This letter proposes a multimodal data-driven reinforcement learning (RL) method for operational decision-making in industrial processes. The method, named feedstock-guided multimodal actor-critic (FGM-AC), integrates feedstock properties and multimodal data (e.g., images, audio, sensor data) into the state space to enhance decision-making accuracy and adaptability. The FGM-AC algorithm is designed to address the challenges of feedstock fluctuations and underutilized multimodal data in industrial processes. The method uses a reward function and state representation tailored to handle the complexity of multimodal data, improving the performance of the RL framework. The proposed method is applied to an industrial flotation process, where the goal is to optimize the concentration and grade of flotation froth. The state space includes operational conditions, feedstock conditions, and the target grade of flotation froth. The reward function is designed to maximize froth concentration while meeting grade specifications. The FGM-AC algorithm is trained using a combination of actor and critic networks, with experience replay to improve learning efficiency. Experiments on real industrial data show that the FGM-AC method outperforms manual operations and other RL-based methods (e.g., DQN, AC) in improving froth concentration and grade. The results demonstrate the effectiveness of the FGM-AC method in optimizing industrial processes. The method's ability to utilize multimodal data and adapt to feedstock fluctuations makes it a promising solution for industrial decision-making. Future work will focus on enhancing the security of online RL algorithms and extending the method to other industrial processes with available multimodal data.This letter proposes a multimodal data-driven reinforcement learning (RL) method for operational decision-making in industrial processes. The method, named feedstock-guided multimodal actor-critic (FGM-AC), integrates feedstock properties and multimodal data (e.g., images, audio, sensor data) into the state space to enhance decision-making accuracy and adaptability. The FGM-AC algorithm is designed to address the challenges of feedstock fluctuations and underutilized multimodal data in industrial processes. The method uses a reward function and state representation tailored to handle the complexity of multimodal data, improving the performance of the RL framework. The proposed method is applied to an industrial flotation process, where the goal is to optimize the concentration and grade of flotation froth. The state space includes operational conditions, feedstock conditions, and the target grade of flotation froth. The reward function is designed to maximize froth concentration while meeting grade specifications. The FGM-AC algorithm is trained using a combination of actor and critic networks, with experience replay to improve learning efficiency. Experiments on real industrial data show that the FGM-AC method outperforms manual operations and other RL-based methods (e.g., DQN, AC) in improving froth concentration and grade. The results demonstrate the effectiveness of the FGM-AC method in optimizing industrial processes. The method's ability to utilize multimodal data and adapt to feedstock fluctuations makes it a promising solution for industrial decision-making. Future work will focus on enhancing the security of online RL algorithms and extending the method to other industrial processes with available multimodal data.