<p>Fused Magnesium Furnaces (FMFs) are critical energy-intensive equipment, where abnormal detection is essential for efficient, stable, and safe operation. Recently, multimodal data have been increasingly used for diagnosis and monitoring. However, existing approaches still suffer from two major challenges: (1) the insufficient fusion of heterogeneous modalities, which limits the ability to capture informative cross-modal interactions, and (2) the uncertainty inherent in multimodal feature representations, which undermines the reliability and interpretability of the decision process. To overcome these limitations, we propose a novel deep multimodal fuzzy neural network (DMFNN) for abnormal condition recognition in FMFs, by jointly leveraging 1-D current signals and 3-D video data. The framework first employs a Transformer-based encoder with tokenization to produce uniform representations, and a bidirectional cross-attention mechanism to explicitly model cross-modal dependencies. To handle uncertainty and enhance interpretability, a neuro-fuzzy classifier is introduced for the final decision-making stage. Experimental results on a benchmark dataset demonstrate the superiority of our approach over existing techniques, achieving an accuracy of 93.51%, an F1 score of 94.98%, a false detection rate of 5.58%, and a missed detection rate of 1.90%. Beyond FMFs, the proposed framework holds potential for broader applications in multimodal anomaly detection in industrial scenarios. Code released at <a href="https://github.com/xichaoo/DMFNN.">https://github.com/xichaoo/DMFNN.</a></p>