To address the limited generalizability of electric-vehicle fault warning caused by large differences among vehicle models, restricted access to on-board private data and insufficient extraction of temporal features, the paper proposes a hybrid deep-learning fault-warning model based on grid-side charging-process data. First, model-independent voltage, current and other time-series data readily collected at charging piles are processed by the Transformer self-attention mechanism to capture global temporal dependencies and long-range correlations among key components such as batteries and motors. Scecond, a Bidirectional Long Short-Term Memory (BiLSTM) network models local temporal features in both directions to enhance the contextual representation of fault signatures. Finally, a Support Vector Machine (SVM) classifies the fused features to warn of fault levels during charging. Visualization of attention weights confirms that the model focuses effectively on fault-sensitive features such as voltage anomalies and abrupt temperature-gradient changes. The results show that the method enables unobtrusive "inspection while charging" fault warning using the charging process.