# Voltage Level Classification Module - PD-Defect

Collection, licensing, and citation are documented on the [main dataset card](../README.md).

## TL;DR

Object detection on pole-top crops that labels each Crossarm or V-Crossarm instance as high tension (HT) or low tension (LT). In this release, HT comprises 11 kV and 33 kV, whereas LT comprises 220 V and 440 V service. The subset contains 1,772 crops and 2,496 annotated instances.

## Task Definition

Detect voltage-informative pole-top structures and assign an HT or LT label from the visible insulator configuration, following APEPDCL field practice (insulator count, geometric arrangement, and placement relative to the crossarm).

## Label Definitions

- `high_tension` (HT): insulator configurations associated with 11 kV and 33 kV assets (not distinguished as separate classes).
- `low_tension` (LT): insulator configurations associated with 220 V and 440 V service in the collected APEPDCL network (not distinguished as separate classes).



## Annotation Protocol

Voltage is assigned solely from the visible insulator configuration. Mild perspective distortion is permitted when that configuration remains distinguishable.

Component bounding boxes include their associated insulators, as the insulator configuration provides the basis for voltage-level classification. Each box tightly encloses the structural member together with its associated insulators and any contiguous field-fabricated or welded extensions, which inherit the voltage label. This scope differs from Crossarm and Top-Cleat Tilt Assessment, where such extensions are excluded so that Straight / Tilted judgment is not influenced by attachments that do not share the member's orientation.

Top Cleat instances are not annotated as independent objects. In post-processing, each Top Cleat inherits the voltage label of its associated V-Crossarm.

## Annotation Format

YOLO object detection:

```text
<class_id> <center_x> <center_y> <width> <height>
```

- Class 0: `high_tension`
- Class 1: `low_tension`



## Dataset Structure

```text
voltage_level/
├── data.yaml
├── dataset.json
├── train/
│   ├── images/
│   ├── labels/
│   └── train.json
├── valid/
│   ├── images/
│   ├── labels/
│   └── val.json
└── test/
    ├── images/
    ├── labels/
    └── test.json
```



## Train / Validation / Test Split

Train 1,337 (75.45%) / valid 218 (12.30%) / test 217 (12.25%); total 1,772.

- Annotated instances: 2,496 (`high_tension` 1,585; `low_tension` 911)



## Baseline Benchmarks

YOLO12m (`high_tension`, `low_tension`). Overall box metrics are given in the split headers; class-wise metrics follow.

`valid` — Precision 91.12%; Recall 91.04%; `mAP@0.5` 95.73%; `mAP@0.5:0.95` 82.87%


| Class        | Precision | Recall | `mAP@0.5` | `mAP@0.5:0.95` |
| ------------ | --------- | ------ | --------- | -------------- |
| high_tension | 94.50%    | 91.40% | 96.24%    | 86.90%         |
| low_tension  | 87.74%    | 90.68% | 95.22%    | 78.85%         |


`test` — Precision 93.83%; Recall 91.40%; `mAP@0.5` 95.72%; `mAP@0.5:0.95` 81.68%


| Class        | Precision | Recall | `mAP@0.5` | `mAP@0.5:0.95` |
| ------------ | --------- | ------ | --------- | -------------- |
| high_tension | 94.62%    | 93.63% | 97.76%    | 88.87%         |
| low_tension  | 93.04%    | 89.17% | 93.68%    | 74.48%         |




### Image-level metrics

Image-level presence/absence per class at detection confidence 0.35 (binary Accuracy, Precision, Recall, and F1 over images).

`valid` (218 images)


| Class        | Accuracy | Precision | Recall | F1     |
| ------------ | -------- | --------- | ------ | ------ |
| high_tension | 95.41%   | 98.15%    | 95.78% | 96.95% |
| low_tension  | 92.66%   | 91.53%    | 94.74% | 93.10% |


`test` (217 images)


| Class        | Accuracy | Precision | Recall | F1     |
| ------------ | -------- | --------- | ------ | ------ |
| high_tension | 97.70%   | 98.82%    | 98.24% | 98.53% |
| low_tension  | 91.71%   | 90.76%    | 93.91% | 92.31% |




## Limitations

Limitations specific to this module:

- Samples with slight perspective distortion were included, because such distortion did not materially affect voltage-level identification from the visible insulator configuration.



## Examples

Random predictions on the test set. Boxes show predicted class and confidence.

![Voltage Level Classification examples](./examples/voltage_level_collage1.jpg)

![Voltage Level Classification examples (continued)](./examples/voltage_level_collage2.jpg)

## Citation

Please cite PD-Defect ([DOI 10.6084/m9.figshare.34329402](https://doi.org/10.6084/m9.figshare.34329402)); BibTeX is given on the [main dataset card](../README.md).