Artificial intelligence is the general trend in the field of power equipment fault diagnosis. However, limited by operation characteristics and data defects, the application of the intelligent diagnosis method in power transformers is still in the initial stage. To fill this research gap, in this article, a novel double-stacked autoencoder (DSAE) is proposed for a fast and accurate judgment of power transformer health conditions with an imbalanced data structure. Three problems affecting the diagnosis effectiveness are overcome by a DSAE framework, an aging-tolerance criterion, and an advanced sparse deep clustering network. The proposed DSAE method is validated by two case studies based on an actual power transformer dataset. The results indicate that the proposed DSAE method can achieve a fairly reliable diagnosis with a higher accuracy and less time than the other methods, which demonstrates the superiority and effectiveness of the proposed approach.

Early stage fault detection plays a pivotal role in Industrial equipment accidents avoidance and scientific maintenance. While limited by the complex operation background, its application encounters with the conundrum of fault feature indistinctness. To address the challenge, a time-series transfer learning (TSTL) method is proposed, which contains two phases: first, early stage series are transferred to their corresponding serious stage for fault feature enhancement. Moreover, due to the improvement of model structure and loss function, the limitation of mismatched working condition is well-weaken. Second, a transferred fault mode recognition model is trained by using transferred normal series that provides a novel solution for data imbalance. Finally, the TSTL method is verified by actual vibration datasets of power pole tower bolts. Its superiority in feature transfer and fault detection is confirmed by several groups of comparative experiments and results demonstrate TSTL outperforms mainstream methods.

When power transformer failure occurs, it can often be considered in two parts, which are electrical accident and mechanical breakdown. Researches usually classify the faults by dissolved gas analysis (DGA) and locate the trouble position by analyze mechanical vibration. However, under the background of power big data, the raw data with multi-source and heterogeneous characteristics bring difficulties to diagnosis. Deep learning (DL) provides an effective way to extract the features of raw data automatically. Convolutional neural network (CNN) is an effective DL method. In this paper, characteristic gas and vibration sensor locations are selected by analyzing fault evolution mechanism. Then, to improve diagnosis accuracy and efficiency as well as reduce noise interference, fault classification and location on line are realized by using LeNet-5 and proposed Fault Detection and Location CNN (FDL-CNN) respectively. At last, the method effectiveness is validated in the experiment.

A method for time interpolation based on the climatological speed of large-scale atmospheric waves that is empirically determined is proposed. When tested on a 7-yr dataset this method is found to be easy to use, has good accuracy, and is, in fact, considerably more accurate than the much used linear time interpolation. The gain in accuracy is particularly large for mobile synoptic waves. Several applications of a time-continuous description of the atmosphere are discussed. Abstract A method for time interpolation based on the climatological speed of large-scale atmospheric waves that is empirically determined is proposed. When tested on a 7-yr dataset this method is found to be easy to use, has good accuracy, and is, in fact, considerably more accurate than the much used linear time interpolation. The gain in accuracy is particularly large for mobile synoptic waves. Several applications of a time-continuous description of the atmosphere are discussed.