基于时序窗口Transformer的冷轧连轧机电机异常预警方法研究OA

To address the challenges of sparse motor anomaly samples,time-varying operating conditions,and temporal boundary overlap in short-term early warning for cold rolling tandem mills,this paper proposes a short-term motor anomaly early warning method under strict temporal isolation.Based on a benchmark dataset from a five-stand cold rolling tandem mill,the method first partitions the training,validation,and test sets along the original timeline,sets isolation gaps between adjacent data blocks to avoid window sharing caused by global sliding windows prior to splitting,and then constructs temporal window samples within each block.Using motor power,torque,rolling speed,tension,and strip specification variables from the most recent 10 samples,the model predicts whether a motor anomaly will occur in the next sample,with Logistic regression serving as an interpretable linear baseline.Experimental results show that on the strict temporal isolation test set,the temporal window Transformer achieves an F1 score of 0.802 7,outperforming Logistic regression at 0.729 7;precision improves from 0.760 6 to 0.842 9,recall from 0.701 3 to 0.766 2,and the area under the precision-recall curve from 0.717 0 to 0.768 7.Further experiments with extended prediction horizons reveal that when the horizon increases to 3,both models exhibit significant performance degradation;the Transformer still surpasses Logistic regression in precision,recall,and F1 score,but ROC-AUC becomes comparable while PR-AUC falls below that of Logistic regression.These findings indicate that the temporal window Transformer holds practical value for near-horizon motor anomaly early warning,yet further validation incorporating stand structure information,trend features,and field data is required for far-horizon applications.