Deep-Learning-Based Flare-Smoke Detection Allows Real-Time Application

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 220857, “Real-Time Application of Deep-Learning-Based Flare-Smoke Detection,” by Kratika Gupta, Purba Chatterjee, SPE, and Xianzhong (John) Chen, ExxonMobil, et al. The paper has not been peer reviewed. _ This work focuses on developing a smoke-detection model that can be used in an automated, end‑to‑end flare‑smoke‑detection, alerting, and distributed-control solution that leverages existing closed-circuit television (CCTV) cameras at manufacturing facilities. At the core of this solution is a deep-learning computer-vision model that leverages an extensive and diverse data set. The smoke‑ detection model uses a novel approach for real-time detection of hydrocarbon emission, using semantic segmentation through compact vision transformers (ViTs). Introduction The authors propose a novel application of efficient ViTs for predicting smoke regions in flare images. Because smoke detection is formulated as a semantic segmentation task, crisper boundaries and more-accurate predictions for smoke areas and densities are obtained. Use of ViTs over convolutional neural networks (CNNs) enables a global receptive field, which is particularly important when considering large-resolution images. A lightweight ViT architecture is used in this work that features hardware‑efficient operations, resulting in latency reduction and support of model inference in real time. Methods Lightweight Multiscale Attention (MSA). This module uses lightweight rectified‑linear‑unit (ReLU)-based global attention to enable a global receptive field with linear computational complexity. The similarity function generally used in transformers is replaced by a ReLU-based similarity described in the complete paper. Because the computationally expensive exponential function is replaced by hardware-efficient ReLU operation, the latency involved in the attention mechanism is significantly reduced. The difference in latency between softmax attention and ReLU‑based linear attention has a significant effect on real‑world applications. ReLU‑based linear attention is from 3.3 to 4.5 times faster than softmax attention because of the elimination of hardware‑unfriendly exponential operations. Such efficiency is crucial for applications with limited hardware resources. For instance, on a Qualcomm Snapdragon 855 central processing unit using TensorFlow Lite, ReLU‑based linear attention is markedly faster than softmax attention. This makes it especially suitable for mobile and edge devices where low latency is essential. Generating Multiscale Tokens. The ReLU-based attention mechanism has limited capacity, which is addressed by enhancing it with small-kernel depthwise convolution to improve local information extraction. To boost multiscale learning, information is aggregated from nearby query, key, and value tokens into multiclass tokens. This process is performed independently for each query, key, and value token in each head, using small‑kernel convolutions for efficiency. To improve graphics‑processing‑unit (GPU) efficiency, group convolution infrastructure is used to combine depthwise and 1×1 convolutions into a single operation. After obtaining multiscale tokens, global attention is applied to extract multiscale global features, which are then concatenated along the head dimension and fed to a final linear projection layer to merge the features.