7.7.4. Generic Time Series Anomaly Detection

This example serves as a “Hello World” introduction to time series anomaly detection using the TinyML ModelZoo toolchain. It demonstrates how to use autoencoder-based anomaly detection with our toolchain.

7.7.4.1. Overview

Task: Time series anomaly detection (autoencoder-based)
Dataset: Synthetic sinusoidal signals with frequency/amplitude anomalies
Model: AD_17k (~17,000 parameters)
Target: F28P55 (NPU device)

7.7.4.2. Running the Example

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

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

7.7.4.3. Understanding the Dataset

The dataset uses a synthetic sinusoidal pattern:

Normal Pattern:

Signal: y = 1.2 sin(2pi*f*t) + 0.8 cos(2pi*f*t)
Base frequency: f = 1.0 Hz
Amplitude variation: +/- 10%
Frequency variation: +/- 5%

Anomaly Types:

Type	Description
Frequency Faster	Signal frequency increases by 30-70%
Frequency Slower	Signal frequency decreases by 20-50%
Amplitude Higher	Signal amplitude increases by 60-120%
Amplitude Lower	Signal amplitude decreases by 40-70%

Dataset split:

Normal samples: 60 files (train: 50%, val: 10%, test: 40%)
Anomaly samples: 32 files (all in test set)

Each file contains 5000 samples (50 seconds at 100 Hz).

7.7.4.4. Dataset Format

For anomaly detection, ModelZoo expects this folder structure:

dataset_name/
|
|-- classes/
      |-- Normal/
      |     |-- normal_0000.csv
      |     |-- normal_0001.csv
      |     |-- ...
      |
      |-- Anomaly/
            |-- freq_faster_0000.csv
            |-- freq_slower_0000.csv
            |-- amp_higher_0000.csv
            |-- amp_lower_0000.csv
            |-- ...

Important

The autoencoder is trained only on Normal data. All Anomaly files are used exclusively for testing.

7.7.4.5. Configuration

common section:

common:
  target_module: 'timeseries'
  task_type: 'generic_timeseries_anomalydetection'
  target_device: 'F28P55'

dataset section:

dataset:
  enable: True
  dataset_name: generic_timeseries_anomalydetection
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/01_03_00/datasets/generic_timeseries_anomalydetection.zip
  split_factor: [0.5, 0.1, 0.4]

data_processing_feature_extraction section:

data_processing_feature_extraction:
  data_proc_transforms:
    - SimpleWindow
    - Downsample
  frame_size: 100
  sampling_rate: 100
  new_sr: 10
  variables: 1

training section:

training:
  enable: True
  model_name: 'AD_17k'
  batch_size: 64
  training_epochs: 50
  num_gpus: 1
  learning_rate: 0.001
  quantization: 1
  output_int: False

Important

output_int must be False for anomaly detection since reconstruction error must be in float for threshold comparison.

7.7.4.6. How Autoencoder Detection Works

Training Phase (Normal Data Only):

Autoencoder learns to compress and reconstruct normal patterns
For normal samples: Reconstruction error is LOW (~0.008)
Model learns: “This is what normal looks like”

Testing Phase (Normal + Anomaly Data):

Normal sample:
  Reconstruction Error: ~0.008 -> Correctly identified as normal

Anomaly sample:
  Reconstruction Error: ~0.69 -> Flagged as anomaly (88x higher!)

Detection Logic:

if reconstruction_error > threshold:
    -> ANOMALY
else:
    -> NORMAL

7.7.4.7. Why Frame Size Matters

Anomaly detection requires temporal context - the model must see multiple consecutive samples to understand the pattern.

Frame Size	Cycles Visible	Detection Quality
1 - 10	< 0.1 cycles	Insufficient
20 - 50	0.2 - 0.5	Weak (partial cycle)
100	1.0 cycle	Good (full cycle visible)
>200	2.0+ cycles	Excellent

Rule of thumb: Frame size should capture at least 1 complete cycle.

7.7.4.8. Expected Results

Reconstruction Error Statistics:

Data Type	Mean	Std Dev
Normal training data	0.0078	0.0022
Normal test data	0.0081	0.0023
Anomaly test data	0.6898	0.4616

Performance at Threshold k=4.5:

Metric	Value
Threshold	0.0178
Accuracy	99.97%
Precision	99.95%
Recall (Detection Rate)	100.00%
F1 Score	99.97%
False Positive Rate	0.07%

Confusion Matrix:

	Predicted Normal	Predicted Anomaly
Ground Truth: Normal	9,617 (TN)	7 (FP)
Ground Truth: Anomaly	0 (FN)	12,832 (TP)

The model successfully detects all 4 anomaly types with 100% recall!

7.7.4.9. Output Location

Results are saved to:

../tinyml-modelmaker/data/projects/{dataset_name}/run/{date-time}/{model_name}/

Key outputs:

training/base/ - Float model training results
training/quantization/ - Quantized model results
training/quantization/post_training_analysis/ - Threshold analysis
training/quantization/post_training_analysis/threshold_performance.csv - Metrics table
training/quantization/post_training_analysis/reconstruction_error_histogram.png - Error distribution
compilation/artifacts/mod.a - Compiled for device

7.7.4.10. Next Steps

After successfully running this generic example:

Try Anomaly Detection Example - DC arc fault anomaly detection
Try Fan Blade Fault Classification - Fan blade anomaly detection
Try Motor Bearing Fault - Motor vibration anomaly detection
Read Anomaly Detection - Learn more about anomaly detection