电熔镁炉多模态信息工况智能识别方法研究

Research on intelligent recognition method of multimodal information working condition in fused magnesium furnace

中文关键词  电熔镁炉  多模态  强化学习  Transformer  工况识别

英文关键词  fused magnesium furnace  multi-modal  reinforcement learning  Transformer  working condition recognition

中文摘要

      针对电熔镁炉识别过程中出现的模态完备性不足问题, 本文阐述了一种电熔镁炉多模态信息工况智能识 别方法. 首先, 通过语义神经网络提取工况图像特征和双向编码语言模型提取语言特征, 构建多模态工况的完备联 合特征向量, 并通过Transformer编码层实现全局交互, 捕捉视觉与语言信息的细粒度对齐. 引入自适应Transformer 解码层的自注意力机制, 采用全连接网络获得多模态工况识别结果. 基于强化学习定义门控单元评估策略, 实时评 估不确定工况识别结果, 构建解码层的动态调节机制, 以获取多模态工况的细粒度特征, 并采用模糊积分集成模型 库的识别结果. 实验证明了本方法的有效性和鲁棒性.

英文摘要

      Aiming at the lack of modal completeness in the identification process of fused magnesium furnace, an intelligent recognition method of multi-modal information working condition of electric smelting magnesia furnace is described in this paper. Firstly, we use semantic neural network to extract image features and two-way coding language model to extract linguistic features to construct a complete joint feature vector of multimodal working conditions, and then we realize global interaction through the Transformer encoding layer to capture the fine-grained alignment of visual and linguistic information. The self-attention mechanism of the adaptive Transformer decoding layer is introduced to obtain the multimodal work condition recognition results using fully connected networks. Based on reinforcement learning, we define the evaluation strategy of gating unit, evaluate the recognition results of uncertain working conditions in real time, construct the dynamic adjustment mechanism of the decoding layer to obtain the fine-grained features of multimodal working conditions, and use the fuzzy integration to integrate the recognition results of the model library. Experiments demonstrate the effectiveness and robustness of this method.