In partial discharge (PD) monitoring systems, multiple PD sources and external noise can be simultaneously detected, resulting in different numbers and densities of pulses depending on the signal source. This paper proposes an automatic PD diagnosis method based on dual density-based clustering and phase-resolved partial discharge (PRPD) pattern analysis to improve the diagnostic reliability of online PD monitoring. First, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) are complementarily applied to cluster pulse signals with non-uniform densities, and PRPD patterns are reconstructed from the resulting clusters. Next, to account for variations in the phase reference, trigger level associated with noise magnitude, and input signal magnitude under field measurement conditions, the PRPD data are augmented using X-axis phase shifting, Y-axis zero padding, and vertical scaling. The augmented PRPD patterns are then classified into Noise, Corona, Internal, and Surface classes using a Conv 2D-based convolutional neural network (CNN), with the classification results represented as probability scores in percentage form. Finally, Pearson correlation-based PRPD similarity analysis is applied to the CNN classification results. A True Alarm is generated when highly similar PD patterns are repeatedly detected, while PRPD patterns identified as False Alarms are stored in a Reference Memory to suppress repeated alarms caused by similar patterns. Experimental results demonstrate that the proposed method improves clustering performance under non-uniform pulse-density conditions and achieves a PD classification accuracy of 92.9%. Furthermore, the application of the PRPD similarity algorithm improves the overall diagnostic accuracy to 94.5%. By integrating pulse clustering, PRPD classification, and similarity-based alarm decision into a unified diagnostic procedure, the proposed method improves the reliability of automatic diagnosis and alarm generation for online PD monitoring systems.