Artificial Intelligence Based Approach for Classification of Human Activities Using MEMS Sensors Data.

Area of Science:

• Computer Science
• Biomedical Engineering
• Signal Processing

Background:

• Micro Electronic Mechanical Systems (MEMS) sensors in smartphones enable advanced Human Activity Recognition (HAR).
• Machine Learning (ML) techniques are crucial for classifying human motion activities using sensor data.
• Existing wearable technologies often rely on traditional algorithmic approaches for activity detection.

Purpose of the Study:

• To develop and evaluate advanced ML models for HAR using smartphone sensor data.
• To compare the performance of various ML classifiers against a novel Recurrent Neural Network (RNN) model.
• To achieve higher accuracy in classifying a diverse set of daily human activities.

Main Methods:

• Collected a dataset of nine daily activities (Laying Down, Stationary, Walking, etc.) using smartphone and wearable sensors (accelerometer, gyroscope, magnetometer).
• Trained and evaluated several ML models: Decision Tree, Random Forest, K Neighbors, Multinomial Logistic Regression, Gaussian Naive Bayes, and Support Vector Machine.
• Developed and tested a custom Bidirectional Long-Short-Term Memory (Bi-LSTM) model, a type of RNN.

Main Results:

• The Random Forest algorithm achieved a test accuracy of 95%.
• The proposed custom Bidirectional Long-Short-Term Memory (Bi-LSTM) model achieved a superior test accuracy of 98.1%.
• The Bi-LSTM model demonstrated improved performance compared to traditional ML algorithms.

Conclusions:

• Advanced ML models, particularly RNNs like Bi-LSTM, significantly enhance HAR accuracy.
• Smartphone sensor integration with sophisticated ML offers a promising avenue for next-generation activity recognition systems.
• The developed Bi-LSTM approach provides a more accurate alternative to current algorithmic-based HAR in wearable devices.