H2020 STAR: Leading Edge Research for Trusted Artificial Intelligence in Production Lines
In recent years, we are witnessing the digital transformation of production lines as part of manufacturers’ transition to the fourth industrial revolution (Industry 4.0). Based on Cyber Physical Systems (CPS) and digital technologies like cloud computing, the Industrial Internet of Things (IIoT) and Artificial Intelligence (AI), Industry 4.0 is enabling flexible production lines and supporting innovative functionalities like mass customization, predictive maintenance, zero defect manufacturing and digital twins. AI is currently the most disruptive digital enabler of the Industry 4.0 era and enables novel use cases like predictive quality management (Quality 4.0), effective human robot collaboration, agile production, and generative software design. State of the art AI systems in industrial plants operate in rather controlled environments. Nevertheless, AI systems in industrial plants must be safe, trusted, and secure, even when operating in dynamic, unstructured and unpredictable environments. Ensuring the safety and reliability of these systems is a key prerequisite for deploying them at scale and for fully leveraging the benefits of AI in manufacturing. This observation is fully in-line with the guidelines of EU’s Expert Group on AI, which mandate that AI systems are robust from a technical perspective and take into account their social environment.
STAR is a joint effort of AI and digital manufacturing experts towards enabling the deployment of standard-based secure, safe reliable and trusted human centric AI systems in real-life manufacturing environments. STAR researches, develops, validates and make available to the AI and Industry4.0 communities novel technologies that enable AI systems to acquire knowledge in order to take timely and safe decisions in dynamic and unpredictable environments. Moreover, the project researches technologies that will AI systems to confront sophisticated adversaries and to remain robust against security attacks. In this way STAR’s solutions eliminate security and safety barriers against deploying sophisticated AI systems in production lines. The project’s results will be fully integrated into existing EU-wide Industry 4.0 and AI initiatives (notably EFFRA and AI4EU), as a means of enabling researchers and the European industry to deploy and fully leverage advanced AI solutions in manufacturing lines.
STAR acts as a catalyst for ethical AI deployments in production lines, given that the project’s results are fully aligned to the recently published ethical guidelines of EU’s HLEG on AI. Specifically, STAR will produce technical solutions that boost the safety, robustness, and trustworthiness of systems AI in dynamic, real-life settings, while at the same exploring the legal implications of a safe and secure AI in prominent manufacturing scenarios.
To address the challenges of ethical, trusted, and secure AI systems STAR carries out leading-edge AI research and innovation activities in the following areas:
1) Explainable AI: STAR researches and will provide a library of explainable AI (XAI) techniques for manufacturing use cases such as Quality4.0 and human robot collaboration. The library will contain algorithms that explain the operation of deep learning systems based on their dominant features (e.g., Deep Learning Important Features (DeepLIFT) and Prediction Difference Analysis techniques), other popular XAI techniques (e.g., LIME), as well as algorithms for explainable robotics. STAR will enable recording and replaying the computations that are associated with decisions as a means of understanding how specific inputs lead to given results. The latter process will also facilitate the transparency and predictability of the AI systems in the shop-floor, while at the same time boosting security and safety.
2) Active Learning (AL) and Simulated Reality (SR) for Fast, Safe and Efficient On-Line Learning and Knowledge Acquisition: STAR researches AI systems that operate in dynamic manufacturing environments, while acquiring knowledge in a fast and safe matter. Specifically, STAR researches advanced and efficient forms of Reinforcement Learning (RL), including: (i) Active Learning (AL) approaches that enable robots and other AI systems to query human experts about their next course of action. Such AL interactions will be employed in cases where robots and other AI systems have low confidence about what to do next, but also as a means of accelerating acquisition of knowledge; and (ii) Simulated Reality (SR) approaches enhance RL systems with on-line simulations as a means of enabling RL agents to simulate the outcomes of their next action before actually taking it.
3) Human Centric Digital Twins for Simulation of Safe and Trusted AI Applications with the Human-in-the Loop: STAR researches and will provide advanced Simulation and Digital Twins solutions for AI-based “human in the loop” processes, including human robot collaboration. The project will provide technologies that simulate human behaviour and the interactions between humans and robots, as a means of detecting safety issues in the AI-based manufacturing process. The simulation and digital twins technologies of the project will address several use cases and problems, including: (i) The detection of safety zones for humans; (ii) The optimal deployment of fleets of Automated Mobile Robots (AMR) based on RL techniques; and (iii) The identification of safety issues through monitoring of the worker’s activities and status (e.g., performance, fatigue) in a given task context. To support such simulations, the project will specify and model the digital image of the human worker, while using it to develop Human Centric Digital Twins.
4) Security for AI systems: STAR will research, implement and validate solutions for securing AI systems in manufacturing, including technologies that address attacks at both the training (i.e., poisoning) and the operational (i.e., evasion) phase of Deep Neural Networks (DNNs). The project’s AI security solutions will boost the robustness of DNNs against adversarial inputs and attempts to contaminate the training datasets. Among other techniques they will leverage the project’s XAI library towards identifying and remedying hacked AI systems. Furthermore, the project will provide a decentralized solution for data reliability, which shall ensure the availability of high-quality data for training and operating AI algorithms regardless of the distribution and heterogeneity of the data sources (e.g., ERP, smart objects, automation devices, sensors etc.). The solution will leverage a blockchain infrastructure and will exploit background experiences of the project’s partners. It will boost the reliable storage and management of industrial data and of AI algorithms configurations.
STAR’s focus on the above-listed research areas (i.e., explainable AI (XAI), Active Learning (AL) and Simulated Reality (SR) in manufacturing use cases, human centred digital twins for AI in manufacturing, security and trust for AI systems in manufacturing) places the project at the forefront of the global research in AI in general and in digital manufacturing in particular. The project leverages background projects and results of the partners in the above areas, which ensures research excellence and will enable STAR to stand out from similar research initiatives worldwide.
To remain up-to-date about the project’s outcomes, please follow us:
By: Dr John Soldatos, INTRASOFT International