Examining the Effects of Deep Learning Model Structure on Model Interpretability for Time-Series Classifications in Fire Research DOI Open Access
Wai Cheong Tam, Linhao Fan, Qi Tong

et al.

Journal of Physics Conference Series, Journal Year: 2024, Volume and Issue: 2885(1), P. 012097 - 012097

Published: Nov. 1, 2024

Abstract This present work utilizes an interpretability model to understand and explain the decisions of deep learning models. The use DeepLIFT is proposed attributions a study case are obtained. Benchmarking against two other models, namely Grad-CAM dCAM, conducted. Results show that can provide precise inputs in both temporal spatial directions. A parametric also carried out effects structure on obtained from model. Ten different convolutional neural network structures considered. Three important observations made: 1) changes have minor direction, but 2) they negligible 3) layers need be fixed avoid attribution discrepancies. By understanding decision resulting structure, it hoped this contribute development trustworthy models for fire research community.

Language: Английский

Forecasting backdraft with multimodal method: Fusion of fire image and sensor data DOI
Tianhang Zhang, Fangqiang Ding, Zilong Wang

et al.

Engineering Applications of Artificial Intelligence, Journal Year: 2024, Volume and Issue: 132, P. 107939 - 107939

Published: Jan. 27, 2024

Language: Английский

Citations

8
FireDM: A weakly-supervised approach for massive generation of multi-scale and multi-scene fire segmentation datasets DOI
Hongtao Zheng, Meng Wang, Zilong Wang

et al.

Knowledge-Based Systems, Journal Year: 2024, Volume and Issue: 290, P. 111547 - 111547

Published: Feb. 20, 2024

Language: Английский

Citations

8
Modelling flame-to-fuel heat transfer by deep learning and fire images DOI Creative Commons
Caiyi Xiong, Zilong Wang, Xinyan Huang

et al.

Engineering Applications of Computational Fluid Mechanics, Journal Year: 2024, Volume and Issue: 18(1)

Published: March 21, 2024

In numerical fire simulations, the calculation of thermal feedback from flame to solid and liquid fuel surface plays a critical role as it connects fundamental gas-phase burning condensed-phase gasification. However, is computationally intensive task in CFD modelling methods because requirement high-resolution grid for calculating interface heat transfer. This paper proposed real-time prediction flame-to-fuel transfer by using simulated images computer-vision deep learning method. Different methanol pool fires were selected produce image database training model. As diameters increase 20 40 cm, dominant shifts convection radiation. Results show that AI algorithm trained can predict both convective radiative flux distributions on condensed with relative error below 20%, based input morphology be captured larger size. Regardless growing or decaying puffing flames induced buoyancy, this method further non-uniform distribution coefficient rather than empirical correlations. work demonstrates use computer vision accelerating simulation, which helps simulate complex behaviours simpler models smaller computational costs.

Language: Английский

Citations

5
Development of an early-stage thermal runaway detection model for lithium-ion batteries DOI
Wai Cheong Tam, Jian Chen, Hongqiang Fang

et al.

Journal of Power Sources, Journal Year: 2025, Volume and Issue: 641, P. 236714 - 236714

Published: March 23, 2025

Language: Английский

Citations

0
SegLD: Achieving universal, zero-shot and open-vocabulary segmentation through multimodal fusion via latent diffusion processes DOI
Hongtao Zheng, Yifei Ding, Zilong Wang

et al.

Information Fusion, Journal Year: 2024, Volume and Issue: 111, P. 102509 - 102509

Published: June 5, 2024

Language: Английский

Citations

2
FlareNet: A Feature Fusion Based Method for Fire Detection under Diverse Conditions DOI
Balal Yousaf, Adeel Feroz Mirza, Muhammad Irfan

et al.

Research Square (Research Square), Journal Year: 2024, Volume and Issue: unknown

Published: Oct. 3, 2024

Abstract Fire detection is crucial for safeguarding lives and property. In recent years, advancements in deep learning Internet of Things (IoT) architectures have significantly enhanced the accuracy fire smoke systems. this study, we introduce FlareNet, a feature fusion based model that leverages DenseNet architecture combined with Spatial Pyramid Pooling (SPP) Contextual Feature Network (CFPN). FlareNet further augmented dual attention mechanisms Enhancement Attention (FEA) mechanism to selectively emphasize critical features distinguishing between non-fire scenes. Our proposed rigorously evaluated across five diverse datasets: Sharma, Deep Quest, BoWFire, FD dataset, our novel MixFire achieving an impressive average 99.2%. A comparative evaluation against state-of-the-art (SOTA) algorithms reveals outperforms existing methods notable improvement accuracy, precision, recall, F1-score, thereby setting new benchmark domain detection. Furthermore, comprehensive analysis baseline models such as VGG16, VGG19, ResNet18, MobileNetV2, also presented. These underscore FlareNet’s capability enhance systems more sustainable environment. code dataset can be accessed by https://github.com/adeelferozmirza/FlareNet.

Language: Английский

Citations

1
Hyper real-time flame detection: Dynamic insights from event cameras and FlaDE dataset DOI

Saizhe Ding,

Haorui Zhang, Yuan‐Ting Zhang

et al.

Expert Systems with Applications, Journal Year: 2024, Volume and Issue: 263, P. 125746 - 125746

Published: Nov. 17, 2024

Language: Английский

Citations

0
Examining the Effects of Deep Learning Model Structure on Model Interpretability for Time-Series Classifications in Fire Research DOI Open Access
Wai Cheong Tam, Linhao Fan, Qi Tong

et al.

Journal of Physics Conference Series, Journal Year: 2024, Volume and Issue: 2885(1), P. 012097 - 012097

Published: Nov. 1, 2024

Abstract This present work utilizes an interpretability model to understand and explain the decisions of deep learning models. The use DeepLIFT is proposed attributions a study case are obtained. Benchmarking against two other models, namely Grad-CAM dCAM, conducted. Results show that can provide precise inputs in both temporal spatial directions. A parametric also carried out effects structure on obtained from model. Ten different convolutional neural network structures considered. Three important observations made: 1) changes have minor direction, but 2) they negligible 3) layers need be fixed avoid attribution discrepancies. By understanding decision resulting structure, it hoped this contribute development trustworthy models for fire research community.

Language: Английский

Citations

0