Model Drift Early Warning Method Based on Operational Monitoring Data Sensing
Hongxia Guo, Peng Wang, Xi Chen, Y. Sheng, X. Li, Cuicui Wang, Fangfang Qin, Shan Gao · 2025
A model drift early warning method based on operation monitoring data sensing is constructed to address the problem of performance degradation in task scheduling models over time, which affects the rational allocation of communication front-end resources. Firstly, by analyzing data characteristics, the performance monitoring indicators of the task scheduling model are screened and their thresholds are defined. When the monitoring indicators of the task scheduling model are abnormal, drift detection is performed on the input data. In performing drift detection, the old and new data distributions are obtained, and the model output is calibrated using an unsupervised calibration method to output the predicted confidence value. Subsequently, the data frequency histogram is constructed, and the difference between the old and new data distributions can be obtained through the KL dispersion, so as to determine whether model drift has occurred. Finally, when the model drift is confirmed, timely warning should be given, and the Salp Swarm algorithm should be used to optimize the parameters and reconstruct the model to adapt to the new data distribution. The experimental results show that the proposed method can effectively achieve model drift detection and early warning, improve the continuous availability of task scheduling models, and assist power enterprises in optimizing and reconstructing the allocation of front-end communication resources.