Industrial manufacturing systems increasingly rely on large-scale sensor networks to monitor equipment states and production processes. For example, in continuous casting, casting-speed signals are continuously acquired by encoders and transmitted to a programmable logic controller (PLC) or a distributed control system (DCS) through industrial sensing links. These sensor-generated time-series signals provide essential information for intelligent perception and fault diagnosis, yet often lack high-quality annotations. To address this issue, in this paper, we proposed the Industrial Deep Time Series Detection-Q-Network (IDTS-DQN), a semi-supervised deep reinforcement learning algorithm, for anomaly detection in industrial sensing systems. The method integrates deep learning and reinforcement learning through a dual reward mechanism and an anomaly-based experience replay strategy, enabling the agent to learn temporal patterns with limited labeled samples and dynamically improve detection performance. Extensive experiments on real-world industrial datasets demonstrate that IDTS-DQN outperforms baseline methods in recall, precision, and F1-score, providing an adaptive sensing-data analysis approach for industrial monitoring and demonstrating the potential of deep reinforcement learning in sensor-based anomaly perception. However, the framework has been evaluated only offline on historical sensing data, and its applicability to broader signal types and fault modes remains to be examined. Future work will focus on multimodal sensor fusion, edge-based online deployment, and domain-knowledge integration for industrial sensor-signal analysis.