Scholarlink

<p>Analyzing the complex speckle patterns from Fiber Specklegram Sensors (FSS) for physical parameter estimation remains a significant challenge. This study proposes and demonstrates the application of a Vision Transformer (ViT) architecture for high-accuracy temperature regression using a benchmark synthetic FSS dataset. The proposed model achieves state-of-the-art performance, yielding a Root Mean Square Error (RMSE) of 0.511 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(^{\circ }\)</EquationSource> </InlineEquation>C and a coefficient of determination (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>) of 0.9998. This result significantly outperforms other established deep learning models, including various CNN and Transformer-based approaches. To ensure the model’s reliability, an Explainable AI (XAI) analysis using Integrated Gradients was conducted. The analysis confirmed that the ability of the model to associate predictions with physically relevant structures in the speckle patterns.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable vision transformer based temperature sensing with fiber specklegram sensors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月10日

查看全文
  • Serif Ali Sadik

摘要

Analyzing the complex speckle patterns from Fiber Specklegram Sensors (FSS) for physical parameter estimation remains a significant challenge. This study proposes and demonstrates the application of a Vision Transformer (ViT) architecture for high-accuracy temperature regression using a benchmark synthetic FSS dataset. The proposed model achieves state-of-the-art performance, yielding a Root Mean Square Error (RMSE) of 0.511 \(^{\circ }\) C and a coefficient of determination ( \(R^2\) ) of 0.9998. This result significantly outperforms other established deep learning models, including various CNN and Transformer-based approaches. To ensure the model’s reliability, an Explainable AI (XAI) analysis using Integrated Gradients was conducted. The analysis confirmed that the ability of the model to associate predictions with physically relevant structures in the speckle patterns.