In order to solve the anomaly problem in the carbon emission monitoring data of the sugar industry, based on its characteristics of strong timing and multi-source correlation, this paper proposes a set of anomaly detection and repair framework of “feature construction—dual module collaboration—decision verification”. By constructing a multi-dimensional feature system, the improved time-series isolated forest (TS-IF) is used for anomaly detection, combined with sliding window weighted interpolation (SW-WI) to achieve data repair, and finally the reliable data is output through dynamic threshold verification. Experiments show that the framework can effectively identify and repair anomalies in both synthetic and real scenes, the F1 score of anomaly detection is 0.92, and the repair error is controlled within 5%. The framework is easy to adjust parameters, low complexity and strong reproducibility, which provides theoretical support for the quality control of carbon emission data of sugar enterprises.