Engine Order Analysis

Extracts engine order magnitudes and positional deviations from audio recordings via pitch-adaptive spectral analysis.

This tool implements the spectral analysis pipeline described in:

Doerfler & Wyse, "Analysis-Driven Procedural Generation of an Engine Sound Dataset with Embedded Control Annotations", EUSIPCO 2026. [arXiv]

Overview

Engine sounds are characterised by their engine orders -- sinusoidal components at integer and half-integer multiples of the crankshaft rotation frequency. This package extracts per-order features from multi-channel engine recordings (stereo audio + RPM/torque control channels) through three processing stages:

1. Angle-domain resampling: Time-warps each frame so the fundamental frequency is held constant, using cubic-spline interpolation.
2. Frequency-aligned FFT analysis: Scales the Blackman analysis window to span P = 20 fundamental periods and zero-pads by factor z = 4, locking each order to a fixed bin index regardless of RPM.
3. Centroid-based order estimation: Estimates true order positions via spectral centroid within tapered rectangular regions, with magnitude recovery through parabolic interpolation.

The output is a JSON file containing per-order magnitudes A_h and deviations delta_h indexed by (RPM, torque) operating point.

Installation

pip install -e .

Requirements

• Python >= 3.9
• numpy, scipy, librosa, matplotlib

Usage

Command Line

# Analyse a single file
engine-order-analysis /path/to/recording.wav

# Analyse a folder of WAV files
engine-order-analysis /path/to/folder/

# Specify output path and generate plots
engine-order-analysis /path/to/recording.wav -o results.json --plot

# Custom sample rate and chunk size
engine-order-analysis /path/to/folder/ --sample-rate 16000 --chunk-size 65536

Python API

from engine_order_analysis import run_analysis

# Run on a single file
output_path = run_analysis(
    input_path="/path/to/recording.wav",
    output_path="results.json",
    plot=True,
)

# Run on a folder
output_path = run_analysis(input_path="/path/to/folder/")

Input Format

Each WAV file must contain 4 channels at 48 kHz (resampled to 16 kHz for analysis):

Channel	Content
1-2	Stereo engine audio
3	RPM signal (value x 0.0001)
4	Torque signal (value x 0.001)

This is the format used by the Procedural Engine Sounds Dataset.

Output Format

The JSON output maps "RPM,torque" keys to per-order analysis results:

{
  "3200,150": {
    "magnitudes": [[0.023, 0.015, ...], ...],
    "deviations": [[0.001, -0.003, ...], ...]
  }
}

• magnitudes: List of frames, each containing 128 values (orders 0.5 to 64.0).
• deviations: Order-normalised offset between estimated and nominal order positions. Positive = sharp, negative = flat.

With --plot, four heatmap PNGs are generated: magnitudes and deviations each plotted against RPM and torque.

Citation

@inproceedings{doerfler2026eusipco,
  title={Analysis-Driven Procedural Generation of an Engine Sound Dataset
         with Embedded Control Annotations},
  author={Doerfler, Robin and Wyse, Lonce},
  booktitle={Proc. 34th European Signal Processing Conference (EUSIPCO)},
  year={2026},
  address={Bruges, Belgium}
}

License

CC BY-NC 4.0