Deep Transfer Learning for Time Series Data Based on Sensor Modality Classification.

Area of Science:

• Ubiquitous Computing
• Machine Learning
• Pattern Recognition

Background:

• Limited labeled time-series data hinders deep learning model training, particularly in ubiquitous computing.
• Wearable device data analysis requires robust pattern recognition for meaningful applications.

Purpose of the Study:

• To propose a transfer learning method for training deep neural networks (DNNs) with scarce labeled time-series data.
• To develop a general architecture adaptable to various sensors for multichannel data in target fields.
• To evaluate the method's effectiveness in Human Activity Recognition (HAR) and Emotion Recognition (ER).

Main Methods:

• A transfer learning approach is proposed, leveraging sensor modality labels from large unlabeled time-series datasets.
• A DNN is pre-trained on general time-series characteristics and then transferred to a specific target problem DNN.
• The method utilizes a general architecture adaptable to different sensors, suitable for multichannel data.

Main Results:

• The transfer learning approach significantly outperforms baseline DNN training without transfer learning for both HAR and ER.
• A new dataset, Cognitive Village-MSBand (CogAge), comprising 61 atomic activities from three wearable devices, was introduced for HAR.
• The proposed method demonstrates effectiveness across different ubiquitous computing tasks and wearable sensor modalities.

Conclusions:

• Transfer learning offers a viable solution to the challenge of limited labeled time-series data in ubiquitous computing.
• The proposed method enhances the performance of deep learning models for pattern recognition tasks like HAR and ER.
• The adaptable architecture makes the approach suitable for diverse wearable sensor data and multichannel applications.