7.9.4. Generic Timeseries Anomaly Detection

A “Hello World” introduction to autoencoder-based timeseries anomaly detection, using a synthetic signal with four kinds of anomalies.

Example folder: examples/generic_timeseries_anomalydetection (GitHub)

7.9.4.1. Overview

  • Task: timeseries anomaly detection (generic_timeseries_anomalydetection), autoencoder trained on normal data only

  • Dataset: synthetic signal y = 1.2 sin(2 pi f t) + 0.8 cos(2 pi f t), 100 Hz sampling

  • Configs:

    • config.yaml: device F29H85, model AD_17k

    • config_NPU.yaml: device F28P55 (NPU), model AD_11k_NPU

7.9.4.2. The Dataset

Normal signals have a base frequency of 1 Hz (about +/-5% variation), an amplitude variation of up to +/-10% and Gaussian noise with a standard deviation of 0.05. The four anomaly types are:

Parameter

Normal range

Anomaly range

Frequency

0.95 - 1.05 Hz

Faster 1.3 - 1.7 Hz, slower 0.5 - 0.8 Hz

Amplitude (sine)

1.08 - 1.32

Higher 1.92 - 2.64, lower 0.36 - 0.72

Amplitude (cosine)

0.72 - 0.88

Higher 1.28 - 1.76, lower 0.24 - 0.48

Noise std dev

0.03 - 0.07

Same for all types

There are 60 normal files and 32 anomaly files (8 per type), 92 in total. Each file has 5000 samples (50 seconds at 100 Hz). The config downloads https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/generic_timeseries_anomalydetection.zip automatically.

7.9.4.3. Dataset Format and Splitting

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

Each CSV holds a single column of signal values. ModelZoo splits the data automatically with dataset.split_factor:

  • Training: 50% of the normal files only (the autoencoder learns normal behavior)

  • Validation: 10% of the normal files only

  • Test: the remaining 40% of the normal files plus all anomaly files

7.9.4.4. Why Frame Size Matters

Anomalies here are changes of frequency and amplitude, so the model must see at least one full cycle of the signal. With a 100 Hz signal and 1 Hz base frequency, a frame of about 100 samples covers one cycle. A single sample cannot distinguish a 1 Hz signal from a 1.5 Hz signal. As a rule of thumb, choose a frame that spans at least one full cycle of the pattern.

7.9.4.5. Running the Example

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

./run_tinyml_modelzoo.sh examples/generic_timeseries_anomalydetection/config.yaml

For the NPU variant use examples/generic_timeseries_anomalydetection/config_NPU.yaml.

7.9.4.6. Configuration

common:
  target_module: 'timeseries'
  task_type: 'generic_timeseries_anomalydetection'
  target_device: 'F29H85'
  run_name: '{date-time}/{model_name}'
dataset:
  enable: True
  dataset_name: generic_timeseries_anomalydetection
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/generic_timeseries_anomalydetection.zip
  split_factor: [0.5, 0.1, 0.4]
data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  - Downsample
  frame_size: 100
  sampling_rate: 100
  new_sr: 10
  variables: 1
training:
  model_name: AD_17k
  model_config: ''
  batch_size: 64
  training_epochs: 50
  early_stopping: true
  early_stopping_patience: 5
  num_gpus: 0
  quantization: 1
  output_int: false
testing:
  enable: True
compilation:
  enable: True

  • Downsample reduces the data from sampling_rate to new_sr before windowing.

  • output_int must be false for anomaly detection because the reconstruction error is compared to a float threshold.

  • config_NPU.yaml differs in the target device (F28P55), frame_size: 128, model_name: AD_11k_NPU, 30 training epochs (no early stopping), quantization: 2 and auto_quantization: False.

7.9.4.7. How It Works

  1. The autoencoder is trained on normal data only and learns to reconstruct it with a low error.

  2. After training, ModelZoo computes the mean and standard deviation of the reconstruction error on the normal training data and creates thresholds mean + k * std for a range of k values.

  3. On the test set, a window is flagged as an anomaly if its reconstruction error exceeds the threshold. Accuracy, precision, recall, F1 score and false positive rate are reported per k.

7.9.4.8. Results

The example readme documents a reference run with AD_17k (float best epoch 46 with validation MSE 0.008; quantized best epoch 5 with validation MSE 0.009):

Quantity

Value (readme run)

Reconstruction error, normal training data (mean / std)

0.007815 / 0.002208

Reconstruction error, normal test data (mean)

0.008051

Reconstruction error, anomaly test data (mean)

0.689792

Threshold at k = 4.5

0.017751

Accuracy / precision / recall / F1 at k = 4.5

99.97% / 99.95% / 100.00% / 99.97%

False positive rate at k = 4.5

0.07% (7 of 9624 normal windows)

The automated 1.5.0 test run of config.yaml also finished with 100% accuracy and F1 at k = 4.0. The test run of config_NPU.yaml failed with a CUDA device mismatch error during quantization-aware training (a tensor on CPU versus one on cuda:0). Treat the NPU variant as unverified in 1.5.0.

7.9.4.9. Output Location

Results are written to ./data/projects/generic_timeseries_anomalydetection/run/<date-time>/AD_17k/. The training folders contain the reconstruction-error plots and the threshold performance table; compiled artifacts are in compilation/artifacts/. See Understanding the Run Output.

7.9.4.10. Running on Device

Follow Deploying Anomaly Detection Models from ModelZoo to Device.

7.9.4.11. Next Steps