7.9.6. AC Arc Fault

AC arc fault detection is an Edge AI solution for the MSPM0G5187 microcontroller with an on-chip NPU. It detects series arc faults in residential and commercial electrical systems, a leading cause of electrical fires, using ML inference on data captured through the TIDA-010971 analog front end with a PCB Rogowski coil. The goal is Arc Fault Circuit Interrupter (AFCI) products that comply with NEC requirements.

Example folder: examples/ac_arc_fault (GitHub). For DC arc faults, see DC Arc Fault Detection.

7.9.6.1. Overview

Task: binary classification (task_type: arc_fault)
Device: MSPM0G5187 (Arm Cortex-M0+ with NPU), config config_MSPM0.yaml
Dataset: ac_arc_fault_log300_example_dsk
Model: CLS_1k_NPU

7.9.6.2. Why AC Arc Fault Detection?

Arc faults cause more than 30,000 home fires each year in the United States, and traditional circuit breakers are not designed to detect low-level arcs. Edge AI enables complex pattern recognition, multi-feature analysis and adaptive detection across load types. The targets of the reference solution are:

Metric	Target
Response time	less than 10 ms
Detection accuracy	more than 95%
Compliance	UL 1699

7.9.6.3. System Components

Hardware: MSPM0G5187 and the TIDA-010971 analog front end (PCB Rogowski coil)
Software: Code Composer Studio 12.x or later, MSPM0 SDK 2.08.00 or later

7.9.6.4. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

./run_tinyml_modelzoo.sh examples/ac_arc_fault/config_MSPM0.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\ac_arc_fault\config_MSPM0.yaml

7.9.6.5. Configuration

common:
  task_type: arc_fault
  target_device: MSPM0G5187
dataset:
  dataset_name: ac_arc_fault_log300_example_dsk
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/ac_arc_fault_log300.zip
data_processing_feature_extraction:
  feature_extraction_name: ArcFault_512Input_FE_RFFT_32Feature_8Frame_1InputChannel_removeDC_Full_Bandwidth
  store_feat_ext_data: false
training:
  model_name: CLS_1k_NPU
  batch_size: 30
  training_epochs: 30
  num_gpus: 0
  quantization: 2
  learning_rate: 0.001
  weight_decay: 0.00001
testing: {}
compilation: {}

quantization: 2 enables quantization-aware training (INT8), which is required for the MSPM0G5187 NPU.

7.9.6.6. Dataset

ac_arc_fault_log300.zip contains 100+ captures per load type:

Normal operation: inductive loads, LED lighting, switched-mode power supplies, mixed loads
Arc faults: series and parallel arc faults
Masking loads: dimmer, SMPS, drill, compressor, resistive loads, vacuum cleaner

To prepare and label your own AC arc fault captures use examples/ac_arc_fault/AFCI_LabellingScript.py; instructions are in examples/ac_arc_fault/readme_labelling.md.

7.9.6.7. Feature Extraction

The preset ArcFault_512Input_FE_RFFT_32Feature_8Frame_1InputChannel_removeDC_Full_Bandwidth mirrors the on-device pipeline:

ADC capture: 512 samples at 107 kSps (about 4.76 ms per frame)
512-point real FFT (ARM CMSIS-DSP on device)
Complex magnitude
DC removal
Binning of 8 adjacent bins, giving 32 features per frame
Normalization to the INT8 range
Concatenation of 8 frames, giving 256 features per input

7.9.6.8. Models

Any enabled classification model can be used. In 1.5.0 the arc fault specific models are enabled in the ModelZoo and can be selected with training.model_name, next to the generic NPU models.

Model	Approx. parameters	Comment
`ArcFault_model_200_t`	200	Smallest and fastest
`ArcFault_model_300_t`	300
`ArcFault_model_700_t`	800	Balance of speed and memory
`ArcFault_model_1400_t`	1,600	Most accurate
`CLS_500_NPU` / `CLS_1k_NPU` / `CLS_2k_NPU` / `CLS_4k_NPU`	500 to 4,000	Generic NPU models (`CLS_1k_NPU` is the default in the config)

7.9.6.9. Results

The example readme.md reports these figures for the reference solution on MSPM0G5187 with NPU (recommended model, 8-frame voting):

Metric	Value
End-to-end latency	less than 150 ms
Detection accuracy	more than 99%
False positive rate	0.01%
Precision / Recall / F1	99.97% / 99.72% / 99.84%

In the 1.5.0 config test run, config_MSPM0.yaml with CLS_1k_NPU reached a test accuracy of 99.64% with an AUC ROC of 1.000.

7.9.6.10. Training and Deployment Flow

Running the config performs feature extraction, training, INT8 quantization and compilation with the TI Neural Network Compiler, producing the model library and interface headers for the MSPM0 SDK. See Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check) and NPU Device Deployment.