There are two primary methods to collect datasets for grid fault detection:

1) Using commercial grid monitoring devices

• Easiest approach due to devices automatically collecting and reporting grid data during adverse events our provided dataset was obtained using this method

2) Using grid-connected converters

• Deploy the same converter that will eventually run the detection algorithm?
• Operate in grid-connected mode for extended periods

Correctly identifying and tagging anomalous event intervals

The classification of AC grid faults presents significant challenges due to:

• High diversity in fault characteristics across datasets
• Absence of standardized fault classes
• Need to prioritize faults relevant to On-Board Charger (OBC) operation

To overcome these challenges, a proprietary semi-supervised annotation technique was developed that combines:

• Human expert annotationLeverages domain expertise to identify critical fault patterns
• Unsupervised hierarchical clustering
• Automatically groups similar fault patterns
• Helps identify natural categories in the data

The annotation quality is rigorously evaluated using dimensionality reduction visualization techniques which allows for verification of clear separation between fault classes and helps identify potential misclassifications or ambiguous cases.

This hybrid approach ensures fault categories are both technically meaningful for OBC operation and statistically distinguishable in the data.

?

With your dataset ready, explore, train, and evaluate models using the TI tinyML Model Zoo?

Access our library of optimized generic time-series models, scalable across performance and power requirements.

For this use case , refer to the model in the TI tinyML Model Zoo? under Examples → Grid fault detection.

Start with our command line tools, our tinyml-modelzoo hosts a more extensive and flexible set of model-development tools, including Bring your own data (BYOD) or? Bring your own model (BYOM)