3.2. First Example

A detailed walkthrough of DC arc fault detection, from config to compiled model.

3.2.1. Overview

DC arc faults are dangerous electrical arcs in DC systems (solar, batteries). The model classifies windows of current samples as class_0_normal or class_1_arc. The outputs shown below are from a real run of this example on the F28P55 with CLS_1k_NPU.

3.2.2. Step 1: The Configuration

examples/dc_arc_fault/config_dsk.yaml (comments omitted):

common:
  task_type: arc_fault
  target_device: F28P55
dataset:
  dataset_name: dc_arc_fault_example_dsk
data_processing_feature_extraction:
  feature_extraction_name: FFT1024Input_256Feature_1Frame_Full_Bandwidth
  variables: 1
training:
  model_name: CLS_1k_NPU
  batch_size: 32
  training_epochs: 10
  num_gpus: 0
  early_stopping: true
  early_stopping_patience: 5
testing: {}
compilation: {}
common

task_type: arc_fault selects the arc fault classification task and target_device: F28P55 is an NPU device.

dataset

dataset_name: dc_arc_fault_example_dsk is a built-in dataset preset that is downloaded and extracted automatically.

data_processing_feature_extraction

A 1024-point FFT per frame with 256 features, one frame, one variable (current).

training

The 1K-parameter NPU classification model. Training stops early when validation accuracy has not improved for 5 epochs.

3.2.3. Step 2: Run

./run_tinyml_modelzoo.sh examples/dc_arc_fault/config_dsk.yaml

The console shows, in order: dataset download, feature extraction, float training, quantization-aware training, testing, and compilation.

3.2.4. Step 3: Read the Training Log

The log reports validation accuracy per epoch and a summary of the best epoch, with accuracy, F1 score, AUC ROC and the confusion matrix. When early stopping triggers, the log says so (Early stopping at epoch N: no improvement for 5 epochs) and the best checkpoint is kept. The test-set result is printed as Test Data Evaluation Accuracy.

3.2.5. Step 4: Examine the Output Folder

The output goes to ./data/projects/dc_arc_fault_example_dsk/run/<date-time>/CLS_1k_NPU/.

Path

Contents

training/base/

Float model (model.onnx, checkpoint.pth), run.log, PCA plots, golden vectors

training/quantization/

Quantized model, golden vectors, post_training_analysis/ (ROC curve, score histogram, fpr_tpr_thresholds.csv)

compilation/artifacts/

mod.a (compiled model library) and tvmgen_default.h (C API)

compilation/<date-time>_CLS_1k_NPU_F28P55.zip

Bundle with everything a device project needs

compilation/run.log

Compiler log with the memory usage of the model library

Every item is explained in Understanding the Run Output.

3.2.6. Step 5: Customize

training:
  model_name: CLS_4k_NPU       # a larger model (check the model name exists in the model zoo)

common:
  target_device: F2837         # a non-NPU C2000 device

Use your own data by changing the dataset section (see How to Create a ModelZoo Config YAML and Bring Your Own Data). Model names available for your task and device are listed under Model Zoo Reference.

3.2.7. Next Steps