This repository showcases cutting-edge R&D projects in Digital Signal Processing (DSP) and Machine/Deep Learning (ML/DL), focusing on noise reduction and custom signal transforms (e.g., modified CWT/STFT) for detecting anomalies in acoustic and vibration signals. Drawing on expertise in optimization, calculus, and linear algebra, these developments enable real-time feature extraction for applications like environmental sound analysis and classification. They also support industrial sound monitoring, with potential extensions to rotating machinery failure detection (bearings, motors, rotors), HVAC fault detection and diagnosis (pumps, compressors, valves), and drilling telemetry monitoring.
Automatic Environmental Sound Classification (ESC) leverages the ESC-50 dataset (and its ESC-10 subset) developed by Karol Piczak, as detailed in his paper titled:
"ESC: Dataset for Environmental Sound Classification." by Karol J. Piczak. 2015. In Proceedings of the 23rd ACM international conference on Multimedia (MM '15). Association for Computing Machinery, New York, NY, USA, 1015–1018. https://doi.org/10.1145/2733373.2806390"
This dataset serves as a foundation for research in audio event recognition.
Advancements in ESC Using Multi-Feature CNNs:
We propose a two-stages classification approach with Multi-feature Convolutional Neural Networks (CNNs), achieving near-perfect accuracy rates, specifically reaching up to 99%. This high accuracy is attributed to innovative pre-processing techniques that combine mel-spectrograms with complex wavelet transforms (CWT).
Resolution of Remaining Classification Challenges:
A notable challenge in ESC-10 sound classification was the confusion between "sea waves" and "rain" sounds. This issue was addressed by developing an original transformation of the complex CWT, termed aT-CWT. This transformation replaces the phase component of the CWT for stationary and pseudo-stationary sounds with a Gaussian distribution, enhancing the model's ability to differentiate between similar sounding environmental events.
By integrating the aT-CWT transformation, the multi-feature CNN model has now achieved 100% accuracy in classifying environmental sounds from the ESC-10 dataset.
|
|
|
|---|---|---|
aT-CWT transform: "seawave" |
aT-CWT transform: "rain" |
Confusion Matrix with aT-CWT |
- Rotating machinery Failure Detection: bearings, motors,rotors.
- Oil & Gas drilling for real-time telemetry.
A scalable pipeline for predictive maintenance in rotating machinery, using time-domain features (RMS, kurtosis) and optimized Kalman denoising on NASA PRONOSTIA datasets to detect faults ~2 minutes early (frame 600). Extends to:
LSTM/Prophet for anomaly forecasting and
CWT/aT-CWT for phase-coherent scalograms, achieving >95% accuracy on high-harmonic preservation (236-980 Hz BPFO).
Applications
- Rotating machinery Failure Detection: bearings, motors,rotors.
- HVAC Fault detection and diagnosis (FDD): pumps, compressors, valves.
-
This project develops an automatic unsupervised classification model to diagnose valve faults in industrial machinery using acoustic signals from an 8-microphone circular array, leveraging the MIMII dataset (CC BY-SA 4.0, Hitachi, Ltd., https://zenodo.org/records/3384388).
We introduced a novel ACSTFT transform, achieving an impressive ROC AUC of 0.99 on the +6dB valve dataset, and trained a CNN-Autoencoder for robust anomaly detection.
Unlike standard MIMII challenge approaches that classify noisy signals directly, we prioritize denoising using MVDR beamforming combined with a custom Generalized Sidelobe Canceler (GSC), transforming the array into a noise-robust “sensor.”
Focused exclusively on valves, this model enhances fault detection in challenging industrial environments, offering a practical solution for machinery monitoring. Explore the code, ROC curves, and spectrograms showcasing our results.
Applications
|
|
|
|
|
|---|---|---|---|---|
Novel ACSTFT Transform |
ROC-AUC= 0.99 |
Reconstruction error (MSE) |
MVDR beamforming |
Denoised Valve Sound Signals |
In this project, we are developing effective methods for classifying mitochondrial genomes (DNA sequences) using Digital Signal Processing (DSP), Machine Learning (ML), and Deep Learning (DL). This research is ongoing, and we plan to publish our results regularly. As a starting point, we analyzed the paper titled:
"ML-DSP: Machine Learning with Digital Signal Processing for ultrafast, accurate, and scalable genome classification at all taxonomic levels" by Gurjit S. Randhawa , Kathleen A. Hill and Lila Kari. https://doi.org/10.1186/s12864-019-5571-y
The alignment-free DNA sequence classification approach: ML-DSP, proposed by Gurjit S. Randhawa has proven to be very effective.
By introducing a simple alignment technique alongside short Fast Fourier Transforms (FFTs), termed ML-FFT + SoftAlign, we have surpassed the performance of ML-DSP, particularly with challenging datasets such as those from Fungi and Insects.
-
This repository is a handpicked anthology of intricate mathematical and logical problems designed to probe the limits of AI reasoning capabilities. From differential equations and series convergence to probabilistic asymptotics and spectral inversions, each challenge demands multi-step deduction, symbolic manipulation, and creative insight—perfect for benchmarking LLMs like Grok or GPT.
Whether you're an AI researcher honing models or a math enthusiast seeking tough puzzles, DeltaNotes will offer 50+ problems with hints, solutions, and code snippets for verification. Under construction, problems are added on a weekly basis.
Explore, test, and contribute: Turn theory into AI stress-tests!
This section is a portfolio of Machine Learning projects with Python and various visualization and analysis tools. Most of these projects were carried out within the framework of IBM certifications. They are presented with Jupyter Notebooks.
Some projects have been improved by incorporating more in-depth data analysis, better graphs, advanced ML techniques.
In this project, we predict if the Falcon 9 first stage will land successfully. Project includes: SpaceX data collection, Data Wrangling, Webscraping, EDA with SQL Queries & Data visualization, SpaceX Launch Records Dashboard, Launch Sites Locations Analysis with Folium, Machine Learning classification with optimization of hyperparameters and selection of best model: KNN, Decision Tree, SVM, Logistic Regression. 
A widerange of small projects with various ML techniques, prediction, supervised and unsupervised classification: Linear Regression, Polynomial Regression, Non-Linear Regression, Recommandation Systems, KNN, Customer Segmentation with K-Means, Hierarchical Clustering, Density-Based Clustering, Logistic Regression. 
The project consists of finding the best model for predicting home prices in King County, USA in Washington State, based on a dataset of homes sold between May 2014 and May 2015. Prediction accuracy was improved by implementing a spline regression model.One Jupyter Notebook includes interactive Folium maps (interactive maps will not display on Github).
Loan Status Prediction using Supervised Classification Algorithms: KNN, Decision Tree, SVM, Logistic Regression.
Old dataset on housing prices derived from the U.S. Census Service to present insights based on our experience in Statistics. Median value of houses bounded by the Charles river, of owner-occupied units built before 1940, relationship between Nitric oxide concentrations and the proportion of non-retail business acres per town, impact of weighted distance to the five Boston employment centres on the median value of owner-occupied homes. 
Dataset: car dataset including various makes, specifications and prices.
After cleaning the dataset, running statistics, identifying the most relevant variables, we develop several models that will predict the price of a car using a set of features/variables.
|
|
|
|---|---|---|
Word Cloud |
Folium with markers |
Choropleth |
|
|
|
|
|---|---|---|---|
"Figure 8" toroid |
Gyroid |
Truncated cuboctahedron |
Helicoid-Catenoid |
Crafting DSP & ML/DL innovations that uncover hidden signals in noise, grounded in rigorous numerical analysis and real-world impact.
Audio & Signal Processing scientist blending mathematical precision, custom transform design, and ML-driven anomaly detection to transform chaotic data into actionable insights for industrial and environmental applications.
What drives me most in DSP/ML is the thrill of R&D: Distilling cutting-edge theory (NR gains, STFT, wavelets, synchrosqueezing) into tools that solve tough problems, like detecting valve faults or drilling vibrations before they escalate.
⇒ Signal Analysis Rigor - Advanced time-frequency methods (CWT, STFT, Mel-Spectrograms) for non-stationary phenomena
⇒ Noise Reduction & Enhancement - Beamforming, VAD, and adaptive filters achieving >20 dB SNR gains in multi-sensor setups
⇒ Mathematical Foundations - Expertise in calculus, linear algebra, and optimization for feature extraction and model stability
⇒ ML/DL Innovation - CNN, CNN autoencoders and genetic algorithms for unsupervised anomaly detection in acoustics/vibrations
⇒ Industrial Applications - 25+ years in electroacoustics, computational mathematics, DSP, in Telecom & Consumer Electronics industries, Intelligent Building and Safety Industry. | PhD in Mech. Engineering: Numerical Analysis (BEM/FEM) and Optimization.
I develop open-source pipelines that bridge numerical rigor with practical engineering, specializing in custom transforms for anomaly detection in noisy environments. From environmental sound classification to industrial valve diagnostics, my work advances real-time ML for telemetry and beyond.
▶ Mission: Empowering engineers with math-backed DSP tools that turn signal chaos into predictive power—building the future of acoustic, vibrations, signals intelligence, one transform at a time.
🔭 I’m currently working on advanced projects in ML & DL
👯 I’m looking to collaborate on Digital Signal Processing, Machine Learning, Deep Learning
📫 How to reach me: stephane.dedieu@bloo-audio.com