Mental workload (MWL) is the cognitive effort required to manage information in working memory. Excessive MWL increases the error risk and, when prolonged, can impair appetite, sleep, and overall health. Therefore, an objective and real-time MWL estimation is crucial. In this study, we introduce an MWL estimation method during floor-cleaning tasks using inertial sensor data collected from the body and cleaning tools. We introduce conventional statistical features from inertial sensor signals and two types of feature derived from symbol sequences via vector quantization. We construct regression models to estimate MWL and compare their errors using various combinations of these three feature types. The models consistently achieve lower errors than a naive baseline, which always predicts the training data median. We also compare results from different perspectives, such as sensor placement in each scenario and the computation time required for feature extraction. The findings suggest that the proposed approach has practical potential for daily monitoring and visualization of MWL.