Monitoring and evaluating the in-service condition of equipment in real time is essential for maintaining the stable operation of large-scale complex electromechanical systems. Inductive oil debris sensors detect wear particles in lubricating oil via electromagnetic induction, providing a reliable basis for assessing the wear condition of critical mechanical components, and have been widely applied in the maintenance of large-scale mechanical equipment. However, the induced voltage signals generated by debris are usually weak and are difficult to identify accurately by manual feature extraction under the influence of various interferences, which limits the identification accuracy and generalization capability of inductive oil debris sensors. To address this issue, this paper proposes an intelligent recognition method for oil debris signals, referred to as PCatten. Firstly, the morphological stability of debris-induced signals under multiscale filtering is exploited to construct multiscale filtering features, which characterize the key geometric profiles and energy distribution of debris events and provide physically meaningful input representations for subsequent deep learning. Subsequently, a parallel convolution module is designed to perform branch-wise deep convolutional modeling of features at different scales, and an improved fusion attention module is introduced to adaptively recalibrate feature weights along the channel and temporal dimensions, thereby highlighting debris-sensitive components and suppressing complex background interference. Finally, the reconstructed multiscale feature sequence is fed into a Vision Transformer, which captures long-range dependencies and cross-scale correlations through the self-attention mechanism, enabling accurate discrimination of debris-induced voltage signals under strong interference and low signal-to-noise ratios. Experimental results demonstrate that the proposed model achieves excellent performance on both three-coil sensor dataset and high-gradient static magnetic field sensor dataset, the interference elimination rate, debris identification rate, and debris identification accuracy are 99.72%, 98.94%, and 99.44%, respectively, and the proposed method still exhibits superior debris-detection performance compared with traditional algorithms under low signal-to-noise ratios ranging from -5 to 0 dB.