To improve the detection efficiency and accuracy of series arc fault (SAF) in low-voltage alternating current power system, this study takes industrial motor distribution circuits as the object, constructs datasets through arc fault experiments, and designs a lightweight SAF identification model based on MobileViT architecture. The model uses lightweight convolution modules and transformer modules to extract local and global features from the current signal respectively, and uses the unfold-transformer-fold mechanism and global average pooling to achieve parameter and complexity reduction. Further, the TensorRT inference optimizer and engine are used to deploy and optimize the model, which significantly improves the inference speed of the model in embedded devices, and based on this, the full-circuit SAF on-line detection device is developed. The detection device has flexible deployment characteristics: when installed at the front end of the frequency converter, it can simultaneously monitor the SAF of the front and back end of the frequency converter. It can also achieve precise monitoring of the SAF of the back end when installed at the back end. The test results show that the average runtime of the device is less than 0.874 ms, and the accuracy is above 97.20%, which can meet the requirements of IEC62606 standard and industrial scenarios. In addition, the comparison experiments show that the device is superior to the existing arc fault detector products and can provide a reference for the development of industrial arc fault circuit breakers.