7.7.1. Generic Time Series Classification

This example serves as a “Hello World” introduction to time series classification using the TinyML ModelMaker toolchain. It demonstrates how to use any generic time series classification task with our toolchain.

7.7.1.1. Overview

Task: Time series classification (3 classes)
Dataset: Synthetic waveforms (sine, square, sawtooth)
Model: CLS_1k_NPU (~1,000 parameters)
Target: F28P55 (NPU device)
Training time: ~2-5 minutes on CPU

7.7.1.2. Running the Example

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

cd tinyml-modelzoo
run_tinyml_modelzoo.bat examples\generic_timeseries_classification\config.yaml

7.7.1.3. Understanding the Dataset

The dataset contains three classes of synthetic waveforms:

Sine wave - Smooth sinusoidal pattern
Square wave - Alternating high/low values
Sawtooth wave - Linear ramp pattern

Each waveform is a 1D time series with 1024 samples.

7.7.1.4. Dataset Format

For classification tasks, ModelMaker expects the dataset in this structure:

{dataset_name}.zip/
|
|-- classes/
      |-- class_1/
      |     |-- file1.csv
      |     |-- file2.csv
      |
      |-- class_2/
      |     |-- file1.csv
      |     |-- file2.csv
      |
      |-- class_N/
            |-- ...

Note

For classification tasks, the folder names under classes/ become the class labels. No separate annotation files are required.

7.7.1.5. Configuration

common section:

common:
  task_type: generic_timeseries_classification
  target_device: F28P55

dataset section:

dataset:
  dataset_name: generic_timeseries_classification

data_processing_feature_extraction section:

data_processing_feature_extraction:
  feature_extraction_name: Generic_1024Input_FFTBIN_64Feature_8Frame
  variables: 1

training section:

training:
  model_name: CLS_1k_NPU
  batch_size: 256
  training_epochs: 20
  num_gpus: 0

testing: {}
compilation: {}

7.7.1.6. Feature Extraction

The Generic_1024Input_FFTBIN_64Feature_8Frame preset:

Applies 1024-point FFT to convert to frequency domain
Bins the FFT output into 64 frequency features
Concatenates 8 frames for temporal context
Total input to model: 64 x 8 = 512 features

7.7.1.7. Evaluation Metrics

Accuracy

Percentage of correctly classified samples.

Range: 0% to 100%
Ideal value: 100% (higher is better)

F1-Score

Harmonic mean of precision and recall.

Range: 0 to 1
Ideal value: 1 (higher is better)

7.7.1.8. Expected Results

Float32 Model:
Accuracy: 98-100%
F1-Score: ~1.0

Quantized Model:
Accuracy: 95-100%

7.7.1.9. Output Location

Results are saved to:

../tinyml-modelmaker/data/projects/generic_timeseries_classification/run/<timestamp>/CLS_1k_NPU/

Key files:

training/base/best_model.pt - Trained model
training/quantization/best_model.onnx - Quantized ONNX
compilation/artifacts/mod.a - Compiled for device

7.7.1.10. Variations to Try

Different Feature Extraction

Try raw time domain instead of FFT:

data_processing_feature_extraction:
  feature_extraction_name: Generic_512Input_RAW_512Feature_1Frame

Different Model Size

Try a smaller or larger model:

training:
  model_name: CLS_100_NPU  # Smaller
  # or
  model_name: CLS_4k_NPU   # Larger

Different Target Device

Compile for a different device:

common:
  target_device: MSPM0G3507

7.7.1.11. Next Steps

After successfully running this generic example:

Try Arc Fault Detection - Real-world arc fault detection
Explore Classification Dataset Format - Use your own classification data
Read Understanding Config - Learn all config options