Recurrent Trend Predictive Neural Network for Multi-Sensor Fire Detection

Mert Nakip,Osman Yildiz,Cuneyt Guzelis

doi:10.1109/access.2021.3087736

Abstract

We propose a Recurrent Trend Predictive Neural Network (rTPNN) for multi-sensor fire detection based on the trend as well as level prediction and fusion of sensor readings. The rTPNN model significantly differs from the existing methods due to recurrent sensor data processing employed in its architecture. rTPNN performs trend prediction and level prediction for the time series of each sensor reading and captures trends on multivariate time series data produced by multi-sensor detector. We compare the performance of the rTPNN model with that of each of the Linear Regression (LR), Nonlinear Perceptron (NP), Multi-Layer Perceptron (MLP), Kendall- <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\tau $ </tex-math></inline-formula> combined with MLP, Probabilistic Bayesian Neural Network (PBNN), Long-Short Term Memory (LSTM), and Support Vector Machine (SVM) on a publicly available fire data set. Our results show that rTPNN model significantly outperforms all of the other models (with 96% accuracy) while it is the only model that achieves high True Positive and True Negative rates (both above 92%) at the same time. rTPNN also triggers an alarm in only 11 s from the start of the fire, where this duration is 22 s for the second-best model. Moreover, we present that the execution time of rTPNN is acceptable for real-time applications.

Highlights

Nowadays, most fire detectors only detect smoke, the fire is a process that consists of smoke, flame, several kinds of gasses, temperature and humidity [1]
For the multi-sensor detector, we compare the performance of our Recurrent Trend Predictive Neural Network (rTPNN) model with that of each of the Linear Regression (LR), Nonlinear Perceptron (NP), Multi-Layer Perceptron (MLP), Kendall-τ combined with MLP (Kendall-MLP), Probabilistic Bayesian Neural Network (PBNN), Long-Short Term Memory (LSTM), and Support Vector Machine (SVM)
We have proposed a novel Recurrent Trend Predictive Neural Network which captures the trends of time series data via its recurrent internal architecture. rTPNN performs data fusion on the multivariate time series with its captured

Summary

INTRODUCTION

Most fire detectors only detect smoke, the fire is a process that consists of smoke, flame, several kinds of gasses, temperature and humidity [1]. In order to decrease both FNR and FPR, in this paper, we propose the Recurrent Trend Predictive Neural Network (rTPNN) architecture for multi-sensor fire detectors. The recurrent internal architecture of the rTPNN model successfully captures the trends in the time series data of each sensor that minimizes the error for the output of rTPNN; it significantly improves the overall prediction performance of the neural network. We evaluate the performance of the multisensor fire detector based on rTPNN for the 9 different reallife fire experiments from the publicly available data set [8], [9] For these experiments, for the multi-sensor detector, we compare the performance of our rTPNN model with that of each of the Linear Regression (LR), Nonlinear Perceptron (NP), Multi-Layer Perceptron (MLP), Kendall-τ combined with MLP (Kendall-MLP), Probabilistic Bayesian Neural Network (PBNN), Long-Short Term Memory (LSTM), and Support Vector Machine (SVM).

RELATED WORKS

RECURRENT TREND PREDICTIVE NEURAL

PARAMETER TUNING FOR STATE-OF-THE-ART MODELS AGAINST WHICH rTPNN IS COMPARED

Findings

CONCLUSION