Enhanced Time-Series Activity Recognition with a Hybrid CNN-SVM Model: Insights from Wearable Sensor Analysis

The fundamental requirement of time-series activity detection that uses wearable sensors spreads across healthcare monitoring alongside fitness tracking and smart environments applications. The proposed model fuses CNN together with SVM for improving the precision and dependability of activity recognition procedures. The model first feeds sensor signal data to CNN to extract spatial-temporal features automatically before utilizing SVM for final class determination. The UCI Human Activity Recognition dataset serves as the test platform for the proposed method where it evaluates 10,299 samples belonging to six activity classes. The CNN-SVM combination model achieves superiority over conventional models based on experimental testing by producing 95.6% overall accuracy which exceeds CNN at 91.4% and LSTM at 89.7% as well as CNN-LSTM at 92.5%. The model achieves a precision value of 95.3% together with recall at 95.5% and calculates an F1-score of 95.4% while maintaining high performance across all activity classes. Walking received accurate classification 520 times from a total of 527 samples along with sitting that achieved precise identification in 370 out of 374 cases. Conducting an ablation study shows that combining CNN-extracted features with SVM results in better accuracy performance which surpasses standalone CNN by 4.2% as well as standalone SVM using handcrafted features by 7.1%. Computational evaluation of the model shows training duration at 150 seconds and inference per sample operation taking 2.7 milliseconds which makes the system suitable for real-time use. The hybrid CNN-SVM model demonstrates its potential value as an effective solution for precise time-series activity detection through wearable devices.