Characterization of padel ball/racket impact using artificial intelligence methods

Junia ISEN – Master 1 project under the supervision of Arthur PATÉ

25th November - 20th December 2024 ➡️ Research Part

17th March - 25th April 2025 ➡️ Application Part

👥 Authors:

ANTONIUK Pavlo, DAMERY Vincent, LAMBERT Edouard, MANY Hugo, OMS Henri, ZAKI Ilias

📑 Table of Contents

1. 🎯 Project Objective
2. ✨ Features
3. 🤖 Machine Learning Models
4. 📊 Data
5. 🛠️ Tools and Functions
6. 📁 Project Structure
7. 📊 Results and Evaluation

🎯Project Objective

The goal of this project is to predict the impact position of the ball on the padel racket, the type of racket used, and the racket’s age, based sound or vibrations.

✨Features

Energy

Energy per frequency band is extracted using the FFT and segmented using customizable bandwidths. This highlights how much energy is distributed across specific frequency regions.

Envelope

The spectral envelope of the audio signal is calculated using the Hilbert transform. It captures how the amplitude change over time.

MFCC

Mel-Frequency Cepstral Coefficients are extracted to represent the spectral content in a perceptually relevant way. Averaged MFCCs are used as features for classification tasks.

Peaks

Using FFT, the most prominent frequency peaks (position and magnitude) are extracted from the audio signal. These peaks shows the dominant acoustic components of each impact.

Attack Time

This feature represents the time it takes for the sound to rise from silence to its peak amplitude — a key characteristic in assessing impact sharpness and racket responsiveness.

🤖Machine Learning Models

KNN

KNN is a simple model. It looks at the closest examples and choose the most common label. We used it as a baseline.

RTF

Random Forest builds a bunch of decision trees and combines their results. it's a good model for noise and avoids overfitting.

SVM

SVM tries to find the best boundary to separate classes. It's a good model when the data has many features.

XGBoost

XGBoost builds trees one at a time. Each new tree focuses on fixing the errors from the last one. It’s fast and mostly gives strong results.

📊Data

The dataset consists of audio and vibration recordings of padel ball impacts on rackets. The data is categorized based on the following attributes :

• Racket Types:
  Four types of rackets are used, each differing in size, shape, number of holes, and other characteristics :

  • B : Blue
  • O : Orange
  • R : Red
  • V : Green

• Impact Zones:
  The ball impacts are recorded at three distinct zones of impact on the racket :

  • S : Smash
  • C : Center
  • V : Volley

• Racket Generations:
  The rackets are categorized into three generations based on their time of use :

  • P1 : New rackets (Fixed condition)
  • P2 : Moderately used rackets (Free condition)
  • P3 : Heavily used rackets (Free condition)

For each combination of racket type, impact zone, and condition, 9 impacts are recorded.

This structured dataset provides a comprehensive basis for analyzing the relationship between sound/vibration features and racket characteristics.

Sound

We use the sound data in temporal and frequency domains to extract features for the different models. The Fast Fourier Transform (FFT) was used to compute the spectrum.

Each audio file is named using the following format : generation_racketType_numOfRecord_zoneOfImpact_numOfTry.wav

Example : P1_RB_1_C_1.wav

Vibration

We use the frequential data in frequency domain to extract features for the different models, using Fast Fourier Transform (FFT) to compute the spectrum.

The vibration datas are stored in the All_Data_combined.csv file. Below is an overview of the columns in this file :

Column Name	Description
Raw Signal Ch0	Temporal signal data captured from the vibration sensor.
Spectrum	Frequency domain representation of the vibration signal.
freqs	Frequency values corresponding to the spectrum data.
File Name	Name of the file containing the recorded vibration data. generation_racketType_numOfRecord_zoneOfImpact_numOfTry.csv
Position	Impact zone on the racket where the ball made contact (C, S, V).
Racket Type	Color-coded identifier for the type of racket used (B, O, R, V).
Age	Generation or usage condition of the racket (P1, P2, P3).

🛠️Tools and Functions

Signal Processing Functions

• readWavFolder(folderPath)

  Reads all .wav files from a folder and returns their sample rates, data arrays, and filenames.

• spectrumFromWav(wavFile)

  Computes the FFT magnitude spectrum of the first audio channel of a WAV file.

• spectrumFromSignal(signal, sample_rate)

  Computes the FFT magnitude spectrum and filters it between 150 Hz and 1000 Hz.

• energy_per_frequency_band_from_spectrum(spectrum, freqs, band_width)

  Divides the frequency spectrum into bands and computes the energy (sum of squares) in each.

• envelope_from_signal(signal)

  Computes the amplitude envelope of a 1D signal using the Hilbert transform.

• extractPeakFromSignal(signal, smoothing=1, num_peaks=None)

  Extracts the most prominent peaks from a 1D signal, with optional smoothing and limit on number of peaks.

• plot_spectrum_with_freq(signal, freqs, title="Spectrum Plot")

  Displays the magnitude of a frequency spectrum against its corresponding frequencies.

Vibration Data Functions

• readAllFileVibration(folderPath)

  Recursively loads all .csv vibration files from a directory and returns the folder name and the corresponding pandas DataFrame.

📁Project Structure

├── Data
├── Functions
├── SoundPart
│   ├── ModelMLAgeRacket
│   │   ├── KNN
│   │   |   ├── S_KNN_Age_P1.P2.P3_Peaks.ipynb
│   │   |   ├── S_KNN_Age_P1.P2.P3_Peaks.csv
│   │   |   ├── xxxx.ipynb
│   │   |   └── xxxx.csv
│   │   └── RTF
│   ├── ModelMLPositionRacket
│   └── ModelMLTypeRacket
├── VibrationPart
│   ├── Deprecated
│   ├── ModelMLAgeRacket
│   ├── ModelMLPositionRacket
│   └── ModelMLTypeRacket
└── Visualization

📊Results and Evaluation

The best results for each model are stored in the AllResults.xlsx file. All of them was trained and tested on sound and vibration datasets with selected features to evaluate its effectiveness.

The file contains multiple sheets : sound, vibration, both, comparisons between sound and vibration and comparison between KNN and RTF.

The table below show some example of performance for different machine learning models from this file :

Sound

Model Type	Objectives	Dataset	Features	Training Accuracy	Test Accuracy
KNN	Position	P1	Energy	99.7%	100%
KNN	Age	P1.P2.P3	Peaks	100%	77.3%
RTF	Age	P1.P2.P3	Peaks	99.8%	91.4%
RTF	Racket Type	P1.P2.P3	AttackTime	100%	67.5%
...	...	...	...	...	...

Vibration

Model Type	Objectives	Dataset	Features	Training Accuracy	Test Accuracy
KNN	Position	P1	Envelope	99.4%	100%
KNN	Racket Type	P1.P2.P3	MFCC	100%	100%
RTF	Position	P1	Energy	100%	95.5%
...	...	...	...	...	...