With the widespread adoption of embedded sensors

human activity recognition (HAR) technology achieves significant progress.Compared with wearable sensors

smartphones become an ideal platform for HAR monitoring due to their built-in multi-sensor capabilities and high penetration rate.To address challenges in practical applications

including limited labeled samples

underutilized sensors

and imbalanced class distributions

a smartphone-sensor-based HAR method is proposed.By employing a multidimensional feature enhancement mechanism that integrates human attributes and combines it with a recursive feature elimination (RFE) algorithm based on random forest(RF)

highly discriminative features are selected and input into a stacked learning model that fuses extreme gradient boosting (XGBoost) and a one-dimensional convolutional neural network (1DCNN)

thereby achieving recognition of 12 daily activities.Cross-validation results indicate an average recognition accuracy of 96.92%

which validates the method’s superiority in handling complex activities and individual differences

and provides both theoretical innovation and practical value for smartphone-sensor-based HAR methods.