spectrogram-anomaly-ae

Code accompanying the paper on unsupervised anomaly detection in vibration signals using convolutional autoencoders and spectrogram representations.

Dataset

This repository applies the proposed methodology to the public turning/chatter diagnosis dataset hosted on Mendeley Data:

Khasawneh, F., Otto, A., & Yesilli, M. (2019). Turning Dataset for Chatter Diagnosis Using Machine Learning (Version 1) [Data set]. Mendeley Data. https://doi.org/10.17632/hvm4wh3jzx.1

The dataset is distributed as .mat files containing the time vector (t) and multi-sensor vibration signals (d). In this repository, we extract the accelerometer channels, convert windowed time-series into spectrogram representations (RGB images), and use these inputs to run unsupervised anomaly detection experiments with convolutional autoencoders.

Notebook Workflow

The experiment workflow is organized as a numbered notebook series:

Notebook	Purpose
notebooks/01_Load_Data_Segmentation_Labeling.ipynb	Download/extract the original Mendeley dataset and create labeled 2.5-second vibration windows.
notebooks/02_Create_Frozen_Splits_and_Manifests.ipynb	Create deterministic train/validation/test manifests.
notebooks/03_Create_Spectrogram_Datasets.ipynb	Generate spectrogram image datasets from the frozen manifest.
notebooks/04_Train_CNN_AE_BN16_150x100px.ipynb	Train the main CNN autoencoder on nominal training samples.
notebooks/05_Evaluate_AE_Scores_and_Thresholds.ipynb	Score AE reconstructions and freeze validation-selected thresholds.
notebooks/06_Baseline_Comparisons.ipynb	Evaluate classical anomaly-detection baselines on the same split.
notebooks/07_VER_Ablation_and_Sensitivity.ipynb	Run VER ablation and segmentation sensitivity experiments.
notebooks/08_Bootstrap_CIs_and_Report_Tables.ipynb	Generate metrics, confidence intervals, figures, and report tables.
notebooks/09_Resolution_Contamination_Axis_Studies.ipynb	Run resolution, contamination, and vibration-axis studies.
notebooks/10_Error_Analysis_and_Deployment_Benchmark.ipynb	Analyze errors and benchmark inference.
notebooks/11_Method_Documentation_and_Citation_Cleanup.ipynb	Generate method documentation tables and citation cleanup notes.
notebooks/12_Publication_Quality_Figures_and_Tables.ipynb	Create paper-ready PDF/SVG/PNG figures and CSV/LaTeX tables.

The first notebook downloads the original Mendeley dataset, extracts the raw .mat files, and creates labeled 2.5-second windowed vibration segments.

The notebook writes extracted raw data to data/raw_mat and saves processed segments to data/01_windowed_labeled_2,5s.

File Format (.npz)

Each file contains a single 2.5-second vibration segment:

Field	Description
t	time vector
X, Y, Z	accelerometer signals (3-axis vibration)
label	binary class (chatter / no_chatter)
A_time	RMS vibration amplitude
is_chatter	boolean decision from spectral analysis