New M.E. Thesis Submitted from ECE Student


With the explosive growth of communication system and many of the communication system today utilize digital signal processors (DSP) to resolve the transmitted information. Finite impulse response (FIR) filters have been and continued to be important building blocks in many digital processing systems (DSP).For the design of Low pass FIR filters complex calculations are required. Mathematically, by substituting the values of Pass band, transition width, pass band ripple, stop band attenuation, sampling frequency in any of the methods from window method, frequency sampling method or optimal method we can get the values of filter coefficients h(n).
Here, Kaiser Window method has been chosen preferably because of the presence of ripple factor (β). Considering Low pass Filter design, the range of values for the parameters required are calculated.A data sheet through programming is performed on the platform of Matlab. For 30 different range of parameters, the values of h(n) i.e. coefficients of FIR filter, named desired result have been calculated named Artificial Neural Network is a highly simplified model of the structure of the biological neural network. It consists of interconnected processing units. In this thesis, ANN model has been designed which is used to design the low pass FIR which in the specified range of parameter which has been used to train the neural network. Basically, ANN can be trained by many methods like Feed forward neural network, Feedback neural network. But in this is thesis the feed forward neural network has been chosen to train the network. Here Multi Layer perception method is used in which BPN (Back Propagation network) algorithm has been implemented for the training of neural network. BPN method has been chosen because the objective of thesis is to use learning algorithm for a Multi Layer Feed Forward Neural Network, so that the network can be trained to capture the mapping implicit in the data set of input-output pattern pairs. The approach to be followed is basically a gradient descent along the error surface to arrive at the optimum set of weights. The error is defined as the square to arrive at the optimum set of weights. The error is a squared difference between the desired output (i.e. mentioned in data sheet using Matlab) and the actual output obtained at the output layer of the network due to application of the input patterns from the given input-output pattern pair. Firstly, we have calculated the output using the current setting of weights in all the layers. But optimum result has been achieved by adjusting the weights of the network by using the features of Back Propagation Network.The Artificial Neural Network has been trained using datasheet which we have made with the help of programming in the Matlab. The network has been trained using Multilayer Perceptron in which Error Back Propagation Algorithm has been specifically used to design Low Pass FIR filter. Using “Levenberg –Marquardt” (trainlm) in the neural network feed forward (newff) the goal meet condition has been achieved by substituting the values from datasheet input parameters:-Pass Band Frequency(PBF),Transition Width(TW) ,Pass Band Ripple (PBR) , Stop Band Attenuation (SBT), Sampling Frequency(SF) and Filter Length(N) as ‘p’ and filter coefficients (h(n)) as target ‘t’ in the program. The other parameters like, net.trainParam.epochs, net.trainParam.goal, have been set accordingly so that goal can be achieved/met. When we simulate this network in the Matlab command window the values of filter coefficients(h(n)) comes in the command window come in the command window which is the required result for the filter design.Secondly, there are two types of graphs which are coming, first graph is performance graph which has come by training the network means after the meeting the goal. Second graphs are error graphs means the error between actual (results comes from ANN) and desired results (Matlab data sheet results).Now, using Artificial Neural network low pass FIR filter can be designed. Secondly Error graphs show the error between actual and desired results are very less. Thirdly, with in the range of trained network filter coefficients for unknown input parameters can be found out and the results come from ANN and calculations are almost same even graphs are also available for that.

Leisure Readings :