A Novel CNN-based Bi-LSTM parallel model with attention mechanism for human activity recognition with noisy data

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Human Activity Recognition (HAR) using smartphones is crucial for mobile applications.
• Challenges in HAR include classification accuracy and processing time for long, noisy sequences.
• Existing Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) models face limitations with noisy or incomplete sensor data.

Purpose of the Study:

• To propose a novel model, ConvBLSTM-PMwA, for efficient and accurate HAR.
• To address the challenges of classification time and accuracy in processing noisy sensor data.
• To improve feature extraction and integration for robust HAR.

Main Methods:

• A 1-D Convolutional Neural Network (CNN)-based bi-directional Long Short-Term Memory (LSTM) parallel model with an attention mechanism (ConvBLSTM-PMwA) was developed.
• Sensor data was segmented into sub-segments using a sliding window for accelerated feature extraction.
• An attention mechanism was employed to redistribute feature weights and integrate them for final classification.

Main Results:

• The ConvBLSTM-PMwA model achieved a classification accuracy of 96.71% on public UCI and WISDM HAR datasets.
• The proposed model demonstrated improved computational time complexity, being at least 1.1 times faster than existing CNN and RNN models.
• The model effectively handled noisy HAR data, outperforming baseline methods.

Conclusions:

• The ConvBLSTM-PMwA model offers a significant advancement in smartphone-based HAR.
• The model provides a robust solution for accurate and efficient activity recognition, even with challenging data conditions.
• This approach enhances the feasibility of real-time HAR applications on mobile devices.