Constantin-Bala Zamfirescu

The growing complexity of industrial assembly processes calls for innovative, user-centered solutions, and in recent years, the adoption of assistive systems to support manual assembly tasks has attracted increasing attention. This work investigates the use of Cognitive Control Units, powered by the Soar cognitive architecture, as assistive systems in smart manufacturing. The Cognitive Control Unit has been integrated into an assembly station that can adapt itself and its processes to the user currently utilizing the station. To evaluate the effectiveness of the station, an experimental study was conducted. The evaluation of the station was carried out against a previous experiment in which no guidance was provided. The study reveals both shorter mean completion times and higher-quality assembly performance. Furthermore, participants indicated that the use of the station had a positive influence on their emotional state.

The use of Autonomous Mobile Robots in various industries is on the rise. Accurate estimation of payload delivery times poses a significant challenge for efficient planning across different sectors. This paper addresses this challenge by evaluating the effectiveness of four time forecasting methods: Linear Regression, Random Forest Regression, Transformer and Bidirectional Long Short-Term Memory networks. These models are compared to a previous approach that uses Long Short-Term Memory neural networks. Among the evaluated models, Bidirectional Long Short-Term Memory and Transformers were the only methods able to outperform the Long Short-Term Memory implementation, achieving a Mean Average Error of 1.26 seconds and 1.27 seconds, respectively. In particular, the newly proposed Bidirectional Long Short-Term Memory model could train twice as fast as the Long Short-Term Memory network, despite the increased complexity generated by the addition of the backward layer. To our knowledge, this is the first time BiLSTM and Transformers are evaluated with the aim to identify the best approach for forecasting the time left to reach a destination in a dynamic industrial environment, where the presence of numerous potential obstacles adds to the complexity of the task.

Industrial processes often rely on high-temperature heat, traditionally generated through the combustion of fossil fuels. However, a significant shift towards renewable and sustainable heat sources is underway, supported by environmental policies and actions such as the European Green Deal. These renewable energy systems are complex and characterized by a high degree of interdependencies between various parameters. Optimizing and orchestrating these processes for efficient heat delivery requires careful consideration of factors such as temperature levels, flow rates, and energy demands. Traditional methods often struggle to handle the complexity of these systems, hindering efforts to maximize efficiency and minimize energy waste. This paper addresses these challenges by proposing a modular digital twin framework tailored for high-temperature heat upgrade systems. By integrating with the physical heat upgrade system, the digital twin can create a dynamic and continuously updated representation of its behavior, while also providing additional advantages such as improved process simulation, predictive capabilities, enhanced design, and system integration. Using the SUSHEAT project as a case study, this work advances digital twin methodologies by introducing an architecture applicable in the early product lifecycle phases, addressing a gap in current research.

Autonomous mobile robots (AMRs) are gaining popularity in various applications such as logistics, manufacturing, and healthcare. One of the key challenges in deploying AMR is estimating their travel time accurately, which is crucial for efficient operation and planning. In this article, we propose a novel approach for estimating travel time for AMR using Long Short-Term Memory (LSTM) networks. Our approach involves training the network using synthetic data generated in a simulation environment using a digital twin of the AMR, which is a virtual representation of the physical robot. The results show that the proposed solution improves the travel time estimation when compared to a baseline, traditional mathematical model. While the baseline method has an error of 6.12%, the LSTM approach has only 2.13%.

This paper presents a context-aware adaptive assembly assistance system meant to support factory workers by embedding predictive capabilities. The research is focused on the predictor which suggests the next assembly step. Hidden Markov models are analyzed for this purpose. Several prediction methods have been previously evaluated and the prediction by partial matching, which was the most efficient, is considered in this work as a component of a hybrid model together with an optimally configured hidden Markov model. The experimental results show that the hidden Markov model is a viable choice to predict the next assembly step, whereas the hybrid predictor is even better, outperforming in some cases all the other models. Nevertheless, an assembly assistance system meant to support factory workers needs to embed multiple models to exhibit valuable predictive capabilities.

This article evaluates several machine learning methods to substitute the missing light detection and ranging data for better spatial localization of industrial automated guided vehicles. Decision trees and ensemble of trees using bagging or boosting techniques have been considered. Also, the $k$ -nearest neighbors algorithm was analyzed. Most of the algorithms have been evaluated based on multiple criteria and hyper parameter tuning. The analysis of the results was done in a comparative way, multiple regression evaluation metrics being considered. The experiments have shown that the extreme gradient boosting algorithm provides the best results in terms of performance, but with timing and resource allocation drawbacks. On the other hand, a simple decision tree model seems to give good results if a tradeoff between performance and prediction time must be made. The $k$ -nearest neighbors algorithm is also performing pretty well, especially because we are experimenting in a static environment.