Semi-supervised learning for optical fiber sensor road intrusion signal detection

Area of Science:

• Engineering
• Computer Science
• Signal Processing

Background:

• Road intrusion detection systems are crucial for security.
• Existing unsupervised methods suffer from high false alarm rates with novel non-intrusion data.
• Distributed optical fiber vibration sensors offer a promising data source for intrusion detection.

Purpose of the Study:

• To develop an advanced road intrusion detection model.
• To address the limitations of current unsupervised classification methods, specifically high false alarm rates.
• To enhance the recognition of intrusion signals using a novel semi-supervised approach.

Main Methods:

• Proposal of a one-dimensional semi-supervised generative adversarial network (1D-SSGAN).
• The 1D-SSGAN comprises a generator and a discriminator.
• Training involves the generator and discriminator on an N-class dataset, with the discriminator's output mapped to N+1 classes. The generator synthesizes new samples from limited data to aid discriminator training.

Main Results:

• The 1D-SSGAN effectively expands datasets by generating new samples from a small initial set.
• This data augmentation assists in training the discriminator more robustly.
• Experimental analysis confirmed the proposed model's effectiveness in intrusion signal recognition.

Conclusions:

• The proposed 1D-SSGAN model demonstrates significant effectiveness for road intrusion detection.
• Semi-supervised learning combined with generative adversarial networks offers a viable solution to reduce false alarms in security systems.
• This approach enhances the reliability of intrusion detection systems utilizing distributed optical fiber vibration sensor data.