How hamidreza-karami/application-of-ml-in-partial-discharge-localization works
Purpose and Entry Points
This is a research artifact for evaluating machine-learning approaches to 3D partial-discharge (PD) localization in power-transformer tanks. It predicts PD coordinates from electric-field sensor signals, comparing classical regression, SVR/XGBoost, and CNN approaches. Primary executable entry points are exploratory Jupyter notebooks: codes/machine_learning.ipynb, codes/CNN.ipynb, and codes/Thriple_sensor.ipynb.
Data Preparation and Storage
Raw/derived experimental data is expected under datasets/, whose data contract is described in datasets/README.md. Database/dataset construction is performed interactively through codes/create_DB.ipynb. MATLAB import and visualization scripts provide a separate ingestion/inspection path: codes/Matlab_files/importfile.m and codes/Matlab_files/show_data_in_MATLAB.m.
Single-Sensor Localization Pipeline
The principal flow is electric-field measurement → feature/model input → predicted 3D PD location → comparison against actual location. Single-sensor experiments cover one-, two-, and three-dimensional/case variants, with model evaluation persisted as prediction-vs-actual CSV files, e.g. logs/1D_final_case_study_one/prediction_per_actual.csv, logs/1D_final_case_study_two/prediction_per_actual.csv, and logs/1D_final_case_study_three_reshape_x/prediction_per_actual.csv.
Deep-Learning/CNN Modeling
codes/CNN.ipynb is the advanced model experimentation entry point; it implements the CNN-oriented branch identified as the best-performing method in the paper. CNN case-study outputs are organized by coordinate axis and sensor/geometry condition, notably logs/1D_final_case_study_Three_P1_x/, logs/1D_final_case_study_Three_P2_y/, and logs/1D_final_case_study_Three_P1_z/. Some runs retain training history and hyperparameters, such as logs/1D_final_case_study_three_upscale_x/history.csv and logs/1D_final_case_study_three_upscale_x/hyperparameters.txt.
Three-Sensor and Alternative Signal Architectures
codes/Thriple_sensor.ipynb handles the multi-sensor branch, an important boundary from the usual single-sensor input assumption. Triple-sensor outputs are explicitly segregated in logs/1D_final_case_study_Three_P1_y_triple/ and logs/1D_final_case_study_Three_P2_y_triple/. Additional snn experiment directories, such as logs/1D_final_case_study_snn_three_p1_x/, represent another neural-model family/configuration with the same prediction artifact contract.
Classical-ML Baselines
codes/machine_learning.ipynb supports lower-complexity model comparisons. Baseline result sets are separated by algorithm in logs/linear_Regression/, logs/SVR/, and logs/Xgboost/; each exposes prediction_per_actual.csv for consistent downstream RMSE/correlation analysis. The colocated plotter.py files, e.g. logs/SVR/plotter.py, are result-visualization entry points rather than training infrastructure.
Evaluation and Runtime Boundaries
There is no packaged service, deployment runtime, or declared dependency manifest: execution is notebook-driven, with Python ML/DL runtimes implied by the notebooks and MATLAB used only for data import/inspection. Persistent artifacts are filesystem-based CSV, text hyperparameters, and optional training histories; no database service or cloud infrastructure is represented. An alarm audio asset (codes/Alarm.mp3) may support interactive/notebook workflow notification, not localization inference.
Components 17
Data preparation
- Experimental datasets · Raw and derived sensor data for localization experiments.
datasets/README.md - Dataset construction
codes/create_DB.ipynb - MATLAB import
codes/Matlab_files/importfile.m - MATLAB inspection
codes/Matlab_files/show_data_in_MATLAB.m
Localization experiments
- Electric-field signals
- Classical ML baselines · Regression, SVR, and XGBoost comparisons.
codes/machine_learning.ipynb - CNN localization · Advanced CNN experimentation branch.
codes/CNN.ipynb - Triple-sensor localization · Multi-sensor input branch.
codes/Thriple_sensor.ipynb - Predicted 3D location
Persistent results
- Baseline predictions · Comparable prediction-versus-actual outputs by algorithm.
- Single-sensor predictions
logs/1D_final_case_study_one/prediction_per_actual.csv - CNN case outputs · Axis and geometry-specific CNN experiment outputs.
- CNN training metadata
logs/1D_final_case_study_three_upscale_x/history.csv - Triple-sensor outputs
- SNN experiment outputs · Alternative neural-model prediction artifacts.
- Result visualization · Plots persisted baseline results.
logs/SVR/plotter.py - RMSE and correlation analysis · Compares predicted coordinates with actual locations.
Connections 23
- Experimental datasets → Dataset construction: defines input data
- MATLAB import → Experimental datasets: imports data
- Experimental datasets → MATLAB inspection: inspected in
- Dataset construction → Electric-field signals: constructs
- Experimental datasets → Electric-field signals: supplies
- Electric-field signals → Classical ML baselines: single-sensor input
- Electric-field signals → CNN localization: model input
- Electric-field signals → Triple-sensor localization: multi-sensor input
- Classical ML baselines → Predicted 3D location: predicts
- CNN localization → Predicted 3D location: predicts
- Triple-sensor localization → Predicted 3D location: predicts
- Classical ML baselines → Baseline predictions: persists
- CNN localization → Single-sensor predictions: persists cases
- CNN localization → CNN case outputs: persists cases
- CNN localization → CNN training metadata: records training
- Triple-sensor localization → Triple-sensor outputs: persists
- CNN localization → SNN experiment outputs: runs alternative neural cases
- Baseline predictions → Result visualization: visualized by
- Baseline predictions → RMSE and correlation analysis: compared
- Single-sensor predictions → RMSE and correlation analysis: compared
- CNN case outputs → RMSE and correlation analysis: compared
- Triple-sensor outputs → RMSE and correlation analysis: compared
- SNN experiment outputs → RMSE and correlation analysis: compared