Fire Alarm System Through Smoke Detectio
Fire Alarm System Through Smoke Detectio
Fire Alarm System Through Smoke Detectio
ABSTRACT- A reliable fire alarm system (FAS) is subsequent stage involves dividing the fire regions into
crucial for timely reporting and responding to fires. While spatio-temporal blocks and extracting correlation features
existing techniques can predict undesirable outcomes, they that capture both spatial and temporal characteristics. These
lack guidance on when and how workers should intervene features are utilized to train and evaluate an SVM classifier.
to minimize the associated costs. Recent advancements in Remarkably, the method demonstrates accurate detection of
sensors, microelectronics, and information technology have general flames in the first stage, making it suitable for
significantly improved fire detection technologies. initial flame detection and sample selection in machine
However, the prevalence of synthetic materials in modern learning. Experimental results showcase its superior
homes has increased the danger of fire-related injuries and performance compared to existing methods in terms of true
deaths due to the release of toxic fumes and gases, positive rate, false positive rate, and missing rate when
including carbon monoxide. This highlights the need for tested on fire videos.
ongoing analysis and development of fire detection In this paper [2], the author developed and evaluated an
techniques to ensure the safety of occupants. alarm verification system using real industry data,
combining stream processing, batch processing, and
KEYWORDS- Fire Alarm System, Smoke, Fire Safety, machine learning. The system achieved over 90% accuracy
Synthetic Materials in classifying alarms at a streaming rate of 30K alarms per
second. They also explored integrating unstructured data to
I. INTRODUCTION improve classification accuracy and shared valuable lessons
Fire is a dangerous threat for the safety of residents. learned for researchers and practitioners building similar
Therefore, a form of fire disaster prevention is needed for systems.
residents by designing a fire early warning system. In our In this study [3], they focused on improving safety in smart
project, we have done data pre-processing using smote for homes by integrating gas leakage and fire systems using
sampling, min-max for normalization, outlier’s detection, cost-effective, energy-efficient devices and M2M
correlation matrix. The algorithms used in our project is communication protocols. They proposed a dataflow system
random forest, decision trees and K-Nearest Neighbour. that collects valuable information and sends it to a central
The systematics of this paper is as follow. The first part point in the M2M home network. Through experiments
discusses about motivation. Background, problem with our prototype system and various sensors, we applied
identification, purpose, and method are exposed in this supervised machine learning techniques to identify
section. The second part discusses about related works, unexpected events and predict the level of danger.
previous research and papers that are related to the current Additionally, their system can send timely notifications to
research are discussed. The third part discusses about alert relevant individuals and take appropriate actions.
system design, the hardware specification and the algorithm This research [4] explores the utilization of acoustic sensors
of each classifier are exposed in this part. The fourth part for early fire detection, aiming to enhance the reliability of
discusses about evaluation, the comparison of each existing fire detection methods by incorporating the
classifier are exposed in this part. The Fifth part discusses acoustic signals generated during fire events. The study
about conclusion, important findings are highlighted in this demonstrates that the acoustic vibrations produced by
part. Future work are also discussed in the fifth chapter. heated materials can be employed to train a machine
learning algorithm for fire detection. The results indicate
that the proposed machine learning approach successfully
II. RELATED WORKS identifies fire events based on the measured acoustic
This section goes through some recent works related to fire signals. However, the performance of the convolutional
alarm system. neural network varies depending on the specific material
As per the study of this paper [1], the research introduces a exposed to the heat. It should be noted that this method
novel two-stage fire detection approach utilizing video serves as a proof-of-concept, and further investigation is
processing. In the initial stage, a combination of color and required to examine factors such as different acoustic
dynamic features is employed to identify real flames, while environments and materials, and potential techniques like
the motion orientation determines the fire region. The
Innovative Research Publication 1
International Journal of Innovative Research in Computer Science & Technology (IJIRCST)
transfer learning, domain adaptation, and data This study [9] explores the application of wireless sensor
augmentation. networks and machine learning methods for detecting
This research [5] findings indicate that the Naïve Bayes overheating and fire hazards within buildings. A wireless
classification method demonstrates the highest performance sensor network was deployed to gather temperature data
among the tested algorithms for improving the detection from specific locations of interest. A novel machine
rate of the smart fire alarm system prototype. With an learning algorithm was developed to intelligently analyze
accuracy of 100%, Naïve Bayes outperformed the decision the collected data, taking into account information from
tree (99.5%) and kNN (90%) due to limitations in rule individual sensors as well as their correlations. Through this
complexity and varying data distances, respectively. The algorithm, normal patterns were identified using training
SVM method yielded the poorest results, suggesting that the data, and hazards were detected by measuring the deviation
linear hyperplane approach may not be suitable for data of new observations from these established patterns. The
classification. Future work could involve implementing proposed approach is fully automated, eliminating the need
online learning for real-time operation of the Fire Alarm for human intervention. Experimental tests using real-world
System while dynamically selecting the optimal classifier. sensor data validated the effectiveness of the machine
In this paper [6], author presents the design of an Internet learning algorithm, demonstrating its potential in accurately
of Things (IoT) fire alarm system that effectively alerts both detecting hazards.
the owner and the fire station during emergencies. The This paper [10] presents a novel approach for flame alarm
research focuses on distinguishing between different states systems called New Feature Extraction and Description.
of the surrounding environment, including sleep, passive, The method involves extracting fire candidates in stable
and active states, by utilizing a decision tree algorithm. By positions using a Chroma-intensity map, and utilizing the
employing this approach, a cost-effective IoT system is Brownian Correlation Descriptor as a discriminative feature
developed, offering an ideal fire safety solution for vector. This descriptor incorporates dynamic dependence
residential areas and hospitals. measures based on multiple adjacent frames to classify fire
Researchers [7] have expanded the application of machine and fire-like objects. The proposed method demonstrates
learning to predict fires in a specific scenario where superior performance with a low false positive ratio
multiple fires break out rapidly. Although this may seem compared to recent algorithms evaluated on the dataset.
implausible, it has become a grim reality in certain parts of However, it should be noted that our approach has a higher
the western United States due to the dry and hot conditions computational complexity due to the utilization of
exacerbated by climate change. These circumstances pose additional frames.
significant fire risks and often lead to extensive destruction.
Therefore, timely fire prediction is of paramount III. METHODOLOGY
importance to enable swift action. Conventional fire
monitoring approaches often struggle to detect fires in real- A. Data Collection
time, highlighting the need to address this critical issue. While selecting the data we made sure that our data
Numerous innovative approaches have been proposed to should include all the attributes needed to get as precise
anticipate and predict forest fires, with a growing emphasis result as possible.
on integrating machine learning to enhance detection and
B. Our data includes the following attributes-
prediction capabilities. Machine learning models offer
Temperature[C], Humidity [%], TVOC[ppb],
advantages such as pooling operations, extracting detailed
eCO2[ppm], Raw H2, Raw Ethanol, Pressure[hPa],
features, and leveraging substantial information, resulting in
PM1.0, PM2.5, NC0.5, NC1.0, NC2.5 and CNT.
effective fire prediction outcomes. Furthermore, these
models hold practical value in planning fire prevention and C. Data Preprocessing
management strategies. DATA SELECTION- We used the heatmap method to
The utilization [8] of wireless sensor networks and machine select the data of interest. A heat map plots these
learning in the proposed forest fire detection system has coefficients to visualize the strength of correlation
proven to be highly effective, offering improved accuracy between variables. It helps you find the best features for
in fire detection. The system ensures optimal performance your machine learning modeling. A heat map converts
by conducting analysis at both the sensor node and the base the correlation matrix to a color code.
station, resulting in faster and more precise outcomes. A OUTLIER DETECTION- Outlier Detection is a key
threshold ratio is implemented within the sensor node to stage in numerous information mining undertakings.
adapt to various weather conditions and environments. This The motivation behind the outlier detection technique
system can be deployed in any part of the forest, even here is to recognize the boundaries impacted by the
without pre-existing network connectivity, as it relies on a anomaly device out of thousands of boundaries. There
dedicated built-in network infrastructure. With primary are a few ways to deal with identifying exceptions.
power supplied by rechargeable batteries and secondary Boxplots were utilized for the information. Boxplots,
solar power, the system can operate independently for otherwise called bristle plots, are made to show an
extended periods. During real-life trials conducted in outline of a bunch of information values with properties
tropical forest sites, the integrated communication like least, first quartile, middle, third quartile, and most
infrastructure of the proposed system exhibited faster extreme.
response times, alerting relevant authorities promptly
compared to existing systems.
Figure 1: Methodology
DUPLICATE HANDLING- We utilized Pandas built in SCALER change highlights by scaling each element to a
technique drop duplicates () to drop copy columns. given reach. This assessor scales and deciphers each
Naturally, this strategy returns another Data Frame with element separately to such an extent that it is in the
copy columns eliminated. We can set the contention in given reach on the preparation set, for example
place=True to eliminate copies from the first Data somewhere in the range of nothing and one.
Frame. D. Now, after that, we bifurcated our data into training and
HANDLING NULL VALUES- We utilized drone () for testing data. Using 80% of data for training and 20% of
dealing with invalid qualities. The drone () capability is data for testing.
utilized to eliminate missing qualities. Decide whether E. While training the data, we implemented 3 different
lines or segments which contain missing qualities are models.
eliminated. 0, or 'index’: Drop columns which contain RANDOM FOREST- Random Forest is a powerful and
missing qualities. 1, or 'columns’: Drop sections which versatile supervised machine learning algorithm that
contain missing worth. grows and combines multiple decision trees to create a
SAMPLING-For this we utilized the destroyed strategy "forest". It can be used for both classification and
for Destroyed (Engineered Minority Oversampling regression problems in R and Python. A random forest
Method) is one of the most normally utilized algorithm consists of a collection of decision trees,
oversampling procedures to tackle unevenness issues. where each tree in the ensemble consists of a sample of
The point is to adjust the class conveyance by copying data drawn from the training set using permutations
and haphazardly expanding the minority class models. called bootstrap samples.
SMOTE orchestrates new minority examples between DECISION TREES- A decision tree is like a flowchart
existing minority occasions. where each inner node represents a test of a feature
NORMALISATION- For standardization of the (such as whether a coin is flipped heads or tails) and
information we utilized min-max scaling. MIN-MAX each leaf node represents a class label (the decision
made after all features have been computed) and a In future, we will collect more additional data and apply
branch. structure. Represents the combination of other machine learning algorithms to improve the accurate
features that lead to these class designations. of the model to reduce the false positive. We have also the
K-NEAREST NEIGHBOUR- The k-nearest neighbor’s trend toward real-time analysis using cloud services
algorithm, also known as KNN or k-NN, is a non-
parametric, supervised learning classifier, which uses CONFLICTS OF INTEREST
proximity to make classifications or predictions about
the grouping of an individual data point. The authors declare that they have no conflicts of interest.
F. While testing the data, we applied the above 3 medals to
the rest 20% data. REFERENCES
[1] Hu, Z. W. (2019, September 20). Research on two stage
IV. NOVELTY flame detection algorithm based on fire. Retrieved from
For this project, we have read more than 10 research papers ACM Digital Library:
https://dl.acm.org/doi/10.1145/3366194.3366296
and in all of them the most used method was image sensing
[2] Kaiser, A. S. (2018). A Hybrid Approach for Alarm
and processing (a deep learning model) while we used Verification using Stream. Retrieved from
supervised learning models like Decision trees, KNN and openproceedings.org:
Random Forest (this being the algorithm as far as we have https://openproceedings.org/2018/conf/edbt/paper-246.pdf
researched used nowhere for fire alarms and smoke [3] Lamine Salhi Thomas Silverston, T. Y. (2019). Early
detection). All of our project works textually without use of Detection System for Gas Leakage and Fire in. Retrieved
any kind of images. Our project reduces the complexity in from IEEE Xplore:
solving the project. https://ieeexplore.ieee.org/document/8661990
[4] Mogren, J. M. (2022, july 19). A Novel Method for Smart
Fire Detection. Retrieved from springer:
V. RESULTS https://link.springer.com/article/10.1007/s10694-022-01307-
When we implemented the 3 models to the testing data, the 1
[5] Putrada, G. S. (2019). Comparison of Classification
below table 1 shows the accuracy given by them.
Algorithms to Improve. Retrieved from IEEE Xplore:
https://ieeexplore.ieee.org/abstract/document/8935885
Table 1: accuracy Results [6] Sivaraman, R. A. (2021, December). Machine Learning
ALGORITHM ACCURACY Based Fire Alarming System Using. Retrieved from
Researchgate:
Decision trees 99.98% researchgate.net/publication/357159137_Machine_Learning
Random Forest 99.97% _Based_Fire_Alarming_System_Using_Internet_of_Things
[7] Surya, L. (2017, March 1). Risk Analysis Model That Uses
KNN 99.95%
Machine Learning. Retrieved from IJCT:
The below table 2 shows the precision and recall. https://www.researchgate.net/publication/349310241_Risk_
Analysis_Model_That_Uses_Machine_Learning_to_Predict
Table 2: Evaluation Metrics _the_Likelihood_of_a_Fire_Occurring_at_A_Given_Propert
y
[8] Udaya Dampage *, L. B. (2022, January 7). Forest fre
detection system. Retrieved from
nature.com/scientificreports:
https://www.nature.com/articles/s41598-021-03882-9
[9] Wang, Q. H. (2012, june). Intelligent Building Hazard
Detection Using Wireless Sensor. Retrieved from
researchgate:
https://www.researchgate.net/publication/268586834_Intelli
gent_Building_Hazard_Detection_Using_Wireless_Sensor_
Network_and_Machine_Learning_Techniques
[10] YUN, K. S. (2017, February). An Improved Feature
VI. CONCLUSION Extraction and. Retrieved from space.snu.ac: https://s-
space.snu.ac.kr/bitstream/10371/122863/1/000000142478.pd
In this paper, we have contributed by working with purely
f
textual data, but as far as we've seen, all advances in fire
alarms have been done with image data through image
capture. Our project contributes to the achievement of SDG
#3 “Health and well-being”. Fire hazards are one of the
most well-known disasters and our project helps prevent
them in a more efficient way. We proposed an effective
data flow system for collecting useful information at the
central point of the project. From the dataset, we applied a
supervised machine learning process to multiple algorithms
to observe out-of-pattern events and predict levels of hazard
risk. After applying the algorithms, the following table
shows the accuracy achieved-