Existing methods of extracting V-I trajectory in RGB space used as load signatures for non-intrusive load monitoring (NILM) are complex and not suitable for edge computed devices. To solve this problem, we propose a new method to obtain the RGB V-I trajectory. The new method is simple and efficient by using the time derivatives of instantaneous reactive current and voltage to map the grayscale V-I trajectory to RGB space. Then, a lightweight CNN model, ShuffleNetV2, is selected to carry out a transfer learning study for NILM problem. Using RGB V-I trajectory obtained by using the method in this paper as load signature achieves a 5.1% improvement in F1-score than using grayscale V-I trajectory. A V-I trajectory of larger resolution avails the improvement of transfer learning performance, and the model trained on RGB V-I trajectory seems to be less affected by input resolutions for a smaller variance in F1-score among various input resolutions.

Printed circuit board (PCB) manufacturing is one of the most important parts of electronic production, where a small defect may cause the final product to fail. Therefore, the industry urgently needs a system to detect and locate all manufacturing defects. In this paper, we propose Generative Adversarial Networks (GANs) based learning defect system with an extremely low bit per pixel (BPP) for feature compression. The system includes an encoder, generator, and multi-scale discriminator for generative learned compression and a comparator to distinguish defected components from a compressed feature map generated by GAN. The model synthesizes images at extreme low bitrates where traditional methods such as JPEG show strong artifacts, resulting in a proportional reduction in the storage of feature comparison. Experimental results demonstrate the effectiveness and efficiency of our model with 97.8% mAP at 72FPS.

No abstract