Defect-Repairable Latent Feature Extraction of Driving Behavior via a Deep Sparse Autoencoder

Area of Science:

• Data Science
• Machine Learning
• Automotive Engineering

Background:

• Vehicle sensor data is collected as multi-dimensional time-series, often containing redundant information.
• Redundancy complicates data analysis and can impact results.
• Sensor data can be defective due to sensor failures, necessitating robust analysis methods.

Purpose of the Study:

• To propose a defect-repairable feature extraction method for multi-dimensional sensor time-series data.
• To extract low-dimensional time-series data representing driving behavior.
• To reduce the negative effects of defective sensor data during feature extraction.

Main Methods:

• Utilized a deep sparse autoencoder (DSAE) for feature extraction.
• DSAE is designed to be defect-repairable, handling incomplete sensor data.
• Applied DSAE to multi-dimensional sensor time-series data of driving behavior.

Main Results:

• DSAE demonstrated high-performance latent feature extraction for driving behavior data.
• The method proved effective even with defective sensor time-series data.
• Extracted latent features reduced the negative impact of data defects on driving behavior segmentation.

Conclusions:

• The proposed DSAE method is effective for extracting low-dimensional features from complex driving behavior data.
• DSAE successfully handles defective sensor data, maintaining feature extraction quality.
• The extracted features improve downstream tasks like driving behavior segmentation.