7.9.30. Anomaly Detection Examples Overview

In 1.4.0 this page was a walkthrough of the DC arc fault anomaly detection example. In 1.5.0 the anomaly detection configs are part of several example folders, so this page is an overview that lists all of them and collects the common workflow: normal-only training, thresholds and how to read the outputs.

For a minimal first run use Generic Timeseries Anomaly Detection. For the theory and dataset format see Anomaly Detection and Anomaly Detection Dataset Format.

7.9.30.1. Anomaly Detection Configs in 1.5.0

All of them use task_type: generic_timeseries_anomalydetection and are run from the tinyml-modelzoo folder with ./run_tinyml_modelzoo.sh examples/<folder>/<config>.yaml (Windows: see Running Examples).

Config	Device	Model	Page
`generic_timeseries_anomalydetection/config.yaml`	F29H85	`AD_17k`	Generic Timeseries Anomaly Detection
`generic_timeseries_anomalydetection/config_NPU.yaml`	F28P55	`AD_11k_NPU`	Generic Timeseries Anomaly Detection
`dc_arc_fault/config_anomaly_detection_dsk.yaml`	F28P55	`AD_17k`	DC Arc Fault Detection
`dc_arc_fault/config_anomaly_detection_dsi.yaml`	F28P55	`AD_17k`	DC Arc Fault Detection
`motor_bearing_fault/config_anomaly_detection.yaml`	F28P55	`AD_Linear`	Motor Bearing Fault
`motor_bearing_fault/config_anomaly_detection_MSPM0.yaml`	MSPM0G5187	`AD_500_NPU`	Motor Bearing Fault
`fan_blade_fault_classification/config_anomaly_detection.yaml`	F28P55	`AD_17k`	Fan Blade Fault Classification
`ecg_classification/config_anomaly_detection.yaml`	F28P55	`AD_Linear`	ECG Classification

In the automated 1.5.0 config test run the generic config.yaml, the DC arc fault DSI config, the ECG, fan blade and motor bearing (F28P55) configs completed. The DC arc fault DSK config stopped with a NaN/Inf loss at epoch 0, the motor bearing MSPM0 config failed with a quantization scale/zero-point dimension error, and the generic config_NPU.yaml failed with a CUDA device mismatch. Check the page of each example for its status.

7.9.30.2. Workflow

The autoencoder is trained on the Normal class only. The Anomaly class is used for testing and threshold evaluation.
After training, the mean and standard deviation of the reconstruction error on normal training data are computed and thresholds threshold = mean_train + k * std_train are generated for a range of k.
At inference time a window with a reconstruction error above the threshold is an anomaly.

A typical anomaly detection config (the DC arc fault DSK variant keeps the same structure) has these elements:

common:
  task_type: generic_timeseries_anomalydetection
  target_device: F28P55
dataset:
  dataset_name: <name>
  input_data_path: <zip url or path>      # classes/Normal and classes/Anomaly
data_processing_feature_extraction:
  data_proc_transforms: [...]             # or feature_extraction_name: <preset>
  variables: 1
training:
  model_name: AD_17k
  quantization: 1
  output_int: false                       # must be false for anomaly detection
testing:
  enable: True
compilation:
  enable: True

7.9.30.3. Available Models

The anomaly detection models enabled in tinyml_modelzoo/model_descriptions/anomalydetection.py are: AD_500_NPU, AD_1k, AD_2k_NPU, AD_4k, AD_6k_NPU, AD_8k_NPU, AD_10k_NPU, AD_11k_NPU, AD_16k, AD_17k, AD_20k_NPU, AD_Linear and Ondevice_Trainable_AD_Linear (used by the on-device learning example, see On-Device Training Documentation). Models with the _NPU suffix are designed for devices with an NPU. Start with a small model for simple, repetitive patterns and move to a larger one for complex signals.

7.9.30.4. Threshold Selection

Lower k detects more anomalies but raises more false alarms.
Higher k reduces false alarms but may miss subtle anomalies.
threshold_performance.csv in post_training_analysis/ lists threshold, accuracy, precision, recall, F1 score and false positive rate for each k. Pick the row that balances them for your application.
reconstruction_error.png and reconstruction_error_log_scale.png in the same folder show the reconstruction error distributions of normal and anomaly data. Well separated distributions mean good detection capability; overlapping ones call for different features or a different model.

7.9.30.5. Anomaly Detection versus Classification

Aspect	Classification	Anomaly detection
Training data	All classes needed	Normal data only
Unknown faults	Cannot be detected	Detected as anomalies
Fault identification	Identifies the fault type	Only flags abnormality

7.9.30.6. Troubleshooting

High false positive rate: increase k, or add more diverse normal training data.
Missed anomalies: decrease k, increase model size, or use features that expose the anomaly.
Unstable reconstruction error: train longer, check preprocessing, and compare with the float model.