Linear modulated test signals

Linear modulated test signals

Modern communication systems are continuously developed to reach higher data rate. Nowadays, digital communications are widely used, taking advantages of digital signal filtering and modulation progression, which offers a greater data rate and a judicious exploitation of the frequency band.
To make the data transmission more resistant to the transmission channel disturbances, extra data bits for encoding, commonly called redundancy bits, are used. Source encoding purpose is to ensure a data transmission more efficient.
by realization of data compression. These characteristics and operations make the amplified and transmitted source of information very redundant ; it is a specification that will be advantageous in realizing the proposed data on-line modeling and predistortion architecture. Linear modulation techniques are widely employed, for their high data rate using a relatively narrow frequency band. In this section, it will be presented briefly the linear modulation techniques tested in the simulations of the present work. Binary symbol encoding generates the symbol Ck from input digital signal ;
it can be complex or real. The symbol rate measured in «baud» is given by the number of symbols transmitted in unit of time. In the case where the symbol Cfc is complex, this value must be coded in amplitude and phase.

Neural network theory

The modeling of microwave PAs is one of the key subject in wireless communication systems. This is mainly driven by the need of precise model to be used for behavioral study and linearization of PAs. Through evolution of technology in DSPs, FPGAs, ADC and DAC, the PA is modeled and dynamically characterized on a wide bandwidth with excellent performances. With the improvement
of modulation techniques, the dynamic behavioral modeling is inescapable. In this section, it will be validated, by simulation, with the test signal (16-QAM) the results given in several works ’41> 28′. A RVTDNMs architecture will be realized with XSG software in the Matlab/Simulink environment.
The proposed model is based on the sampled baseband signal, after being filtered by the pulse shaping filter. The model based on data before the pulse shaping filter can not well model the output-of-band behavioral, because it can not compensate the out-of-band distortions when linearizing .

Stopping criterion

Stopping criterion is a test operation realized at each learning epoch, it can be decided whether or not to stop or not the learning process ; we make the comparison between the calculated MSE and the acceptable MSEacc defined beforehand according to inequality . This value is obtained in function of precision and exactitude of the possible desired results .The following steps are realized during the learning process of the NK :
1. Initialize the synaptic weights wJii)d at random values (generally between -1 and +1) ;
2. Initialize an accumulator of instantaneous errors energy to zero E(0) = 0;
3. Present a sample of the input signal, and calculate the outputs of all output layer neurons using the equations 2.24 and 2.25 ;
4. Specify the desired output and evaluate the local gradient for all neurons using equations 2.35 and 2.37;
5. Adjust synaptic weights according to equations 2.44, 2.45 and 2.46 ;
6. Calculate the error energy from equation 2.28 and add it to accumulator ;
7. Repeat steps 3 through 6 for each instant of the learning epoch ;
8. Calculate the mean square error MSE according to equation 2.29 ;
9. If MSEcai < MSEOCC then go to 10, else go to 2;
10. Save weights and biases of the M and end.
In this algorithm procedure, steps 2 to 8 represent a learning epoch

Back-propagation algorithm modification

Neurai networks are used in linearizing PAs because of their capability of modeling and fitting nonlinearities. Using that, this characteristic can be generalized to memorize severe nonlinearities with memory. Linearization with RVTDNNs is reported in the work of Hwangbo et ai ’28’ by using the indirect learning architecture. It was proved that the trained NN is able to operate with (3G) base-station signals such as WCDMA and CDMA2000t4X1. The proposed NN for linearization consists of a similar architecture (RVTDNNs) used for PAs modeling with the back-propagation algorithm in the adaptation process.  the input and output cartesian signals are compared to train the NN implemented on the Xilinx FPGA, low-pass filters are used to filter the signals after ADC and DAC operations done by the circuit Memec pl60. The proposed NN predistortion architecture is based on sampled baseband signals, the Mod/Dem operations are achieved before using the signals by the FPGA board. This architecture can be applied for any type or class of PAs. It can be achieved with complex signals like CDMA and OFDM without modification of the NN general structure.

Neural network architecture implementation with XSG software

In this section, we present the proposed NN architecture implementation with XSG software which will be implemented on FPGA board in future research work. This architecture is designed for applications with Xilinx FPGA circuits, and this software gives the opportunity to realize applications with very high integration level. It has been adopted as a nonlinear activation function of the hidden layer neurons the tangent hyperbolic function tanh(). In digital systems, there are many kinds of representation for data, values of synaptic weights, biases. Input and output signals are real-valued and they can be presented even in digital or analog forms. When using digital representation, signal values can be presented on floating or fixed-point format and they can be serial or parallel. Ideally, design with floating point precision is more flexible and easy for manipulation. However, this advantage has a price in term of silicon surface and number of input/output pins. For hardware implementation and because of the high price of arithmetic with floating point precision, it is necessary to use the fixed point simple precision presentation with two’s complement Fix[30-28] format which can provide a precision of about 4 x 10~9 on the parameter variations and data are between -1 and 1. The work frequency of the FPGA board depends on the signal frequency and the rapidity of the ADC/DAC circuits. Because Matlab uses double precision presentation, the XSG software provides the use of blocks Gateway-in and Gateway-out like communication point between the FPGA and Simulink parts. The general presentation of data with simple fixed point precision consists of two parts : signed or unsigned numbers. Two’s Complement Fixed Point Format can even represents negative and positive numbers.

Table des matières

CHAPITRE I:INTRODUCTION 
1.1 Généralités
1.2 Problématique
1.3 Hypothèses
1.4 Méthodologie retenue
1.5 Contributions apportées
CHAPITRE II:A NEURAL NETWORK APPROACH FOR THE LINEARIZATION OF RADIO FREQUENCY POWER AMPLIFIERS WITH ADAPTIVE PREDISTORTION 
2.1 Abstract
2.2 Introduction
2.3 Linear modulated test signals
2.4 Power amplifier original model
2.4.1 Memoryiess nonlinear subsystem
2.4.2 Memory linear subsystem
2.4.3 Memory nonlinearity system
2.5 Linearization performance criteria
2.6 Memory power amplifier data on-line modeling architecture
2.6.1 Neural network theory
2.6.2 Back-propagation algorithm
2.6.3 Simulation results
2.6.4 Discussions and conclusion
2.7 Memory power amplifier data on-line predistortion architecture
2.7.1 Back-propagation algorithm modification
2.7.2 Simulation results using 16-QAM test signal
2.7.3 Test of the architecture with other PA model and modulated signal
2.7.4 Discussions and conclusion
2.8 Neural network architecture implementation with XSG software
2.8.1 Forward propagation module
2.8.2 Backward propagation module
2.8.3 Decision module
2.8.4 Discussions and conclusion
2.9 General conclusion and future works
CHAPITRE III:CONCLUSION GÉNÉRALE

Télécharger le rapport complet