flight test. Data were collected to evaluate the accuracy and performance of the I-AHRS designed system. The data were processed for validate the Neural Network data fusion strategy solution through a software function running on a PC computer using the data received from the sensors developing “in-live” testing of the intelligent-algorithm and optimizing it. Figure 20 shows the result values in one of the system axis performed in the test starting in a level position and moving in each direction (clock-wise and counter clock-wise).

Figure 20. Data fusion output angle result for one axis.

VI. CONCLUSION

This paper has presented the analysis, design and construction of an attitude and heading reference system based on low-cost MEMS transducer constructed in-house and developed for the EFIGENIA UAV autonomous Unmanned Aerial vehicles UAV EJ-1B Mozart and EJ-2B Maria designed autopilots. Additionally, an improved method for combining data from the accelerometers, gyros and magnetometers based on a digital Neural Network has been developed to produce an accurate attitude angle measurements optimized for accelerated and non-accelerated UAV operation conditions.

ACKNOWLEDGMENT The author thanks to General. José Manuel Sandoval from the Colombia Air Force, Mr. Sergio Abramoff, from Analog Devices, Mrs. Carol Popovich from Microchip Technology, the Byte-Craft Limited, Altera Corporation and the University of Cauca.

REFERENCES [1] D. J. Biezad, Integrated Navigation and Guidance Systems. Reston,

Virginia, American Institute of Aeronautics and Astronautics AIAA, 1999.

[2] A. B. Chatfield, Fundamentals of High Accuracy Inertial Navigation. Progress in Astronautics and Aeronautics, Volume 174. Reston, Virginia, American Institute of Aeronautics and Astronautics AIAA, 1997.

[3] R. M. Rogers, Applied Mathematics in Integrated Navigation Systems. Restom, Virginia. American Institue of Aeronautics and Astronautics AIAA, 2000.

[4] P. Zarchan, H. Mussoff, Fundamentals of Kalman Filtering: A practical Approach. Reston, Virginia. American Institute of Aeronautics and Astronautics AIAA, 2000.

[5] R. Pratt, Flight Control Systems. Progress in Astronautics and Aeronautics. Restom, Virginia, American Institute of Aeronautics and Astronautics AIAA, 2000.

[6] K. C. Chang, Digital Systems Design with VHDL and Synthesis: An Integrated Approach. Los Alamitos, California. IEEE Computer Society, 1999.

[7] R. G. Bown, P. Y. C. Hawang, Introduction to Random Signals and Applied Kalman Filtering, John Wiley & Sons, 1997

[8] S. Haykin, Neural Networks, Toronto, Canada, Macmillan College Publishing Company, 1994

[9] R. J. Schalkoff, Artificial Neural Networks, Clemenson University, McGraw Hill International Editions, 1997

[10] B. L. Stevens, F. L. Lewis, Aircraft Control and Simulation, 2nd edition, John Wiley & Sons, 2003

[11] A. K. Jain and J. Mao, “Artificial Neural Networks: A Tutorial,” IEEE Computer Magazine, pp. 31-44, March 1996.

[12] Altera QuartusII Reference Manual. Available: http://www.altera.com [13] Altera DSP Builder Toolbox. Available: http://www.altera.com [14] Microchip dsPIC Digital Signal Controllers Reference Manuals.

Available: http://www.microchip.com [15] Analog Devices Sensor Products. Available: http://www.analog.com [16] Bytecraft C programming tools. Available: http://www.bytecraft.com


T02-007

[IEEE 2007 Mediterranean Conference on Control & Automation - Athens, Greece...

Documents

Transcript of [IEEE 2007 Mediterranean Conference on Control & Automation - Athens, Greece...