Acoustic emission-based partial discharge localization in oil using neural network approaches

The primary challenge in partial discharge (PD) acoustic emission (AE) localization systems lies in the complexity of the measurement methods. Existing systems rely on a substantial number of AE sensors, which, in turn, impact the size of the data acquisition unit. PD localization techniques within AE systems that are rely only on time difference approaches like Time of Arrival (TOA) principles may not offer the most effective solution for accurately pinpointing PD sources, especially when dealing with uncommon and weak PD AE signals. This study presents an examination on the PD AE localization in oil by combining the Time Difference of Arrival (TDOA) approaches with Adaptive Neuro-Fuzzy Inference System (ANFIS), Artificial Neural Network (ANN) and Optimized-Artificial Neural Network (Op-ANN) to increase the accuracy of PD localization. Needle-plane electrode configuration with 30 mm and 50 mm gap was used to initiate PD signal at three difference locations known as PD Location 1, 2 and 3. The needle tips is 3 µm whereby the plane diameter is 50 mm. Impedance matching circuit (IMC) was used to measure the PD electrical signal. The PD AE signal was acquired through three unit of AE sensors and preamplifier gain units. Once the voltage reached 30 kV, the PD electrical and AEs were recorded. Next, the PD electrical and AE signal were denoised by moving average (MA), finite impulse response (FIR) type of low pass filter (LPF) and high pass filter (HPF), and discrete wavelet transform (DWT). The distance between PD source and AE sensor was calculated based on TDOA to determine the PD location. Comparison between two type of time difference techniques known as first peak method (FPM) and cumulative energy method (CEM) were performed to identified the time difference for measured PD AE signal. Results obtained by FPM and CEM were then applied for analysis on PD AE localization utilizing TDOA. PD AE localization results based on TDOA were used as an input to train the ANFIS, ANN and Op-ANN in order to locate PD source. ANN utilized the single hidden layer to locate PD whereby for Op-ANN, the learning rate and multi hidden layer were optimized for PD AE localization. High accuracy in locating PD based on AE is demonstrated by ANFIS and ANN 4 multi hidden layer. ANFIS achieved a mean difference of 0.02 m at PD Locations 1 and 2, and 0.03 m at PD Location 3. The ANN single hidden layer displays a well-performing range of RMSE values between 0.06 and 0.12, and R² values between 0.59 and 0.92. In terms of mean computation time, ANN learning rate achieving the fastest with 0.68 s. For ANN learning rates, the range of R² and RMSE is from 0.53 to 0.96 and from 0.02 to 0.17, respectively. Although reasonable prediction error and fast computation time are exhibited by ANN learning rates, a relatively high mean difference for PD AE localization with the value of 0.031 m at PD Locations 1, 2, and 3 is observed. In contrast, low mean differences are maintained by the ANN 4 multi hidden layer, specifically 0.026 m for all PD locations. An R² of 0.76 is achieved at the y-axis for PD Location 1, and the lowest RMSE is observed at the z-axis for PD Location 3. The mean computation times is 2.26 s. Overall, the ANN 4 multi hidden layer have successfully locates the PD AE signal with 95% level of accuracy whereby ANN single layer and learning rate approaches is 92% and 93%. ANN 4 multi hidden layer managed ability to locate PD source through multiple layers of abstraction and learning ability on denoised PD AE signal data. The complex interactions between neurons and number of hidden layers make ANN 4 hidden layer unique to perform PD AE localization predictions. This study shows that the used of three AE sensors with AI for PD AE localization is beneficial in reducing the complexity of measurement approaches. The denoising algorithms developed and examined shows a flexibility to denoise PD AE signal. The combination between TDOA and neural network approaches provides an optimum solution to perform PD AE localization that was utilizing three AE sensors. The neural network approach has effectively reduced location and relative error rates, demonstrating the ability of the ANN 4 multi hidden layer to learn complex patterns and relationships in the PD AE signal data. This will help the related industry player to increase the effectiveness of conducting PD AE localization system using three AE sensor, flexi denoising method, and application of neural network in PD AE localization.

Text FK 2024 10 - Declaration Form.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (1MB) Text FK 2024 10 - Full Text.pdf Restricted to Repository staff only until 31 September 2029. Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (13MB) Text FK 2024 10.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (13MB) Official URL or Download Paper: http://ethesis.upm.edu.my/id/eprint/18974

Actions (login required)

View Item