Objective: Neural network models have shown strong performance in fault arc detection. However, these models often relied on fragmented, single-modality features—such as time-domain, frequency-domain, or time-frequency representations of one-dimensional time series—making it difficult to capture transient high-frequency oscillations at the microsecond level. This resulted in the loss of critical detail, limiting the ability to predict arc occurrence precursors and weakening emergency response. To address this issue, this paper proposed a fault arc detection method based on multi-modality feature fusion. Methods: An experimental circuit simulating parallel breakdown arc induced by overcurrent was built, with 75 effective working conditions designed and over 100, 000 data points collected per scenario. Based on the typical characteristics of pre-fault and fault waveforms, 5, 456 one-dimensional time-series signal samples were constructed. Five conversion methods—Markov transition field (MTF), recurrence plot (RP), Gramian angular field (GAF), short-time Fourier transform (STFT), and continuous wavelet transform (CWT)—were used to convert the transient current signals into time-frequency-space feature maps (TFS-Maps). Each mapping method involved multiple parameters, and their effectiveness in feature extraction varied, necessitating the selection of optimal settings. For MTF, parameters such as the number of bins, interval division strategy, and color mapping scheme were chosen. For GAF, the visualization results of the summation and difference angular fields were compared. For STFT, window lengths of 16, 32, and 64 were tested. For CWT, the wavelet basis, scale, center frequency, and bandwidth were optimized. For RP, the signal's standard deviation was used. The resulting multi-modality dataset—containing original signals and their corresponding TFS-Maps—was split into training, validation, and test sets in a 7∶2∶1 ratio. A gated recurrent unit (GRU) was used to model sequence dependencies in the original signals. A Swin transformer integrated with the convolutional block attention module (Swin Transformer-CBAM) was applied to highlight key regions within the TFS-Maps. The outputs from GRU and Swin Transformer-CBAM were fused via cross-modality concatenation to perform arc detection. Accuracy, precision, recall, F1-score, and comprehensive evaluation visualization graphs were used to assess the algorithm's performance. Results: The experimental results showed that (1) among various TFS-Maps, GADF achieved the highest performance, with 98.07% accuracy, 97.52% F1-score, and 98.01% recall; 2) Swin Transformer-CBAM outperformed the convolutional neural network, with improvements of 0.37% in accuracy, 0.17% in F1-score, and an increase in recall from 97.67% to 98.01%; and (3) the confusion matrix indicated very few misclassifications, with over 98% agreement between predicted and actual labels. Conclusions: Time-frequency imaging enhanced sensitivity to high-frequency transient features. The attention mechanism effectively captured fault arc features by focusing on critical frequency bands and time-domain segments. The proposed detection method met expectations, improved detection efficiency, and provided a more reliable technical solution for identifying parallel breakdown arcs induced by overcurrent.