<p>Fault diagnosis in complex industrial systems plays a critical role in ensuring operational reliability, improving predictive maintenance, and enhancing energy efficiency. In cement rotary kilns, continuous thermal monitoring is essential to maintain process stability, reduce energy losses, and prevent unexpected shutdowns caused by abnormal temperature distributions. This work introduces a hybrid CNN–DDQN framework that transforms thermal fault classification in cement rotary kilns from a static prediction task into an adaptive decision-making process, achieving superior accuracy and robustness on real industrial data. The proposed approach integrates a convolutional neural network (CNN) architecture for robust feature extraction, with a double deep Q-network (DDQN) reinforcement learning model to enhance adaptive fault classification and decision-making. The models were trained and validated using labeled thermal data collected from a real cement production environment comprising multiple fault categories. Comparative evaluation was conducted using classification accuracy and performance efficiency metrics, and compared with the CNN and residual neural network (ResNet) models. The results demonstrate that the DDQN-based model outperforms the conventional CNN architecture in terms of fault classification accuracy and adaptability to thermal variations, highlighting the advantage of integrating deep reinforcement learning with convolutional feature extraction. The proposed framework provides an intelligent and scalable solution for kiln condition monitoring, contributing to improved preventive maintenance strategies, reduced fault occurrence, extended equipment lifespan, and enhanced operational and economic performance in rotary kiln-based industrial systems.</p>