An Automated CNN–BiLSTM Pipeline for Online Detection of Partial Discharge Pulses From HFCT Measurements With PRPD-Based Experimental Validation

Reliable detection of partial discharges (PD) is essential for condition monitoring of high-voltage electrical equipment. In applications based on high-frequency current transformer (HFCT) sensors, the construction of phase-resolved partial discharge (PRPD) diagrams depends on the accurate discrimination of pulses in signals often contaminated by noise and interference. Traditional signal processing methods present limitations regarding parameterization, scalability, and automation, particularly in continuous monitoring scenarios. This work proposes a new automated pipeline for detecting time windows containing PD pulses from time-domain raw signals acquired with an off-the-shelf 8-bit data acquisition system coupled with an HFCT sensor, integrating deep-learning-based detection with the construction of PRPD diagrams. An aspect that characterizes the originality of this work lies in the fact that a hybrid convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) model are employed to classify temporal segments without the use of fixed statistical thresholds or heuristic rules. The identified windows are used to construct PRPD diagrams, explicitly preserving the phase and amplitude information associated with each pulse. In the test set, composed of acquisitions in a high-voltage (HV) laboratory involving equipment taken out of operation, the model achieved an accuracy of 97.49%, with a recall greater than 97% for the positive class and an area under the curve (AUC) of 0.993. The amplitude estimate derived from the maximum absolute value of the pulse is consistent with baseline reference results and conventional measurements in accordance with the IEC 60270 standard, providing quantitative and qualitative validation of the constructed PRPD patterns. The results indicate that the proposed original method increases the level of automation, preserves physical interpretability, and shows strong potential for application in intelligent continuous monitoring systems for HV equipment.