Publications and conference contributions

• Human impact in complex classification of steel coils  (CIST2023, 11st World Conference on Information Systems and Technologies 4 – 6 April 2023 – Pisa, Italy )
• This work explores different artificial intelligence based alternatives to create automatic classification system based on salience maps of coil surface when heavy unbalanced datasets are considered and where the labels have been assigned by human operators, considering different complex rules. After testing the possibilities of classifier setup process, additional effort was spend create synthetic features based on the characteristics of the salience maps and such features have been used to verify the need for additional check of scores coming from the human operators. Although it is a preliminary result, it provide evidences of the significant impact that confusing scores can have on the classifier final performance.

• Use of robust deep learning methods for the automatic quality assessment of steel products (METEC & 6th ESTAD (European Steel Technology and Application Days, 12 – 16 June 2023 – Düsseldorf, Germany )

The use of machine learning (ML) is a change of the classical programming paradigm. Instead of writing code to get desired output from a given input, you provide the input with the desired output (the training data) to the ML system. The system “learns” the desired output from the samples in the training data. Translated to the problem of automatic product release this means that the system gets quality data and associated decisions and automatically learns the rule-base, that was previously written by hand. Learning decisions from data instead of writing complex rules means that the training data becomes the source code and the essential part of the developed solution. A deep learning (DL) solution applicable in the industrial practice cannot just consider the type of DL technology exploited, but also must incorporate the management of the training data. In this work to support quality decision making process a solution is proposed by combining deep learning technology with sophisticated management of underlying training data enabeling the optimal use of all available data sources and simplify the configurability and maintainability of previous decision support systems. The proposed system which is a RFCS research project (DeepQuality), consists of concepts realizing a human-centered lifecycle for the robust industrial application of DL quality models.

• Smart Workflows for Advanced Quality Assessment in Steel Industry: Benefits of I5. 0  (Advances in Manufacturing IV. MANUFACTURING 2024. Lecture Notes in Mechanical Engineering. Springer, Cham. https://doi.org/10.1007/978-3-031-56474-1_5 )

  This paper aims to present the smart workflow proposal in Industry 5.0 context. The adopted example will be the steel industry manufacturing of high-quality coils for the automotive industry, as it is extremely competitive. The proposal will be a self-maintained framework for continuous improvement of artificial intelligence quality models, being enriched by the qualified opinion of operators. Not just the framework but also the general architecture of models will be presented, as well as the way interaction at different frequencies is considered and aside operations related to performance preservation.