Abstract: Partial discharge is an early indicator of insulation failure in power equipment. Accurate detection and pattern recognition are crucial for preventing safety issues caused by insulation failure. Partial discharge generates vibration signals, which can be leveraged for sensing. Owing to the advantages of the distributed feedback fiber laser (DFB-FL), including its high sensitivity, small size, easy integration, and anti-electromagnetic interference, it was employed as a sensing element to detect vibration signals caused by partial discharge. Different types of partial discharge models were constructed. Vibration signals generated by partial discharge were collected using DFB-FL, and the characteristics of these vibration signals were analyzed. The LSC-Classifier algorithm was utilized to adaptively learn vibration signal features under different models for partial discharge pattern recognition. Based on partial discharge theory, three discharge models were established—air-gap, creepage, and sharp-plate—using a power frequency voltage signal as the excitation source. In the air-gap model, the discharge position was aligned with the positive and negative half cycles of the voltage signal, and the partial discharge vibration duration and intensity in both half cycles were similar. For the creepage model, the discharge position also corresponded to the positive and negative half cycles of the voltage signal. The partial discharge vibration generated in the positive half cycle had a shorter duration and stronger vibration intensity, while that in the negative half cycle exhibited a longer duration and weaker vibration intensity. In the sharp-plate model, the discharge position was associated with the negative half cycle of the voltage signal, resulting in a longer duration and weaker vibration intensity of the partial discharge vibration. As the excitation voltage continued to rise, discharge phenomena were also induced in the positive half cycle. Three discharge models, namely suspension, metal particle, and metal strip models, were established based on the practical application scenarios of partial discharge. In the suspension model, when the excitation voltage was elevated, a longer vibration time was observed, while the vibration intensity remained unchanged. For the metal particle model, although the vibration duration showed no alteration with the rising excitation voltage, the vibration intensity increased. In the metal strip model, both the vibration duration and intensity increased as the excitation voltage rose. The incorporation of these composite discharge models provided more comprehensive and realistic discharge data for partial discharge pattern recognition. To evaluate the predictive performance of the support vector machine (SVM) and the long short term memory (LSTM) models, comparisons were conducted based on metrics such as accuracy, precision, recall, and F1 score. It was found that the LSTM model outperformed the SVM model across all indicators; however, its accuracy in recognizing complex partial discharge patterns was insufficient for precise type determination. By integrating the advantages of SVM and LSTM models, an LSC-Classifier fusion algorithm was proposed for pattern recognition. This algorithm enhanced the overall recognition accuracy of the typical discharge model by about 7% and 2%, respectively, compared to SVM and LSTM models. Moreover, the overall recognition accuracy for the composite discharge model was improved by approximately 18% and 10% compared to the SVM and LSTM models, respectively. Experimental results validated that the application of neural network learning algorithms, when applied to the pattern recognition of partial discharge signals based on DFB-FL vibration sensing systems, exhibited superior feasibility and accuracy, thereby demonstrating strong potential for practical partial discharge detection.Current partial discharge detection is predominantly conducted under controlled laboratory conditions, which represent idealized environments. However, empirical observations reveal that temperature and humidity variations significantly influence partial discharge signal detection reliability, occasionally resulting in undetectable signals. These environmental factors should therefore be systematically taken into account in future research to enhance practical applicability.