Existing methods for detecting AC fault arcs in series during the charging of electric bicycles suffer from high computational complexity and cost, potentially yielding suboptimal results in arc fault detection. To address these limitations, this paper introduces a lightweight convolutional neural network model based on depthwise separable convolution and a feature fusion module. This study investigates the proposed detection method. Initially, an experimental platform simulating series arc faults was established to collect charging current data from electric bicycles. Subsequently, arc characteristics of the electric bicycles were extracted from the time domain, frequency domain, and image domain, with image domain feature extraction implemented using the Markov Transition Field (MTF) method. Then, the proposed model takes the obtained MTF images as input data, automatically extracting high-dimensional image features through convolutional and pooling layers. Subsequently, a feature fusion module reduces the dimensionality of the convolutional output features and integrates them with arc time-domain and frequency-domain features. Finally, classification and recognition of normal and arc samples are achieved via fully connected layers and a Sigmoid function. Compared to conventional convolutional networks, the proposed lightweight network significantly reduces model parameters and computational load, achieving a test accuracy of 99.6%.