Pulse eddy current (PEC) is one of the key eddy current testing (ECT) techniques and it is widely used in metal industry. Being able to automatically recognize thickness with PEC can make the manufacturing process more efficient and convenient. In this paper, the main contribution is three-fold. Firstly, a novel portable pulse eddy current device is designed, and a new PEC dataset is constructed with a wide variety of features using our device. Secondly, 1-D convolutional-based deep learning models are utilised to achieve automatic thickness recognition with high accuracy. In addition, models are moderately immune to lift-off and edge effects. Lastly, a compact and lightweight 1-D convolutional neural network is deployed on the STM32 microcontroller in our device, and it achieves real-time, accurate and low-latency automatic thickness recognition.