Signals classification based on IA-optimal CNN.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Existing A-optimal-based Convolutional Neural Networks (CNNs) lack verification across diverse signal datasets.
• Simplified objective functions in current CNNs limit signal classification accuracy.

Purpose of the Study:

• Propose an Improved A-optimal (IA-optimal) CNN for enhanced signal classification.
• Enhance classifier stability and accuracy by employing a precise optimization objective function.

Main Methods:

• Introduced IA-optimal CNN utilizing the trace of the covariance matrix of fully connected layer weights as the objective function.
• Developed a parameter optimization model without simplifying the objective function.
• Employed a novel dual function to convert the optimization problem into a binary function optimization problem.
• Derived accurate weight update formulas using alternate iterative optimization.

Main Results:

• Tested IA-optimal CNN on five diverse signal datasets, demonstrating its universality in signal classification.
• Compared IA-optimal CNN performance against existing A-optimal-based algorithms, showing superior results.
• Validated the theoretical convergence of the trace of the covariance matrix, though strict A-optimal state is unattainable.

Conclusions:

• IA-optimal CNN offers improved stability and accuracy for signal classification tasks.
• The precise optimization approach overcomes limitations of simplified objective functions in prior methods.
• Experimental and theoretical results confirm the effectiveness and universality of the proposed IA-optimal CNN.