The initial stage of semiconductor packaging, wafer grinding, relies critically on in-process sensors to enable the precise thinning of wafers to a target thickness. Specifically, inner and outer in-process gauges (IPGs) provide real-time height measurements that dictate the vertical positioning of grinding wheels. However, the accuracy of these outer height sensors becomes compromised as wafers become increasingly thin owing to stress-induced lifting on the unground side. This sensor inaccuracy can lead to erroneous feedback, resulting in excessive grinding pressure and a heightened risk of wafer cracking or breakage. The height data acquired from these sensors are presented as variable-length, short-term time series, which are affected by fluctuating production demands. Furthermore, machine recalibration based on individual chuck table heights complicates consistent monitoring using absolute sensor readings. To address these sensor-related challenges in acquiring and interpreting reliable height information, we introduce a novel approach utilizing the accumulated generating operation (AGO) to convert the nonpatterned time series data into discernible patterns, effectively reframing the problem as a pattern recognition task based on sensor data analysis. Addressing the inherent class imbalance due to the infrequency of anomalies (246 abnormal instances within a dataset of 55143), the bootstrap method is employed for data balancing. Subsequently, a bagging ensemble of five back-propagation neural networks is trained for anomaly detection using the transformed sensor data. Experimental results showcase a substantial improvement in the identification of abnormal height patterns in these short-term time series derived from sensor measurements, with the F1-score increasing significantly from 0.201 to 0.871. These findings underscore the efficacy of the proposed methodology in enhancing wafer grinding process monitoring and anomaly detection by effectively handling the complexities of in-process sensing.