This letter introduces a multimodal data-driven reinforcement learning (RL) method, named feedstock-guided multimodal actor-critic (FGM-AC), for operational decision-making in industrial processes. The method addresses the challenges of frequent feedstock fluctuations and underutilized multimodal data (images, audio, and sensor data) by incorporating these factors into the state space. The FGM-AC algorithm aims to enhance the adaptability and accuracy of decision-making strategies, leveraging the overall perception of operators. The reward function and state representation are designed to handle the complexity and specific characteristics of multimodal data, improving the performance of the RL framework. The effectiveness of the proposed method is validated through extensive experiments on actual industrial data, demonstrating its potential to provide accurate and dependable strategies for decision-making. The results show that the FGM-AC algorithm outperforms manual operations and existing RL-based methods, achieving significant improvements in froth concentration and grade. Future work will focus on enhancing the security of online RL algorithms and extending the method to other industrial processes.This letter introduces a multimodal data-driven reinforcement learning (RL) method, named feedstock-guided multimodal actor-critic (FGM-AC), for operational decision-making in industrial processes. The method addresses the challenges of frequent feedstock fluctuations and underutilized multimodal data (images, audio, and sensor data) by incorporating these factors into the state space. The FGM-AC algorithm aims to enhance the adaptability and accuracy of decision-making strategies, leveraging the overall perception of operators. The reward function and state representation are designed to handle the complexity and specific characteristics of multimodal data, improving the performance of the RL framework. The effectiveness of the proposed method is validated through extensive experiments on actual industrial data, demonstrating its potential to provide accurate and dependable strategies for decision-making. The results show that the FGM-AC algorithm outperforms manual operations and existing RL-based methods, achieving significant improvements in froth concentration and grade. Future work will focus on enhancing the security of online RL algorithms and extending the method to other industrial processes.