What is a Partial Discharge (PD)?
Partial discharge (PD) is a type of electrical discharges that does not bridge electrodes and it affects only a portion of insulating material while the rest of it is not affected.
PDs can take place in solids, liquids, and gases. The important thing about PDs it their detrimental impact on the health of insulation system. PDs cause the accelerated aging of electrical equipment. There are three types of discharges:
(1) Internal discharge that happens in the cavities and inclusions within the dielectric material.
(2) Surface discharge that happens at the surface of dielectric material, usually in the presence of moisture and contamination.
(3) Corona that happens in the regions where electric field is extremely inhomogeneous and high, usually around the metal tips and edges.
What is it important to investigate PD?
PD phenomenon has been studied since early, mid-19th century as it has been known as the deterioration factor in cables, transformers, motor windings, and almost all the electrical equipment.
Recently, another industry, Aviation Industry, has urged the study of PD phenomenon.
In recent decades, extensive research has been conducted on the electrification of commercial aircraft to reduce the dependency on mechanical, hydraulic, and pneumatic systems and replace them with electrical systems. A primary goal of this path is to make the power density of the more/all-electric aircraft (MEA/AEA) closer to that of conventional aircraft.
The problem is that the technologies and improvements for achieving higher power density escalate the electrical tension on the insulation system by enhancing the risk of partial discharge. One of these technologies is the wide bandgap semiconductors which generates fast-rise, high-frequency voltage pulse. Another one is the higher voltage level required for electric aircraft which increases the probability of PD inception.
Besides, the harsh environmental conditions at high altitudes have proved as negative factors for insulation systems.
What we aim to do here is to make use of the measurements of the propagated signals in electrodes. It has been shown that this signals can be indicator of PD activities. We want to see if we can classify the PD types based on the short-term recording of this signal.
Therefore, the goal of this study is to classifying the PD type based on short-term behavior of insulation in a matter of second. This allows taking emergency actions like quick switching to backup equipment.
How does our input data look like?
In the Figure 1, you can see a generic format of raw data which has rows of time instants and the columns of cycles of data which are usually a 50 or 60-Hz depending on the frequency of power grid. The time instants are scaled based on a single period of voltage. Meaning that for 50Hz signal, all the time instants are between zero and 20 milliseconds.
If we plot the amplitude of signal versus the time instants, we have a figure called phase-resolved partial discharge pattern (see Figure 2). As you can see, this pattern is quite different for each type of discharge.
Figure 2 – Discharge types: (a) corona, (b) surface, (c) internal discharge
Now, let’s see how we can prepare this data for classification methods.
When we are doing short-term analysis in which we have only seconds of data, it is very important to generate meaning datasets that can properly train the classification methods.
For the introduced batch of data, we follow four steps to generate the training, validation, and test data.
-Step 1: We randomly choose 90%, and 10% of cycles for training and testing, respectively.
-Step 2: Sort the cycles of data based on phase (time instants) order
-Step 3: Assuming that the arrays of should have features, we categorize the time instants into groups (see Figure 3).
-Step 4: Finally, each array of training data is formed by choosing a random time instant from each group, and also choosing a random cycle (similar procedure for test data).
The measurements are performed for 4 seconds for each discharge type. There are 200 cycles of data for each discharge type. A cycle of data is a discrete measurement over a period of the 50-Hz voltage signal. Therefore, the time instants in a cycle are between 0 and 20 milliseconds. Moreover, time instants are phase-resolved meaning that they are scaled for the single cycle of 0-20 milliseconds. At each time instant, the signal is measured in volts.
The remaining properties of raw datasets and the final training and test datasets are as follows:
- The number of time instants in the dataset of each discharge type is different and is follows: Corona (2405), Surface (795), and Internal (3521).
- The number of training arrays for each type of discharge is .
- The candidate numbers of features are .
- The number of testing arrays for each type of discharge is .
- For , the training and testing datasets will have shapes of (45000×256) and (3000×256), respectively.
- The labels for corona, surface, and internal discharges are 0, 1, and 2, respectively.
Which classification methods do we use? Results?
Based on the definition of the problem, this is a supervised learning problem, and the candidate methods are as follows:
(1) K-Nearest Neighbor (KNN)
(2) Logistic Regression
(3)Gaussian Naïve Bayes
(4) Random Forest
(5) Bagging Trees
(6) Adaptive Boosting
The accuracy and computation time of these classification methods are shown in Table 1. For all candidates of , the boosting method presents the highest accuracy. However, most computation time also belongs to this method. If we compare the data duration (which is 4 seconds), for the optimal case of , it is about 100 times the data duration. For the time-sensitive application of insulation health for electric aircraft, it does not look promising. Similarly, decision tree bagging offers high accuracy, but also high computation time. Among all the methods, logistic regression has the poorest performance by far, probably due to the fact that it cannot handle Non-linear problems because of its linear decision surface.
Also, KNN does not perform well. Although at first glance, it might seem like a good choice for the assessment of short-term behavior, KNN does not perform well in the presence of noise. Furthermore, KNN can suffer from skewed class distributions. When a certain class is very frequent in the training set, it will tend to dominate the majority voting.
Random Forest and Gaussian Naïve Bayesian classifiers are among the quickest methods with acceptable accuracy. But our choice would definitely be Gaussian Naïve Bayes Classifier; it offers almost 80% accuracy in less than one-tenth of a second which is a bright choice for the case of electric aircraft. In the case of a danger threatening the insulation system, and thus, the whole electrical equipment, Switches can act in less than a second and protect the entire system against a potential fault. Despite its simplicity, Gaussian Naïve Bayes is known to have a promising performance for small datasets which is confirmed in this study.
<i class=”fa fa-arrow-right” aria-hidden=”true”></i>
- Aviation industry aims to electrify aircraft to achieve emission reduction of 50% set by NASA.
- Enhancement of the power density (weight) → enhances the electric tension on insulation systems.
- Partial discharge is an indicator of insulation system’s health.
- This study aims to use short-term response of insulation systems → To predict the type of discharge and take proper mitigating actions in a matter of seconds.
- Bagging and Boosting methods offer highest accuracies → Also, highest computation time ⊗
- Gaussian Naïve Bayes classifier offer nearly as accurate as ensemble methods’ accuracy → In less than a second which makes it a promising solution for aeronautic applications.
Through following links, you have access to the project paper, presentation slides, codes, and data.