Open set recognition algorithm based on Conditional Gaussian Encoder.

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Existing Closed Set Recognition (CSR) methods misclassify unknown jamming signals as known classes.
• Open Set Recognition (OSR) is crucial for accurately identifying novel or uncatalogued signals in dynamic environments.

Purpose of the Study:

• To design a Conditional Gaussian Encoder (CG-Encoder) for 1-dimensional signal Open Set Recognition (OSR).
• To enhance the discrimination between known classes and improve the recognition of unknown jamming signals.

Main Methods:

• Developed a CG-Encoder utilizing a residual network structure and deep neural networks for signal feature extraction.
• Defined a novel Kullback-Leibler (KL) divergence to approximate known classes as Gaussian distributions in latent space during training.
• Implemented a specific OSR algorithm flow for testing unknown signals.

Main Results:

• CG-Encoder demonstrated superior CSR and OSR performance compared to three other network structures across 9 jamming types.
• Achieved over 70% open set average accuracy at maximum openness, significantly outperforming other algorithms.
• Reached over 95% average OSR accuracy at minimum openness.

Conclusions:

• The proposed CG-Encoder effectively addresses the limitations of CSR by improving OSR performance for jamming signals.
• The method shows significant potential for real-world applications requiring robust signal identification in open environments.