Success Stories

This Use Case takes advantage of STAR’s security and data governance technologies in order to ensure the integrity, confidentiality, availability, non-repudiation and authenticity of industrial data used by AI systems and processes. Techniques for decentralized validation of industrial data are employed, notably techniques that audit/validate datasets derived from sensors and unreliable data sources. The UC leverages also the project’s cyber-defence techniques against Deep Neural Networks, with emphasis on confronting attempts for contaminating machine learning and deep learning algorithms.

The integration of AI has unlocked a realm of transformative capabilities such as automatic visual inspection, predictive maintenance, and optimizing production lines. However, it has also introduced new security challenges, as malicious actors exploit the interactions between AI and legacy ICT systems. Adversarial Machine Learning (AML) has emerged as a significant concern in critical AI applications, involving techniques that manipulate data to alter AI algorithm behaviour while it may remain unnoticed by humans.  This use case focusses on the implementation and testing of the cyber-security components developed in the topic of Security and Data Governance for AI Systems in Manufacturing. The goal of this use case is to ensure safe systems, protected against various cyber-attacks, meaning that the implemented AI systems are protected and safe to use within a manufacturing environment.

This use case makes use of "Human intention recognition" and "Robot reconfiguration based on the dynamic layout" use cases in order to provide a safer work environment for all resources. Updated coordinates of all elements (i.e. stations, obstacles, workers) will be put to use for robot path planning. The position and velocity of the objects in the layout will be considered for collision-free path planning. Current human and future behaviour are also taken into account. Relying on the next possible action of the workers, the robot will be routed to the target to prevent the potential collision.

This use case is geared towards utilizing techniques like simulated data, active learning, and XAI (Explainable Artificial Intelligence) to enable model training with limited training data. The incorporation of simulated data enhances the training process by exposing the model to a broad spectrum of data, even when real data is scarce. Active learning enables the selection of the most relevant training examples, increasing training process efficiency. Additionally, XAI provides insight into the model's decision-making, which is crucial for understanding and having confidence in the results.  One of the challenges faced in the future automation of current production lines relates to flexible evaluation of product quality using visual inspection. In high-volume production environments, systems can be trained and optimized on an abundance of labeled examples, which aids the model in understanding data patterns and relationships, allowing it to generalize and make predictions on new, unlabelled data. However, due to a shift in customer demand, and the wish to be able to produce a small series of products more often, lower volume production is becoming more common. This complicates the flexibility, efficiency and scalability of production lines and does not facilitate rapid responses to product adjustments. 

Human factors play a predominant role in IBER’s production processes i.e., the processes are not fully automated. Hence, training of human resources in the various production processes is planned and carried out in advance, in accordance with appropriate operational methods. This will lead to improvement of operative methods and consequently to the reduction of human errors associated with learning and assimilation. Moreover, due to continuous monitoring of production processes, human errors associated with overspecialization (i.e., when an operator is performing the same functions in the same workplace over a given period) will be reduced.

This use case is related to the detection of human presence and prediction of their behaviors. Having the human factor taken into account, together with robot navigation, better planning can be achieved but also a safer environment. In this regard, DFKI-EI simulated a typical working day where ten (10) workers participated. The tasks of the workers were predefined and of similar nature. The order of the workflow was not the same for all the workers in the dataset acquisition process. Data collection for this case was done via wrist sensors. So, all the information and data collected from workers is processed and used to detect the actual activity performed by workers.

Once a model is deployed in production, it requires careful maintenance. During production, unforeseen issues can arise, and humans are essential for immediate problem-solving and recalibration of the model. The synergy of machine learning models with human expertise is therefore crucial to optimize the production quality control process. Human feedback is vital for refining the model's understanding of evolving production conditions. As product variations or new quality issues emerge, human input guides the model's adaptation and validation, ensuring it remains relevant and effective in real-time quality control.  In this use case active learning is being utilized, where humans bring domain knowledge and expertise to label the most relevant data points to the model. This is particularly important as it allows effective allocation of resources, to devote them to labelling data that is most likely to enhance model accuracy.

This Use Case linearizes production planning towards reducing the work peaks and consequently the number of extra work hours, but also towards reducing production down times as well. It leverages advanced online process sensing/monitoring systems and different parameters and variables into the manufacturing process and products. To this end, non-destructive testing techniques are employed in order to measure dimensions and defects, while contactless and agile technologies will be incorporated.

This Use Case develops human-centred digital twins that simulate production processes in real time, towards identifying scenarios that ensure safety and achieve the required flexibility. The digital twin enables the high-level planning and management systems to access detailed real time information from the production line, in order to get the general state of any production cell for effective real-time optimization.

This use case consists in the dynamic update to the navigation route of the mobile robot, by considering human and/or other (non-moving) objects in the environment. It enables easier reconfiguration of the robot in case the layout of the environment (including the production stations) changes. The layout is constantly monitored with the help of two ceiling-mounted cameras and humans. Whenever the setting changes, the coordinates of the changed entity are also updated, so that any planning activity takes into consideration the newest modifications.

This use case emphasizes on the safety of interactions between human and machine during quality inspection tasks. In this use-case the goal is to ensure mental and physical safety of the factory workers, by monitoring workers’ fatigue when performing manual labelling of images.  By detecting workers’ fatigue, attention and/or mental stress during the labelling process it becomes possible to understand whether the labeled data can be trusted or if it should be reviewed by multiple workers to ensure the accuracy of the final label provided. It also enables monitoring worker health and, for example, to suggest a break or a change of activity to avoid disengagement.