Ride anomaly scoring

Code and data to reproduce the results from "Unsupervised Anomaly Scoring of E-Scooter Inertial Signals for Safety Quantification" accepted at ECML PKDD 2026. The method assigns each e-scooter ride an anomaly score from raw IMU data — a 3-axis accelerometer and 3-axis gyroscope sampled at 10 Hz — without supervision, where higher scores indicate more reckless, crash-prone riding.

Pipeline: Raw IMU → FIR low-pass denoising → STFT spectrograms → time-average pooling → dimensionality reduction → anomaly scoring. We refer to the denoised time-average pooled spectrogram features as DTAPS.

Proposed method: DTAPS + Autoencoder + Local Outlier Factor — AP = 0.77 on GBG500; OR ≈ 2.4 per 1-SD (p < 0.001), ≈ 4.7 at the 95th score percentile (p < 0.001) on SRV3490 confirmed-crash data.

Table of Contents

• Ride anomaly scoring
  • Table of Contents
  • Prerequisites
  • Data
  • Notebooks
  • Running notebooks (batch)
  • Generate statistics from output
  • Results
    • GBG500 — Average Precision (30 seeds, 5-fold stratified CV)
    • GBG500 — Latent space (DTAPS + AE, out-of-fold, median AP seed)
    • SRV3490 — Odds Ratio Analysis (30 seeds, 5-fold group CV, DTAPS + AE + LOF)
  • How to cite

Prerequisites

The datasets are stored with Git LFS. Install it first, then clone the repository so the .parquet files are fetched:

git lfs install
git clone https://github.com/voiapp/ride-anomaly-scoring.git
cd ride-anomaly-scoring

Note: If you cloned before installing Git LFS, run git lfs pull to download the data.

Create and activate a virtual environment, then install dependencies:

python3.10 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Note: When running notebooks, ensure the kernel is set to use the .venv environment.

Data

• Location: data/
  • GBG500.parquet: Time index and IMU data for 500 rides in Gothenburg (Safe: 416, Safe Overall: 38, Bad Weather: 22, Tandem: 18, Reckless: 6). Reckless rides are the positive class we aim to detect; Safe rides are normal; the remaining categories (Safe Overall, Bad Weather, Tandem) are counterfactual — they lie near the decision boundary and should not be flagged as anomalies.
  • GBG500_labels.csv: GBG500 labels (used only for evaluation).
  • SRV3490.parquet: Time index and IMU data for 3490 rides in 8 cities across Europe. 698 rides contain topple (fall) events with user-confirmed crashes; the remaining rides have no known or detected topple events. For confirmed-crash rides, crash_ride_id is null; for baseline rides, it points to the associated confirmed-crash ride ID (4 baselines per crash).

Notebooks

• Location: notebooks/
  • data_info.ipynb: Dataset overview — row counts, total hours, ride duration statistics, and missing-value analysis for GBG500 and SRV3490.
  • GBG500_spectral.ipynb: DTAPS feature extraction and anomaly detection for GBG500.
  • GBG500_spectral_pca.ipynb: DTAPS + PCA dimensionality reduction and anomaly detection for GBG500.
  • GBG500_spectral_baselines.ipynb: Simple baseline anomaly detection methods for GBG500.
  • GBG500_spectral_ae.ipynb: DTAPS + AE dimensionality reduction and anomaly detection for GBG500.
  • GBG500_spectral_svdd.ipynb: DTAPS + Deep SVDD (Ruff et al., ICML 2018) and anomaly detection for GBG500.
  • GBG500_ts2vec.ipynb: TS2Vec (Yue et al., AAAI 2022) and anomaly detection for GBG500.
  • GBG500_minirocket.ipynb: MiniRocket (Dempster et al., KDD 2021) and anomaly detection for GBG500.
  • GBG500_latent_viz.ipynb: Visualizes the GBG500 dataset in the AE latent space using out-of-fold encoder representations (proposed method, median AP seed), plotting the two latent dimensions with the highest variance colored by ride category.
  • SRV3490_or_analysis.ipynb: DTAPS + AE + LOF odds ratio analysis on SRV3490 via conditional logistic regression.

Note: Proposed method for GBG500: DTAPS + AE with LOF.

Running notebooks (batch)

Use run_notebooks.py to execute notebooks via papermill multiple times with different seeds; outputs are written to the default output/ directory with _seed_<N> appended.

# GBG500 spectral + PCA (30 runs each) - parallelism defaults to --n_jobs=-1
python run_notebooks.py notebooks/GBG500_spectral{,_pca}.ipynb --n_runs=30

# GBG500 baselines (run once)
python run_notebooks.py notebooks/GBG500_spectral_baselines.ipynb --n_runs=1

# GBG500 AE + Deep SVDD + MiniRocket baselines (30 runs)
python run_notebooks.py notebooks/GBG500_spectral_{ae,svdd}.ipynb notebooks/GBG500_minirocket.ipynb --n_runs=30 --n_jobs=8

# GBG500 TS2Vec baseline (30 runs)
python run_notebooks.py notebooks/GBG500_ts2vec.ipynb --n_runs=30 --n_jobs=1

# SRV3490 analysis (30 runs)
python run_notebooks.py notebooks/SRV3490_or_analysis.ipynb --n_runs=30 --n_jobs=2

Generate statistics from output

After running notebooks, you can extract all scrapbook glue values from the executed notebooks in output/ and aggregate summary statistics to a single CSV (per glue_key: count, mean, std, min, max and quartiles; std may be empty when count=1):

python generate_stats.py

This writes output/statistics.csv.

Results

Pre-computed aggregated statistics are available in results/statistics.csv. GBG500 notebooks use 5-fold stratified CV; SRV3490 uses 5-fold group CV (splitting by crash group to preserve the matched case-control design). Labels are held out and used only for the metrics below. Results are aggregated over 30 random seeds.

GBG500 — Average Precision (30 seeds, 5-fold stratified CV)

Simple baselines

Features	Dim. Reduction	Scorer	AP (mean ± std)
Random	—	—	0.012
DTAPS	—	Mean spectral power	0.667

Ablation study

Features	Dim. Reduction	Scorer	AP (mean ± std)
DTAPS	AE	IF	0.612 ± 0.053
DTAPS	AE	LOF	0.771 ± 0.047
DTAPS	AE	OC-SVM	0.382 ± 0.103
DTAPS	AE	Reconstruction error	0.620 ± 0.042
DTAPS	PCA	IF	0.570 ± 0.064
DTAPS	PCA	LOF	0.695 ± 0.017
DTAPS	PCA	OC-SVM	0.391 ± 0.091
DTAPS	—	IF	0.543 ± 0.035
DTAPS	—	LOF	0.683 ± 0.026
DTAPS	—	OC-SVM	0.364 ± 0.093
DTAPS	Deep SVDD	Deep SVDD	0.565 ± 0.093

Time-domain representations

Features	Dim. Reduction	Scorer	AP (mean ± std)
TS2Vec	—	IF	0.167 ± 0.087
TS2Vec	—	LOF	0.237 ± 0.111
TS2Vec	—	OC-SVM	0.220 ± 0.106
MiniRocket	—	IF	0.276 ± 0.043
MiniRocket	—	LOF	0.477 ± 0.054
MiniRocket	—	OC-SVM	0.518 ± 0.046
MiniRocket	PCA	IF	0.342 ± 0.045
MiniRocket	PCA	LOF	0.487 ± 0.032
MiniRocket	PCA	OC-SVM	0.536 ± 0.022

GBG500 — Latent space (DTAPS + AE, out-of-fold, median AP seed)

Out-of-fold autoencoder latent representations of the GBG500 rides (each ride encoded by the fold in which it was held out), projected onto the two latent dimensions with the highest variance.

SRV3490 — Odds Ratio Analysis (30 seeds, 5-fold group CV, DTAPS + AE + LOF)

Conditional logistic regression on confirmed-crash vs. matched-baseline rides (4 baselines per crash, all rides truncated 10 s before the topple event):

Analysis	OR median [IQR]	p-value (median)	p < 0.001
Continuous (per 1-SD)	2.44 [2.38, 2.53]	2.1 × 10⁻⁶⁰	100%
Top 5 %	4.70 [4.24, 5.16]	2.2 × 10⁻²⁰	100%

How to cite

This work has been accepted at ECML PKDD 2026 (Naples, Italy, 7–11 September 2026). A suggested citation will be provided as soon as the proceedings are published.