Deep Multimodal Data Fusion is a research area that involves integrating data from multiple sources, such as images, text, and sensor data, to improve decision-making. Traditional methods of data fusion, such as early/late fusion, are no longer suitable for the modern deep learning era. Instead, a new fine-grained taxonomy has been proposed, grouping state-of-the-art models into five categories: Encoder-Decoder methods, Attention Mechanism methods, Graph Neural Network methods, Generative Neural Network methods, and other Constraint-based methods. This survey covers a broader combination of modalities, including Vision + Language, Vision + Sensors, and their corresponding tasks, such as video captioning and object detection. It also provides a comparison among these methods, as well as challenges and future directions in this area. The survey highlights the importance of multimodal data fusion in various applications, such as autonomous vehicles and medical imaging, where combining data from different modalities can lead to more accurate and reliable results. The survey also discusses the evolution of AI and the role of multimodal data fusion in achieving more robust and accurate models. The survey emphasizes the need for more sophisticated fusion methods that can automatically learn complementary and redundant information from multimodal data. It also discusses the challenges of multimodal data fusion, such as the need for large amounts of data and the complexity of integrating data from different sources. The survey concludes that the development of deep learning has significantly reshaped the landscape of multimodal data fusion, revealing the inadequacies of traditional fusion methods and the need for more advanced techniques. The survey provides a comprehensive review of deep multimodal data fusion, categorizing models into five classes and discussing the latest advances in this field.Deep Multimodal Data Fusion is a research area that involves integrating data from multiple sources, such as images, text, and sensor data, to improve decision-making. Traditional methods of data fusion, such as early/late fusion, are no longer suitable for the modern deep learning era. Instead, a new fine-grained taxonomy has been proposed, grouping state-of-the-art models into five categories: Encoder-Decoder methods, Attention Mechanism methods, Graph Neural Network methods, Generative Neural Network methods, and other Constraint-based methods. This survey covers a broader combination of modalities, including Vision + Language, Vision + Sensors, and their corresponding tasks, such as video captioning and object detection. It also provides a comparison among these methods, as well as challenges and future directions in this area. The survey highlights the importance of multimodal data fusion in various applications, such as autonomous vehicles and medical imaging, where combining data from different modalities can lead to more accurate and reliable results. The survey also discusses the evolution of AI and the role of multimodal data fusion in achieving more robust and accurate models. The survey emphasizes the need for more sophisticated fusion methods that can automatically learn complementary and redundant information from multimodal data. It also discusses the challenges of multimodal data fusion, such as the need for large amounts of data and the complexity of integrating data from different sources. The survey concludes that the development of deep learning has significantly reshaped the landscape of multimodal data fusion, revealing the inadequacies of traditional fusion methods and the need for more advanced techniques. The survey provides a comprehensive review of deep multimodal data fusion, categorizing models into five classes and discussing the latest advances in this field.