The AI Regio Project
AI REGIO is part of the European Commission’s I4MS initiative
(ICT Innovation for Manufacturing SMEs),
which fosters the integration of digital innovations by manufacturing SMEs in Europe in order to boost their competitiveness. I4MS is currently operating in its fourth phase.
AI REGIO is a project funded by the European Union Framework Programme for Research and Innovation Horizon 2020 under Grant Agreement n° 952003.
The AQUILA Project
The project aims at extending the AI4EU platform with an “executable” asset named AQUILA System, useful to automate Quality Inspection (QI) of industrial products, in particular electrical panels. The AQUILA System will combine computer vision deep learning techniques with automatic reasoning and Answer Set Programming (ASP), with the objective to obtain 30% higher precision in defect detection w.r.t. solutions based exclusively on computer vision
Experiment overview
Artificial Intelligence (AI) acts a key role in supporting and improving smart manufacturing and Industry 4.0 by allowing for automating different types of tasks manually performed by domain experts. Among them, the compliance assessment of a product with the relative blueprint represents a time consuming and prone to error task. An emerging approach relies on computer vision techniques to support the human operator in this process. The experiment aims to compare the design of an electronic product, represented by a CAD file, with the picture of a real artifact of the product.
Starting from the CAD file is possible to annotate each basic component that must be integrated into the electronic product by an automatic parsing of the file or a graphic user interface. The annotated CAD file is then represented as several logical facts, enabling the use of ASP and automated reasoning
A data augmentation approach is then applied to the set of pictures of each basic component to create a dataset of the electronic product on which a deep learning neural network for object detection is trained
The trained neural network is used to process the picture of a real artifact of the electronic product, recognize the components and represent them as logical facts
A logical program compares the logical representation of the CAD file and the logical representation of the real artifact to recognize if all the expected basic components in the CAD are present in the real implementation in the correct position
The output of the logical program is a compliance report illustrating if there are missing or wrong-positioned basic components
Scientific and Technological Excellence
Innovation
In the past few years, a growing number of automatic QI tools and approaches have been proposed, with the aim to optimize the quality control process through computer vision. In the field of electronic products, the innovative nature of AQUILA is that the system uses the “explicit” knowledge about the right way to assemble the product, represented by a CAD file, and the “implicit” knowledge about a real artifact, represented by the product picture. In that way, the comparison between the ideal product and the real one could be more effective
Novel concepts and approaches
The AQUILA experiment integrates two innovative approaches:
• data augmentation, useful to create a synthetic dataset starting from a small number of images of the basic components of the electronic product;
• the combination of deep learning with ASP. Deep learning is used for object detection, while ASP is used for compliance check w.r.t. the CAD file.
Currently, in scientific literature there are very few examples combining of neural approaches with symbolic ones. Thus, the AQUILA system is a state-of-the-art technology.
Implementation
The AQUILA system will be deployed as two Docker containers. The first one is devoted to AI processing, and it will integrate two state-of-the-art technologies for deep learning (Tensor Flow) and ASP (DLV). In particular, the DLV platform is recognized as one of the most suitable engines for industrial applications of ASP (see [1] The ASP System DLV: Advancements and Applications. Künstliche Intell. 32(2-3): 177-179 (2018), [2] Some DLV Applications for Knowledge Management. LPNMR 2009: 591-597). The second Docker container will contain the AQUILA web application, which offers a CAD annotation environment and an interactive interface for QI.
Challenges
Currently, the quality control of electrical panels employs a dedicated operator who takes an average time of 340 s to control each panel. Moreover, the error rate associated with the quality control of electrical panels is approximately 1 error every 250 panels. Through the introduction of the AQUILA system, we estimated to reduce the control time by about 60% and achieve an error rate of about 1 error per 1000 frames. Currently, the AQUILA system has a TRL 4, and is expected to reach a TRL 6 at the end of the experiment.
Impact
Increase the digitalisation level
The use of the AQUILA system in a manufacturing company will increase the digitalisation level of QI processes. Today QI is manually performed and represents a time-consuming task. The AQUILA system will introduce an automatic digital tool for QI, strongly reducing errors and time. In the Elettrocablaggi company, currently, QI requires on average 340 s of a human operator; we estimated to reduce this time by about 60%
Technological impacts
For Revelis, the technological impact of the AQUILA project is foreseen in terms of enhancement of its know-how in the area of Deep Learning and ASP, as well as in Big Data Analytics.
In particular, the outcome of the project will be: (i) at least one Deep Neural Network for object recognition; (ii) at least one ASP program for compliance check.
Commercial impacts
Revelis will use the results to strengthen the existing relationship with manufacturing industrial players. The rigorous validation of the platform will open the possibility of creating integrated solutions/products for the QI domain. Revelis’ primary aim is to become a leading reference company for computer vision AI solutions.
