Machine learning-assisted multi-task learning-based performance monitoring scheme for photonic MM-wave generation systems

Photonic generation of millimeter-wave (mm-wave) signals is a paramount for high-capacity and high-speed communication. Real-time performance monitoring and optimization of system parameters are crucial to ensure system reliability and efficiency. This paper presents a novel approach for machine learning-assisted performance monitoring of Mach-Zehnder (MZ) modulator-based mm-wave generation systems. The multi-task learning convolution neural network (MTL-CNN) model leveraging transfer learning in conjunction with optical spectrum images is proposed to evaluate the key system parameters. Six tasks, comprising two regression tasks of optical sideband suppression ratio (OSSR) and modulation index estimation and four classification tasks of tupling factor, radio frequency (RF), extinction ratio, and biasing of MZ modulators, are used for effective performance monitoring. By employing MobileNet within the MTL-CNN framework, the model achieves nearly 100% accuracy in classification and minimal regression loss of 0.549 and 0.143 across two tasks of OSSR and modulation index respectively. In comparison with VGG-16 and EfficientNet-B0, the MobileNet-based MTL-CNN demonstrates superior accuracy and computational efficiency. The findings underscore the effectiveness of the proposed ML framework in providing precise performance monitoring for mm-wave generation schemes.