A multiple fault diagnosis approach for transformer based on double antenna RFID sensor and PSA-optimized LSTM

Abstract

This work aims to realize multiple mechanical fault diagnosis for transformers in the incubation period. A double antenna radio frequency identification (RFID) sensor is used to obtain vibration signals generated by a transformer. The measured raw vibration signals have prominent nonlinear characteristics; besides, they are mixed with vast noises, such as electromagnetic interference, measurement disturbance and so on, making it challenging to identify distinguishable features from the measured data. The long short-term memory (LSTM), which exhibits satisfactory performance in dealing with large-scale nonlinear time-series signals, is adopted to extract features from high-dimensional raw signals. Because the parameters of the LSTM cell and support vector machine significantly determine the accuracy of diagnosis, these parameters are determined using the chaotic quantum particle swarm optimization algorithm. Moreover, the sparse periodical attention (PSA) mechanism is used to enhance the LSTM model’s performance by focusing on global feature learning and determining the input data length. The experimental results verify that the exploited RFID sensor realizes reliable data wireless transmission within 17.5 m. Moreover, the PSA-optimized LSTM approach achieves satisfactory trade-off between diagnosis accuracy and computation complexity while dealing with multiple mechanical fault diagnosis of a transformer in early stages.