3.2. First Example

This guide provides a detailed walkthrough of running an arc fault detection example from start to finish.

3.2.1. Overview

DC Arc Fault Detection is a common application where we need to identify dangerous electrical arcs in DC power systems (solar panels, batteries, etc.).

We’ll train a model to classify current waveforms as either “normal” or “arc”.

3.2.2. Step 1: Examine the Configuration

Open examples/dc_arc_fault/config.yaml:

common:
  target_module: 'timeseries'
  task_type: 'generic_timeseries_classification'
  target_device: 'F28P55'
  run_name: '{date-time}/{model_name}'

dataset:
  enable: True
  dataset_name: 'dc_arc_fault_example_dsk'
  input_data_path: 'https://software-dl.ti.com/C2000/esd/mcu_ai/01_03_00/datasets/dc_arc_fault_example_dsk.zip'

data_processing_feature_extraction:
  feature_extraction_name: 'FFT1024Input_256Feature_1Frame_Full_Bandwidth'
  variables: 1

training:
  enable: True
  model_name: 'ArcFault_model_200_t'
  batch_size: 256
  training_epochs: 20

testing:
  enable: True

compilation:
  enable: True

Configuration Breakdown:

common Section

  • target_module: 'timeseries' - Working with time series data

  • task_type: 'generic_timeseries_classification' - Binary/multi-class classification

  • target_device: 'F28P55' - Compile for F28P55 (NPU device)

dataset Section

  • The dataset is automatically downloaded from the URL

  • Contains current waveforms labeled as “arc” or “normal”

data_processing_feature_extraction Section

  • Uses a preset that applies 1024-point FFT

  • Extracts 256 frequency features per frame

  • variables: 1 - Single channel (current only)

training Section

  • Uses a specialized arc fault model (~200 parameters)

  • Trains for 20 epochs with batch size 256

3.2.3. Step 2: Run Training

cd tinyml-modelzoo
./run_tinyml_modelzoo.sh examples/dc_arc_fault/config.yaml

You’ll see output showing:

  1. Dataset download and extraction

  2. Data preprocessing and feature extraction

  3. Training progress (loss, accuracy per epoch)

  4. Testing results

  5. Quantization

  6. Compilation

3.2.4. Step 3: Understand Training Output

During training, you’ll see something like:

Epoch 1/20
Train Loss: 0.6543 | Train Acc: 62.34%
Val Loss: 0.4521 | Val Acc: 78.92%

Epoch 10/20
Train Loss: 0.0234 | Train Acc: 99.12%
Val Loss: 0.0312 | Val Acc: 98.87%

Epoch 20/20
Train Loss: 0.0089 | Train Acc: 99.89%
Val Loss: 0.0198 | Val Acc: 99.45%

Best model saved at epoch 18 with val accuracy: 99.52%

Key metrics:

  • Train Loss/Acc: Performance on training data

  • Val Loss/Acc: Performance on held-out validation data

  • Best model: Checkpoint with highest validation accuracy

3.2.5. Step 4: Examine Output Files

Navigate to the output directory:

../tinyml-modelmaker/data/projects/dc_arc_fault_example_dsk/run/<timestamp>/ArcFault_model_200_t/

Training Outputs (training/base/):

best_model.pt

PyTorch checkpoint of best model

training_log.csv

Epoch-by-epoch metrics

pca_on_feature_extracted_train_data.png

PCA visualization of features

One_vs_Rest_MultiClass_ROC_test.png

ROC curve (if classification)

Histogram_Class_Score_differences_test.png

Class score distribution

Quantization Outputs (training/quantization/):

best_model.onnx

Quantized ONNX model

golden_vectors/test_vector.c

Test inputs/outputs for device validation

golden_vectors/user_input_config.h

Feature extraction config for device

Compilation Outputs (compilation/artifacts/):

mod.a

Compiled model library

tvmgen_default.h

C API for model inference

3.2.6. Step 5: Analyze Results

Check Training Curves

Open training/base/training_log.csv or the generated plots:

  • Loss should decrease over epochs

  • Accuracy should increase and stabilize

  • Val metrics should track train metrics (no overfitting)

Check Classification Performance

The ROC curve shows true positive vs false positive rates:

  • Area Under Curve (AUC) close to 1.0 indicates good performance

  • Review threshold selection for your application needs

Check Quantization Impact

Compare float32 vs quantized accuracy in the console output:

  • Small accuracy drop (<2%) is normal

  • Large drop indicates quantization issues

3.2.7. Step 6: Customize the Example

Try a Different Model

Edit the config to use a larger model:

training:
  model_name: 'ArcFault_model_1400_t'  # ~1400 parameters

Try a Different Device

Compile for a non-NPU device:

common:
  target_device: 'F2837'  # Non-NPU C2000 device

Use Your Own Data

Replace the dataset section with your own data path:

dataset:
  dataset_name: 'my_arc_fault_data'
  input_data_path: '/path/to/your/dataset'

See Bring Your Own Data for dataset format requirements.

3.2.8. Next Steps