Improved Neural Controlled Differential Equation for Remaining Useful Life Prediction of Power Transformers

In recent years, researchers have presented numerous deep learning (DL) approaches to provide reliable prediction of remaining useful life (RUL) in prognostic and health management (PHM) applications. Although supervised DL approaches, such as gated recurrent unite, long-short term memory have overcome RUL prediction technology, these methods are still dependent on certainty data. Concerning real-life PHM applications, the machine learning-based RUL prediction of the power transformer methods is still in the initial phase. To solve this issue, this paper presented improved neural controlled differential equations for RUL prediction of power transformers. First, the multi-scale entropy and K-means are used to calculate the health confidence of the power transformer based on the vibration signal. Then, the cross-attention mechanism improves the feature extraction ability of neural controlled differential equation to overcome the influence of uncertain phenomena. Finally, the RUL of the power transformers is obtained by the health index formula. The advantages of the presented approach have been verified on the vibration data from 13 real power transformers. The presented approach compares with the different RUL prediction methods and obtains a stronger performance than the comparison algorithms. Contrastive results demonstrate that the presented approach obtains an accurate RUL of the power transformer online. Moreover, the accuracy of RUL prediction achieves 0.0523.