Les avancées scientifiques

In the highly competitive automotive industry, quality control is essential to ensure the reliability of vehicles and user safety. A failure in quality control can severely jeopardize safety, result in significant financial costs, and cause substantial reputational damage to the company involved.

One of the challenges for Renault is to improve the reliability of quality control for welding seams in automotive body manufacturing. Currently, this inspection is consistently performed by a human operator due to the legal dimension related to user safety. During an industrial process, this task is resource-consuming. The key challenge is to develop an AI-based solution that reduces the number of inspections required by the operator through automated pre-validation.

Within the Confiance.ai
 
(opens new window)Research Program, Renault Group and SystemX worked jointly on the development of trustworthy AI components tackling this problem. Now part of the European Trustworthy AI Association
 
(opens new window), we want to ensure that these tools effectively validate the proposed AI-Component according to the trustworthy criteria defined by the industry (Intended Purpose).

This industrial use case, provided by Renault Group, represents the “Visual Inspection” thematic through a classification problem.

The goal is to be able to assess weld quality from a photo taken by cameras on vehicle production lines.

A weld can have two distinct states:

The main objective of the challenge is to create an AI component that will assist an operator in performing weld classification while minimizing the need for the operator to inspect the images and double-check the classifications.

For defect identification ("KO"), the system should provide the operator with relevant information on the location of the detected defect in the image, hence reducing the control task duration. 

In the context of industrial infrastructures, pipe inspection and welding are of paramount importance in ensuring safety and reliability. Ultrasonic methods have been shown to provide detailed images of welds without compromising the integrity of the inspected structure. However, the interpretation of these images can be challenging, often necessitating a costly and time-consuming analysis. The utilisation of AI vision models offers a potential solution to address these challenges. These models can analyse the images and classify them as either defective or healthy, thereby partially mitigating the issues. Furthermore, they can differentiate between defect types, providing a more comprehensive classification. It is imperative to avoid erroneous classification of defective images as healthy, given the critical safety implications. To address this challenge, effective methods have been proposed in the literature. However, the complexity of these methods and the associated computational burden of portable embedding systems, in relation to online inference time constraints, are rarely met by the hardware requirements.The present study proposes an evaluation of the performance of an AI-based classifier that has been trained and tested on multiple experimental databases. The aim is to achieve a zero-defect classification rate for healthy predictions, while maintaining a relatively high classification accuracy for healthy images. Furthermore, the study demonstrates how the developed machine learning schema can be straightforwardly implemented into an in-house embedded system, while maintaining reasonable inference performances. This capability has the potential to bring significant benefits to the industrial sector. 

The DeepGreen consortium is pleased to announce the release of version 0.6 of the Aidge framework, a modular, open-source deep learning platform dedicated to embedded systems.  

Since its introduction in early 2024, Aidge has aimed to bring high-performance, low-power artificial intelligence to a wide range of hardware, from microcontrollers to GPUs. With version 0.6, the framework takes a significant step forward in both functionality and maturity. 

Key innovations in version 0.6 

Aidge is a novel software development platform for embedded Artificial Intelligence (AI). It is designed to import or even learn Deep Neural Networks (DNN) and generate optimized code for target hardware architectures, in a completely open, transparent, and traceable manner. The purpose is to avoid dependence on opaque and non-sovereign tools or elements, ensuring competitive performance and favoring the certification of embedded Machine Learning (ML) components. In this paper, we present the preliminary analysis on the potential benefits of using this platform in light of the rising aeronautical certification standards concerning the use of ML into critical aeronautical systems, pointing possible steps toward certification, based on the artifacts that can be automatically generated by Aidge. 

3D object detection plays a crucial role in various applications such as autonomous vehicles, robotics and augmented reality. However, training 3D detectors requires a costly precise annotation, which is a hindrance to scaling annotation to large datasets. To address this challenge, we propose a weakly supervised 3D annotator that relies solely on 2D bounding box annotations from images, along with size priors. One major problem is that supervising a 3D detection model using only 2D boxes is not reliable due to ambiguities between different 3D poses and their identical 2D projection. We introduce a simple yet effective and generic solution: we build 3D proxy objects with annotations by construction and add them to the training dataset. Our method requires only size priors to adapt to new classes. To better align 2D supervision with 3D detection, our method ensures depth invariance with a novel expression of the 2D losses. Finally, to detect more challenging instances, our annotator follows an offline pseudo-labelling scheme which gradually improves its 3D pseudo-labels. Extensive experiments on the KITTI dataset demonstrate that our method not only performs on-par or above previous works on the Car category, but also achieves performance close to fully supervised methods on more challenging classes. We further demonstrate the effectiveness and robustness of our method by being the first to experiment on the more challenging nuScenes dataset. We additionally propose a setting where weak labels are obtained from a 2D detector pre-trained on MS-COCO instead of human annotations. 

Key note by F. Rastello, ML compiler experts from INRIA, given during the 5th C4ML workshop at CGO 2024. 

 In a stride towards innovation, DeepGreen proudly announces the launch of Eclipse Aidge's second release, further fortifying its position as an open-source platform for embedded AI. This milestone brings forth a plethora of enhancements aimed at empowering developers and enthusiasts alike. 

What's New? 

Embark on a journey of exploration and innovation with Eclipse Aidge's latest release. Dive deeper into the realm of embedded AI, unlock new possibilities, and elevate your projects to unprecedented heights.

 Eco-designing digital systems requires the system engineer to compute environmental impacts of its system alongside the design flow. Life Cycle Assessment (LCA) is a well proven method to compute environmental impacts of a service or product, but is hardly integrable in designer’s workflow and software environment. We propose Appa LCA, a Python software easing the integration of LCA method in design workflow by decomposing LCA in two steps: building parametric impact models, and executing them. 

This paper studies the exact implementation of the ACAS-Xu ML models (designed using Machine Learning
technique) on several hardware platforms while ensuring some properties: ML model full semantics description, memory footprint optimisation, integer representation, formal verifiability. Certification aspects are also addressed using the EUROCAE/SAE joint group WG-114/G-34 current draft of the future standard ED-324/ARP6983 for embedding ML technology in aeronautical systems.

The rapidly evolving field of computer vision has witnessed a paradigm shift with the introduction of Transformerbased architectures, particularly Vision Transformers (ViTs). As these models expand in complexity, ensuring their efficient deployment on resource-limited devices becomes crucial. This paper proposes a solution for the model compression problem, emphasizing quantization, and highlights a notable gap in current methodologies: their need to consider outliers in the quantization process. We propose a distillation-guided quantization approach for ViTs, leveraging the Centered Kernel Alignment (CKA similarity score. Empirical experiments are carried out on the DeiT architecture using the ImageNet dataset, with our CKA approach demonstrating promising results in retaining model intricacies during compression.