Neural Network Algorithm for Residential Fire Alarm System

Abstract

This paper compares two neural network algorithms using same sensor data set as input to neural network, i.e. temperature, smoke density, and CO concentration level. The neural network simulation model utilizes the strong features of a neural network like adopting and organizing itself with changing environment and quick learning ability for signal fusion, network training and problem simulation providing accuracy and reduced processing time.  We will consider here two models based on neural networks with a non- linear approximation and prediction capability. Neural network system function depends on neuron connectivity structure, weights, biases, type of feed/propagation, number of hidden layers and quantity of neurons in each layer.   Algorithms available in Neural Network are capable of performing classification, regression, clustering, and dynamic system modelling and control. In this paper we will compare and analyse two algorithms used for function approximation and nonlinear regression in terms of approximation capability, classification ability and learning rate to improve the fire detection accuracy in minimum time period. 

 Keywords: Levenberg-Marquardt, MATLAB, Mean Square Error, Regression, WSN.