Voltage transformers (VTs) are critical sensing devices for high-voltage signal acquisition in power systems, ensuring their measurement accuracy is the foundation for downstream applications. In addition to the main body of the VT, whose measurement error is considered and emphasized, the connected secondary circuit is also a potential source of the measurement error but is usually neglected. The error introduced by poor contact in the secondary circuit is characterized by non-periodic and intermittent fluctuations. Unfortunately, the conventional secondary voltage drop test can hardly capture the entire abnormal process, while the previously proposed anomaly detection approach is only applicable to substations with multiple VT sets. In this study, we propose a generic non-periodic anomaly detection method to achieve long-term online monitoring of secondary circuits without installing additional equipment or restrictions on the power system topology. Generally, the cosine similarity is utilized to characterize the time series differences of secondary voltage measurements caused by poor contact in the secondary circuit, and then a Gaussian mixture model is built to cluster the processed features to achieve unsupervised anomaly detection. Experimental results on three real-world cases extracted from three substations indicate the effectiveness of the proposed method compared with six representative unsupervised approaches.