Inter-turn short circuits in diesel generator stator windings occur frequently, and early-stage faults are difficult to detect. Failure to identify such faults promptly may lead to severe consequences. To address this issue, TSLANet(Time Series Lightweight Adaptive Network) is employed to extract and classify features from three-phase stator currents, improving the accuracy of characteristic frequency extraction during early weak fault conditions. Additionally, a Hybrid Positional Encoding (HPE) is proposed to optimize the positional encoding in TSLANet, enhancing its ability to jointly capture local transient features and global temporal patterns. This results in improved accuracy and F1 scores. Furthermore, a diesel generator model is developed in this study to simulate inter-turn short-circuit faults in stator windings, generating a representative dataset of three-phase stator currents under various short-circuit conditions. Experimental validation based on this dataset demonstrates that the TSLANet method integrated with HPE significantly enhances the early fault identification accuracy for inter-turn short circuits in diesel generator stator windings, confirming the effectiveness of the proposed approach.