Industries increasingly demand intelligent monitoring for data anomalies in industrial sensing systems. However, traditional threshold-based sensor data anomaly detection remains prevalent; a rapid shift to fully AI-enhanced systems is often impractical. Instead, incrementally optimizing traditional methods is more feasible. In this study, we enhance traditional dynamic threshold techniques for sensor time series data using large model-driven recommendations and parameter optimization. Our method first relabels and corrects samples from existing threshold outputs to recalculate appropriate values. Through a time-segmented split mechanism, it dynamically adjusts threshold ratios within each segment to better reflect sensor data variations. Additionally, we incorporate a large language model to recommend optimal parameter configurations, improving usability and flexibility. Real-world deployment demonstrates that our approach significantly outperforms traditional dynamic thresholds in detection accuracy and adaptability, offering a scalable solution for transitioning toward intelligent sensor monitoring. While effectively addressing the rigidity and high false-alarm rates of traditional sensor baselines, this study’s limitation is its reliance on human validation for complex sensor drift. The remaining challenge is achieving the fully autonomous handling of multi-sensor correlations.