This article presents a systematic review of knowledge-driven deep learning (DL) in wireless networks, focusing on its application in 6G network optimization. As 6G networks aim to support diverse services in dynamic environments, they require efficient management of multi-dimensional heterogeneous resources. Traditional model-based methods face challenges due to computational complexity and long processing times, while data-driven DL methods rely heavily on training data and lack interpretability. To address these issues, knowledge-driven DL integrates proven domain knowledge into neural networks, enhancing interpretability and reducing the need for large training datasets. The article proposes a holistic framework for knowledge-driven DL in wireless networks, where knowledge sources, representation, integration, and application form a closed loop. It introduces three main knowledge integration approaches: knowledge-assisted, knowledge-fused, and knowledge-embedded DL. Knowledge-assisted DL uses domain knowledge to guide model selection and loss function design. Knowledge-fused DL combines model-based methods with data-driven neural networks to solve complex problems. Knowledge-embedded DL integrates domain knowledge into neural network structures, improving interpretability and performance. The article also discusses open issues in knowledge-driven DL, including handling complex constraints, theoretical analysis, and knowledge selection and aggregation. A case study demonstrates the effectiveness of knowledge-driven DL in resource management, showing its scalability and performance advantages over traditional methods. The study proposes a novel knowledge-driven DL approach for resource management in device-to-device (D2D) networks, integrating graph-structured communication topology and the WMMSE algorithm into neural networks. This approach, known as the WMMSE algorithm unrolled GNN (WUGNN), outperforms other methods in terms of sum rate and online inference time, demonstrating its potential for efficient 6G network optimization. The article concludes that knowledge-driven DL offers a promising paradigm for intelligent 6G networks, combining the strengths of domain knowledge and deep learning.This article presents a systematic review of knowledge-driven deep learning (DL) in wireless networks, focusing on its application in 6G network optimization. As 6G networks aim to support diverse services in dynamic environments, they require efficient management of multi-dimensional heterogeneous resources. Traditional model-based methods face challenges due to computational complexity and long processing times, while data-driven DL methods rely heavily on training data and lack interpretability. To address these issues, knowledge-driven DL integrates proven domain knowledge into neural networks, enhancing interpretability and reducing the need for large training datasets. The article proposes a holistic framework for knowledge-driven DL in wireless networks, where knowledge sources, representation, integration, and application form a closed loop. It introduces three main knowledge integration approaches: knowledge-assisted, knowledge-fused, and knowledge-embedded DL. Knowledge-assisted DL uses domain knowledge to guide model selection and loss function design. Knowledge-fused DL combines model-based methods with data-driven neural networks to solve complex problems. Knowledge-embedded DL integrates domain knowledge into neural network structures, improving interpretability and performance. The article also discusses open issues in knowledge-driven DL, including handling complex constraints, theoretical analysis, and knowledge selection and aggregation. A case study demonstrates the effectiveness of knowledge-driven DL in resource management, showing its scalability and performance advantages over traditional methods. The study proposes a novel knowledge-driven DL approach for resource management in device-to-device (D2D) networks, integrating graph-structured communication topology and the WMMSE algorithm into neural networks. This approach, known as the WMMSE algorithm unrolled GNN (WUGNN), outperforms other methods in terms of sum rate and online inference time, demonstrating its potential for efficient 6G network optimization. The article concludes that knowledge-driven DL offers a promising paradigm for intelligent 6G networks, combining the strengths of domain knowledge and deep learning.