Research on cigarette moisture anomaly risk identification based on an improved NGBoost algorithm

Once quality abnormalities such as tobacco mildew and excessive moisture occur in cigarette production, it often necessitates the lockdown or even scrapping of a significant portion of the inventory. This can result in widespread market complaints. Therefore, timely identification of moisture anomalies in cigarettes is of paramount importance to adjust relevant parameters or operational processes for subsequent batches promptly. This paper proposes a method for identifying cigarette moisture anomaly risks based on an improved NGBoost algorithm. This method focuses on the moisture content of finished products, involves cleansing time-series data of moisture chain-related influencing parameters, extracting feature parameters using SHAP Value, and ultimately establishing a moisture prediction model using NGBoost. Trend analysis is conducted on the residuals between predicted and actual values on a weekly basis. A change in trend in the residuals serves as a timely alert for moisture anomalies. The results indicate that in 2023, the model identified moisture anomaly risks a total of 18 times, with 14 confirmed as actual risky states. There were 6 instances of false positives. The identification accuracy reached 77.8%, effectively mitigating the quality risks associated with moisture anomalies.