Scholarlink

<p>This study developed a deep learning-based multimodal data fusion framework to address the challenge of real-time monitoring in solid-state fermentation (SSF), using lactic acid bacteria-fermented barley bran as the substrate. The framework synchronously acquires multiple data sources, including physicochemical parameters (pH, moisture, temperature, pressure, agitation speed, and fermentation time) and color image data. Deep texture and color features were extracted from the images using a pre-trained ResNet50 model. These features were integrated with temporal dynamics captured by a Long Short-Term Memory (LSTM) network to build a ResNet50-LSTM fusion model for predicting the evolution of key bioactive compounds during fermentation. Experimental results showed that the fusion model achieved a determination coefficient (R²) of 0.9244 and a root mean square error (RMSE) of 0.6802 on the test set. Compared to unimodal methods using only image or physicochemical data, prediction errors were reduced by 58.89% to 236.05%. This improvement highlights the ability of multimodal data fusion to capture complex nonlinear relationships in SSF by combining complementary visual and sensor information, thereby overcoming the limitations of conventional single-mode monitoring. The study demonstrates the potential of deep learning-driven multimodal fusion for intelligent monitoring and control of SSF processes, offering a methodological reference for real-time fermentation optimization and high-value biomass utilization.</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

Multimodal data fusion and deep learning for predicting phenolics dynamics in barley bran solid-state fermentation

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Yansheng Zhao,
  • Pengfei Jin,
  • Guoting Xiong,
  • Jiapeng Cai,
  • Juan Bai,
  • Chunhua Ding,
  • Xiang Xiao

摘要

This study developed a deep learning-based multimodal data fusion framework to address the challenge of real-time monitoring in solid-state fermentation (SSF), using lactic acid bacteria-fermented barley bran as the substrate. The framework synchronously acquires multiple data sources, including physicochemical parameters (pH, moisture, temperature, pressure, agitation speed, and fermentation time) and color image data. Deep texture and color features were extracted from the images using a pre-trained ResNet50 model. These features were integrated with temporal dynamics captured by a Long Short-Term Memory (LSTM) network to build a ResNet50-LSTM fusion model for predicting the evolution of key bioactive compounds during fermentation. Experimental results showed that the fusion model achieved a determination coefficient (R²) of 0.9244 and a root mean square error (RMSE) of 0.6802 on the test set. Compared to unimodal methods using only image or physicochemical data, prediction errors were reduced by 58.89% to 236.05%. This improvement highlights the ability of multimodal data fusion to capture complex nonlinear relationships in SSF by combining complementary visual and sensor information, thereby overcoming the limitations of conventional single-mode monitoring. The study demonstrates the potential of deep learning-driven multimodal fusion for intelligent monitoring and control of SSF processes, offering a methodological reference for real-time fermentation optimization and high-value biomass utilization.