0

I am an undergraduate currently working on a control design project that involves Nonlinear neural internal model control of a Quadruple Tank Process with Dead Times (QTPwDTs). I have been however stuck on performing system identification of a simple single-tank process using neural networks in MATLAB and SIMULINK which has been impeding my progress on my undergraduate project.

I am admittedly a Newbie in this field, and do not have access to a lot of resources for learning.

Here are the steps I have taken so far:

  1. Data Generation
  • I developed a SIMULINK model for the single tank process

SIMULINK model for the single tank process

  • The Random Number block was used as an input to the model with a Mean of 60, Variance of 100, Seed of 0, and a Sample Time of 1s.
  • Here is the MATLAB initialization script for the steady-state values for the model:
clear all;
clc; 

u = 150; %flow rate
K = 0.7; %flow co-efficient
A = 150; %cross-sectional area of the tank
g = 981; %acceleration due to gravity
a = 2.5; %cross-sectional area of the hole through which the water flows out

  • The model was run for a simulation time of 20000s

  • The scope output for the input and output signal is shown below:

  • Dimension of the input and output data: 20001x1 double

  1. Neural Network Training
  • I trained a feedforward neural network with two hidden layers having 100 neurons.
  • I observed that wider networks (more neurons) yielded a better validation result via experimentation
  • The input regressors to the neural network are [u(k), u(k-1), u(k-2), y(k-1), y(k-2)]. Where u is the input and y is the output extracted from the SIMULINK model for the data extraction phase.
  • I used the poslin activation function for the two hidden layers
  • I trained the neural network with 70% of the original data using the nntool
  • I created the SIMULINK block simulating the neural network using the gensim function with a sampling time of 1.
function [nn, performance] = train_mff_network(input, output)
% Create a Neural Network Toolbox block
nn = feedforwardnet([100,100]);

N = length(input);

% define the inputs and the targets 
% input regressors = [u(k),u(k-1),u(k-2),y(k-1),y(k-2)]
inputs = [input [0;input(1:N-1)] [0;0;input(1:N-2)] [0;output(1:N-1)] [0;0;output(1:N-2)]];
%inputs = single_tank_nn_in';
targets = output;

% configure the neural network 
nn = configure(nn,inputs',targets');

% Specify the network architecture and training algorithm
nn.layers{1}.transferFcn = 'poslin';
nn.layers{2}.transferFcn = 'poslin';
nn.trainFcn = 'trainlm';
nn.divideFcn = 'dividerand';
nn.divideParam.trainRatio = 0.7;
nn.divideParam.valRatio = 0.15;
nn.divideParam.testRatio = 0.15;

% Train the neural network
[nn,tr] = train(nn,inputs',targets');

% Validate the trained network
outputs = nn(inputs');
performance = perform(nn,targets,outputs)
end
  1. Validation

  • To validate the performance of the neural network;

  • I implemented the SIMULINK model for validation in MATLAB using the gensim generated neural network block

  • The SIMULINK model for validation is shown below:

  • I tested the original model and the feedforward neural network model with an input signal generated using the Random Number block with the same configuration used for data extraction

  • A simulation time of 10 seconds was used

  • Here is the validation output obtained:

  • I am expecting the output of the feedforward neural network to closely resemble/approximate the output of the original single tank process SIMULINK model used for data extraction

  • I have tried to retrain with other neural network configurations, however, I have not obtained any major satisfactory result yet.

I strongly believe I am doing something wrong. I really hope an expert can examine the steps I have taken and save me from my current dilemma because my supervisor constantly tells me that it is a simple problem, and I am still far from my main objective which is training a neural network for the Quadratic tank process with dead times that has more complex/problematic dynamics.

Similoluwa Okunowo
  • 1

0 Answers0