Exploring
projects in Artificial Intelligence and Mechanics
Exploration of state-of-the-art AI models and design of mechanical systems.
Artificial Intelligence Projects
Creation of a generative artificial intelligence architecture applied to ultrasonic acoustic waves. Tech news, with application of llama 3.1 via Groq API to summarize the latest publications.
Non-Destructive Testing (NDT) Innovation
AI at the
service of the industry
High-fidelity amplitude reconstruction from binary ultrasonic Full Matrix Capture for non-destructive testing.
Research Context
The acquisition of Full Matrix Capture (FMC) with ultrasonic probes generates a colossal amount of information, leading to significant storage issues. To address this, the idea of binarizing these FMC matrices has developed in the literature. An embedded binary acquisition system was thus developed, inevitably implying a significant loss of information on the acquired physical waves. The goal of this research was to create a lightweight AI architecture capable of reconstructing the lost information to be integrated into the new embedded acquisition system.
-89%
Fewer parameters. The architecture was radically compressed, from 166M (U-Net) to 18M parameters, making it eligible for edge-sensor integration.
16x
Faster. Drastic reduction in FLOPs enabling real-time inference throughput, without any loss of structural resolution on defects.
From binary to real amplitude
The model analyzes binarized signals (1-bit) in matrix form, and manages to reconstruct the corresponding FMC. Swipe to compare.
Architecture Optimization
On portable acquisition systems such as the QuickScan iX PA+, AI inference is limited by hardware constraints. The solution lies in MobileNets with Depthwise Separable convolutions.
Lightening techniques
Linear BottlenecksRemoval of ReLU activations on dimensional projection layers, preserving the manifoldity of low-level information without disruption.
Inverted residualsCore block of the network, optimized for time efficiency.
Custom MobileViT-V1-V3-FPN-PixelShuffle
An unprecedented fusion between the robust local extraction of separable convolutions and the global spatial analysis of Vision Transformers. Final image restoration is performed via a Feature Pyramid Network (FPN) followed by a native PixelShuffle algorithm. Zero additional computational parameters.
Trained on steel
The developed architecture was trained exclusively on steel. Its dataset comprises 50% data simulated by advanced finite elements (Pogo FEA) and 50% data measured via a QuickScan iX PA+.
Hybrid database (600 FMC)
The model was trained on a set of 600 FMCs, with 80% used for training and 20% for validation. A unique test dataset was acquired to evaluate the model's performance. The results show excellent generalization to complex geometries as well as to materials never seen during training (copper, titanium alloy).
Multi-mode and multi-frequency adaptation
The dataset includes measurements in contact and with a wedge, allowing the acquisition of all reconstruction modes of the Total Focusing Method (TFM), including L-L, T-T, TT-T, and TT-TT modes. The model also generalizes its inference on transducers at 2.25, 5, and 7.5 MHz, for industrial use independent of the probe used.
End of architecture report
Latest
in AI
Day-by-day tracking of artificial intelligence news. Integration of a Groq API with the llama-3.1-8b model for article summarization.
Live feed · refreshed twice a day
HAND
Initializing the 3D viewer...
Scroll to discover
Articulated Robotic Hand
Mechanical engineering project serving life sciences, conducted by a team of 7 people, covering kinematics, mechanical design, and prototyping, aimed at a medical engineering application integrating its specific technical, ethical, and regulatory constraints.
Structure
4 fingers, 3 phalanges each, opposable thumb, and structural palm.
Manufacturing
Optimized for 3D printing with precise assembly tolerances.
Interaction
Explore the model above: hover to explode, drag to rotate.
Mechanical Architecture
Theoretical modeling and design of the prototype for 3D printing were performed on Onshape. Each finger is modeled using revolute joints between the different phalanges as well as with the body of the hand. The external interface of the joints is tightly fitted, while the internal part has a looser fit to allow movement. A wire can be introduced into each finger, except for the thumb, to actuate flexion and make the hand mobile. The thumb is mounted on a ball joint, allowing several degrees of freedom. This device acts as a first version (V1) of a prosthetic prototype.
Design Process
Initial experimental validations of the prototype consisted of verifying its ability to withstand several kilograms under transverse stress applied to the fingers. Gripping performance was also evaluated; the hand had to be capable of grasping and supporting a cylindrical object of about 1 kg, such as a water bottle.
Interactive 3D Visualization
The 3D model allows you to explore the assembly in detail:
Exploded view on hover
Hover to see the progressive explosion of parts
Free rotation
Click and drag to observe from all angles
Potential Applications:
Prosthetics:
Basis for the development of functional prostheses
Robotics:
Integration into advanced robotic systems
Education:
Educational support for learning mechanics