To address the challenges of fault feature extraction and diagnostic accuracy in complex operating conditions for plunger pumps, this study proposes an intelligent diagnostic method combining Transformer-based feature extraction, improved sparrow search algorithm optimization, and support vector machine classification. The approach focuses on vibration signals by employing Transformer architecture with multi-head self-attention mechanisms to capture temporal dependencies and global correlations, generating high-dimensional, discriminative deep feature representations. During feature classification, an enhanced sparrow search algorithm optimizes support vector machine kernel parameters and penalty coefficients through adaptive optimization, achieving global search in parameter space to improve model generalization. Experimental validation on the UCI hydraulic system dataset demonstrates superior performance in classification accuracy, stability, and noise resistance compared to long/short-term memory networks and unoptimized SVM models. These results validate the effectiveness and advantages of Transformer-based feature modeling combined with ISSA optimization strategies in intelligent plunger pump fault diagnosis.