Laboratoire de Conception Fabrication Commande

Automatic identification of accident causes from textual reports is a major challenge for improving risk management in industrial systems. This article proposes an approach based on automatic natural language processing (ANLP) and machine learning to predict accident causes from textual descriptions. A database of real accident reports was used, containing textual summaries (SummaryCase) and their associated causes (Cause). After text pre-processing including cleaning, tokenization and vectorization (TF-IDF and Bag of Words), several supervised learning models were tested, including random forests (RF), logistic regression (LR), support vector machines (SVM), Decision tree (DT), Naive Bayes (NB) as well as XGBoost (Extreme Gradient Boosting). The performance of the model was examined in terms of the accuracy, precision, recall and F1-score. As a result those work demonstrate that an XGBoost combining with SMOTE (Synthetic Minority Over-Sampling Technique) are more effective when compared to traditional models especially in the situation of long and complex text description. The results of this study have proved the effectiveness of the application of machine learning techniques to the proactive analysis of accident causes and are expected to contribute to the adaptation of these approaches into the realm of industrial risk management systems.

Risk management (RM) is crucial in product development processes in the engineering domain since mitigating risks ensures satisfactory product performance. Existing RM approaches in engineering require numerical inputs converted from textual data, which are manually collected from risk reports and converted into numerical inputs by human risk experts via their experiences. The manual process of doing so is laborious. Since natural language processing (NLP) techniques can process textual data in a similar way that humans comprehend textual data, NLP techniques can potentially automate the process of obtaining numerical inputs from textual data. Therefore, we experimented with multiple NLP techniques to automate the process of collecting numerical data from risk reports that serve as the inputs to RM approaches. Our method performed risk identification and analysis, during which textual data from risk reports were converted into numerical data via NLP techniques like generative pre-trained transformers (GPT) and bidirectional encoder representations from transformers (BERT). Parts of risk identification and analysis were successfully performed, but some results are not accurate due to NLP techniques not being able to understand causal relationships.

International audience

Dealing with the complexity of smart industries, and increasing their productivity has always been the research focus of researchers aiming at meeting the requirements of Production, Quality and Maintenance, as three fundamental pillars of industries. Given the direct link of this triplet, studies acting on these three functions are, generally, divided into two main types of optimization: Partial integration which considers jointly Production-Quality or Production-Maintenance, or Maintenance-Quality for optimizing productivity, and Full integration that consider them all. To understand the publications trends <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">us</sup> ing the joint consideration of Production, Quality and Maintenance of the existing studies and to provide a holistic view, this paper performs a Systematic Literature Network Analysis SLNA that undertakes a bibliometric analysis from 1977 by the mean of: Performance Analysis and Science mapping using VosViewer. The main contribution of this paper is the classification of the investigated papers into thematics and the study of the relationships among the most frequent terms in theses papers. Analyses of 2306 original publications demonstrated that the joint consideration of Production and Quality is the oldest and the Joint consideration of Production-Maintenance is the most widely used, whereas the full consideration of Production, Quality and Maintenance is recent and related to advanced smart tools.

Abstract Several studies have pointed out that Human factors are one of the main significant challenges in the development of the Factory of the Future (FoF). We speak now of Industry 5.0: industry 4.0 with the emphasis of ecology and human aspects. In this context the aim of this paper is to present the new issues for worker’s health and safety which have to be taken into account as soon as possible in the workplace design process and its use. We are speaking of “integrated prevention” which has to be applied from the system requirements to the functional design and the architectural design. Deployment of new technologies and organizations in the frame of FoF introduces hazardous events, problems of health and safety which have to be taken into account in the different life situations. The identification of dangerous situations therefore requires investigation of identification of characteristic design parameters related to risk qualification for the operator tasks. We start from the basic idea that accident is the interaction between a dangerous device and the worker, that can be directly linked to the generalized energy in a defined space. Each energy is characterized by parameters related to hazards for which the severity is defined by the standards and regulations. The methodology proposed combines a generic design approach based on a systems engineering approach to identify the uses that operators will make to ensure production requirements, and an energy approach for identifying potential hazards.KeywordsHealth and safetyHazard identificationFactory of the futureSystems engineeringIntegrated prevention