| Summary: | 碩士 === 國立雲林科技大學 === 電機工程系 === 103 === In modern society, while the usage of natural resources such as fire and electricity has become more and more widespread, fire accidents occur more often than ever, leading to a larger scale of damage and loss of lives. Therefore, it is essential to detect fire accidents and notify people immediately of yet correctly. When it comes to the traditional fire sensor, a physical limitation is its susceptibility to be influenced and affected by surrounding environment, which results in false alarms or even malfunctions.
As a remedy, this thesis presents the implementation of a fire detection system that operates based on the Dempster-Shafer theory allowing for multi-sensor data fusion. Our development designates the Arduino microcontroller to measure and interpret the readings from three types of sensors: temperature, light and gas. All the readings are normalized to a moderate range of values for reasoning purposes and then transmitted to the backend Raspberry Pi over the wireless medium for subsequent inference processing. In our architecture, the Raspberry Pi is tasked to deduce the probability of fire accident occurrences in light of Demspter-Schaffer’s theory. Deduction also evaluates a conflict coefficient (Kconflict) as a pivotal parameter to validate the misbelief of our deduction results. Conflict coefficient is of utility to reduce false alarms or possible malfunctions of our system. Further, we have deployed a web server with a MySQL database as well, so as to interact with the Raspberry Pi for keeping necessary data in storage for future reference.
In addition, our system offers three notification services. First, a web page is maintained to clearly display the system’s operation details. Second, an application enables the smartphone user to check the operational status of our system online anytime as long as there is accessibility to the Internet. Third, a messaging service informs the user of emergent events in real-time over SMS whenever necessary.
We validate our development by field tests to collect realistic statistics. Experiments are conducted in a controllable, safe environment by emulating fire happening in daytime and nighttime, respectively. Each experiment undergoes three phases: no-fire, on fire, and post-fire situations. We are concerned with the detection sensitivity versus accuracy of real fire accidents. Experimental results show that, when Kconflict is set to 0.8, the accuracy of our implementation in no-fire and on-fire circumstances during daytime reaches 98% and the accuracy during nighttime amounts to 97%. When taking all the three phases into account, the accuracy during daytime and nighttime turns out to be 97% and 89%, respectively. Field tests ensure the effectiveness of our implementation, implying its practical use in real life.
|
|---|
