Advanced Indoor Fire Prediction Using AI Models

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Oct 20 2023

In a recent publication in the journal Fire, researchers explored ways to improve indoor fire prediction for quicker rescue decisions, reducing harm. Using a 300-scenario database generated using computational fluid dynamics (CFD) tools, artificial intelligence (AI) models forecast temperature, carbon monoxide (CO) concentration, and visibility.

Background

In the domain of fire safety, where the preservation of structures and lives is paramount, the menace of fire presents significant perils, encompassing loss of life, property, and infrastructure. Real-time monitoring of parameters such as temperature, smoke, and images plays a central role in identifying and predicting indoor fires.

Several previous studies developed models for predicting smoke visibility, fire size, smoke removal rate, and real-time temperature at the field. Yet, further advancements are needed for forecasting toxic gas concentrations and smoke visibility.

Methodology

The current study entails several key stages. Firstly, numerical simulations, underpinned by CFD tools, generate fire scenario databases featuring varying fire locations and severity levels. Subsequently, data preprocessing transforms the original simulation data into formats conducive to predictive modeling. The research proceeds to design data experiments that assess the performance of AI models, considering different sensor data availabilities and prediction horizons. Finally, a comparative study ensued to yield insights into cost-effective solutions for intelligent fire prediction.

Numerical Simulation Models: Recognizing the challenges of executing real fire experiments, researchers leverage numerical simulation techniques to construct physical models. These models facilitate AI-driven fire predictions based on simulation data. The CFD models are created using PyroSim, a fire dynamics simulation (FDS) software. Notably, PyroSim offers a 3D graphical pre-processing feature that streamlines the setup of complex fire models. As a professional fire dynamics tool, PyroSim computes and outputs a multitude of fire-related results and provides post-processing functions to preserve these results.

Deep Learning Algorithms: The current study employs deep learning algorithms, including Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and Transpose Convolution Neural Network (TCNN), for analysis. LSTM excels at capturing non-linear and non-stationary time-series information, making it ideal for processing extensive datasets with lengthy time-series data. In this study, a LSTM with single layer is employed to extract features from time-series data for temperature.

CNN excels at image feature extraction through operations such as pooling, convolution, and full connectivity. The TCNN model has emerged as a variant of CNN, primarily used for upsampling to enhance image resolution. In the current study, a single TCNN layer is used to augment the predicted fire situation distribution array, while a single CNN layer extracts spatial features. This combination enhances fire prediction accuracy.

Experiments: The meticulously crafted simulation model was created using PyroSim with FDS version 6.7.9. The simulated space had a volume of 1280 cube meters with a uniform computational grid. Structural simplifications were made for ease of analysis. The room featured concrete walls, an open door, and a window for natural ventilation.

The simulation included 80 temperature sensors evenly distributed on the ceiling to monitor temperature. A polyurethane burner represented the fire source, adhering to FDS-recommended parameters. Data acquisition occurred at one Hz, with each simulation spanning 300 seconds. Key data slices were recorded at a height of 1.6 meters, and the database encompassed fire source location and severity level variations for analysis.

For deep learning models, a sliding window approach was applied to extract sequence segments from the sensor data tables, generating training and test datasets. The AI model used PyTorch and included an LSTM layer, a fully connected layer, a TCNN layer, and a CNN layer.

Performance evaluation metrics consisted of the mean squared error (MSE) for loss and the coefficient of determination (R square) for model evaluation. The model achieved an R square of approximately 82.9 percent, demonstrating its effectiveness in predicting temperature, CO concentration, and visibility distributions during fires.

Examination of AI Model Performance

The authors undertake a comprehensive exploration of the predictive capabilities of AI models. Their focus lies on varying quantities of sensors and the arrangement of said sensors. The primary goal here is identifying the most cost-effective combination of both the number and placement of sensors for practical applications. This approach is aimed at optimizing resource allocation by reducing investments in hardware and software while upholding prediction accuracy. Furthermore, the investigation delves into the influence of differing prediction horizons, ensuring the reliability of predictions.

Optimizing Sensor Placement: In determining the most optimal sensor arrangement, the study presents its findings through scatter diagrams, which illustrate prediction accuracy for temperature, CO concentration, visibility distribution, and overall accuracy. The study concludes that an arrangement featuring more than four sensors can maintain an overall prediction accuracy of 80 percent or higher.

Additionally, greater spacing between adjacent sensors, especially along the x-axis, enhances prediction accuracy owing to the distinct features captured by sensors positioned at greater distances.

Prediction of Temperature, CO Concentration, and Visibility: Concerning temperature predictions, the study scrutinizes the capacity of the model to forecast the evolution of fire-related temperatures, particularly highlighting the challenge of predicting the precise location of the fire source. Likewise, in the case of CO concentration predictions, the AI model demonstrates effective forecasting, albeit encountering difficulties in predicting high CO concentrations near the fire source. The study also identifies the model's capability to predict visibility distributions, ultimately concluding that it excels in this aspect when compared to temperature and CO concentration predictions. Nevertheless, minor errors may arise, but they fall within acceptable thresholds.

Conclusion

In summary, researchers introduced deep learning models for advanced indoor fire prediction and an optimization approach for fire sensor system layout based on deep learning. The current study aims to predict temperature, CO concentration, and smoke visibility distribution in advance, enhancing fire decision-making.