Wearable Sensor-Based Human Activity Recognition: Performance and Interpretability of Dynamic Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning
• Wearable Technology

Background:

• Human Activity Recognition (HAR) is crucial for healthcare, rehabilitation, and smart monitoring.
• Wearable sensor data offers a rich source for HAR, but selecting appropriate neural network architectures is key.
• Dynamic neural networks are promising for HAR due to their ability to model temporal dependencies.

Purpose of the Study:

• To systematically compare the performance, computational cost, and interpretability of three dynamic neural network architectures for HAR.
• To evaluate Finite Impulse Response Neural Network (FIRNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) for HAR tasks.
• To apply Layer-wise Relevance Propagation (LRP) for enhanced model transparency in wearable sensor-based HAR.

Main Methods:

• Trained 16,500 models across varying delay lengths and hidden neuron counts for FIRNN, LSTM, and GRU.
• Evaluated models based on classification accuracy, computational complexity, and interpretability.
• Utilized Layer-wise Relevance Propagation (LRP) to analyze feature importance and model transparency.

Main Results:

• LSTM achieved the highest classification accuracy (98.76%), followed by GRU (97.33%) and FIRNN (95.74%).
• FIRNN demonstrated the lowest computational complexity.
• LRP analysis identified gyroscope Y-axis as the most informative sensor data, while accelerometer Y-axis was least informative; GRU showed broader relevance distribution than FIRNN.

Conclusions:

• LSTM and GRU offer superior accuracy for wearable sensor-based HAR compared to FIRNN.
• FIRNN provides computational efficiency, making it suitable for resource-constrained environments.
• The study provides practical guidance on balancing performance, complexity, and interpretability in explainable HAR models.