Data-Driven Fault Detection

Insulation failure is a major cause of transformer faults. The most affected part of a transformer is the transformer winding, accounting for 30% of all internal transformer faults. Due to this, inter-turn faults are among the common types of faults that occur in transformers due to the close proximity of turns within the winding. They occur when there is an insulation failure between adjacent turns in a transformer winding.

Well why is this so bad? Because they cause unusually large fault currents to be drawn into the the transformer which can cause a breakdown, or even in worst case scenario, a fire!

Often, incipient transformer winding faults have an insignificant impact on the terminal voltages and currents of transformers. Due to this, they often go undetected by the traditional protection schemes. However, over time these faults worsen, and eventually deal extensive damage to a significant part of the winding. Hence, there is a need for a method that can effectively detect the fault at its incipient stage, as well as its severity and location so that the fault can be addressed quickly to reduce outage time and damage.

We developed a fault detection scheme that could accurately determine fault location and severity, while addressing shortcomings such as inability to determine fault severity, long operation/inference time, large amounts of requisite training data, failure to operate online, and low accuracy.

Our method uses continuous wavelet transforms and convolutional neural networks, which are respective signal processing and machine learning tools. A wavelet transform, put simply, is a mathematical tool used to understand signals and transform them into different frequencies. This is useful because signals often come in analog form. However, the computerized machine learning method speaks a different language (discrete bits). Therefore, in order to get the machine learning models to understand the input signals, wavelet transforms are used to generate coefficients, which are used to plot special images that the computer can see and understand.

Example of scaleogram image plotted with transformer signals, showing a fault on Phase C of the primary side.

Example of scaleogram image showing fault on Phase B of the secondary side.

Six input currents from the transformer are processed using a wavelet transform, with each current corresponding to a specific phase and side of the transformer—hence the total of six. From the first two images, it becomes apparent how the machine learning models are able to identify faults. Even from a human visual standpoint, the scaleogram plots reveal a distinct blue line running through the center of the images. Any disruption or irregularity in this line serves as a clear visual indicator of a fault. Simple enough, isn’t it? Look below for how the scaleogram images would look if the transformer was healthy.

Well, maybe I lied. This transformer is actually faulty too, but the fault is so minor that the human eye can’t easily detect it. However, spoiler alert: the computer can! It all makes sense, doesn’t it?

The figure above shows the flowchart of the proposed fault detection technique, from start to finish.

A 630 kVA, 10.5 kV/0.4kV multi-winding transformer was simulated in Matlab/Simulink to generate data for the CNN. A convolutional neural network is a special type of machine learning (specifically a deep learning) model used specifically for image processing due to a special set of inherent layers that use special mathematical techniques to analyze bit representation of image matrices or tensors. The CNN was set up to produce four distinct outputs; detection of the fault, the corresponding faulted phase, the transformer side where the fault occurred, and the percentage of turns affected.