Scalable Hierarchical Deep Neural Network for Time Series Analysis in Wearable Sensor-based Human Activity Recognition

Zahra GhorratiAAAI 2025 · DOI

TL;DR

This paper introduces a novel hierarchical deep neural network designed for Human Activity Recognition (HAR) using wearable sensor data. By integrating fuzzy logic and adaptive learning, the framework aims to improve accuracy, interpretability, and computational efficiency, making it suitable for real-time applications on resource-constrained devices.

What Problem Does This Paper Address?

Human Activity Recognition (HAR) using wearable sensors is a rapidly evolving field, crucial for applications ranging from healthcare to sports analytics. While deep learning models like RNNs and CNNs have shown promise, they often suffer from a need for massive labeled datasets, significant computational power, and a lack of interpretability, making them "black boxes." Existing models also struggle with capturing long-term dependencies and adapting to real-time changes in data. This research aims to bridge this gap by developing a more efficient, interpretable, and adaptable deep learning framework for HAR.

Methods

The proposed framework introduces a novel adaptive hierarchical deep neural network with several key components:

  • Hierarchical Architecture: This multi-level network processes time series data at various temporal resolutions, allowing it to capture both short-term and long-term dependencies for more accurate activity recognition.
  • Adaptive Learning Mechanism: An adaptive module adjusts the model's parameters in real-time based on evolving input characteristics, enabling continuous learning without constant retraining.
  • Fuzzy Logic Integration: Fuzzy logic principles are incorporated to handle data uncertainty and imprecision, enhancing interpretability by providing more transparent decision-making, especially with noisy or ambiguous inputs.
  • Computational Efficiency: Techniques like hierarchical learning are employed to reduce the number of parameters, leading to less training data requirements and lower computational costs. This makes the model suitable for deployment on edge devices like smartphones and wearables.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Bridging the Interpretability Gap: By integrating fuzzy logic, the model moves beyond the "black box" nature of traditional deep learning, offering more transparent decision-making processes. This is crucial for building trust in AI systems for critical applications like healthcare.
  • Efficiency Through Hierarchy: The hierarchical structure not only helps in capturing complex temporal patterns but also significantly reduces the number of parameters. This directly translates to less data needed for training and lower computational overhead, a major win for edge device deployment.
  • Adaptability for Real-World Dynamics: The adaptive learning mechanism is key to handling the inherent variability and evolving nature of sensor data in real-world HAR scenarios. This allows the model to stay relevant and accurate over time without frequent, costly retraining.
  • Robustness to Imperfect Data: Fuzzy logic's ability to handle uncertainty makes the model inherently more robust to noisy sensor readings, a common challenge in wearable data collection.

Key Results

The paper outlines the anticipated contributions and benefits of the proposed framework rather than presenting specific quantitative results from experiments. These anticipated benefits include:

  • Improved accuracy and interpretability for real-time HAR.
  • Enhanced scalability for deployment on resource-constrained devices.
  • Reduced need for extensive labeled datasets due to efficient hierarchical learning.
  • Increased robustness to noisy inputs.

Limitations and Future Work

Limitations:

  • The paper focuses on the proposed framework and its anticipated benefits, with specific experimental results and quantitative evaluations not yet presented.
  • While the paper mentions "less training data is needed," the exact reduction in data requirements and its impact on performance across different dataset sizes are not quantified.

Future Work:

  • Model Optimization for Wearable Sensors: Further work is needed to optimize the framework specifically for real-time HAR using wearable sensor data, with a focus on minimizing computational overhead.
  • Implementation of Adaptive Learning Module: The adaptive learning mechanism needs to be fully implemented and tested to demonstrate its ability to adjust parameters based on temporal dynamics.
  • Extensive Testing on Real-World Datasets: Comprehensive testing on various real-world HAR datasets is crucial to evaluate the model's performance, robustness, and generalization capabilities.

Summary updated

BibTeX
@inproceedings{ghorrati2025scalable,
  author = {Zahra Ghorrati},
  title = {{Scalable Hierarchical Deep Neural Network for Time Series Analysis in Wearable Sensor-based Human Activity Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35206},
}
Figure