27 KiB
- Mdnotes File Name: trakadasArtificialIntelligenceBasedCollaboration2020
Green Annotations (18/12/2020, 15:08:12)
"important enablers for AI adoption in manufacturing contexts and is promoted in terms of business intelligence optimization, human-in-the-loop and secure federation across manufacturing sites" (Trakadas et al 2020:5480)
"(1) components that extend the functionality of the existing layers in the Reference Architectural Model for Industry 4.0; (2) definition of new layers for collaboration by means of human-in-the-loop and federation; (3) security concerns with AI-powered mechanisms." (Trakadas et al 2020:5480)
"most AI techniques are based on mathematical models that are dicult to understand by the general public, so most people use AI-based technology as a black box that they eventually start to trust based on their personal experience" (Trakadas et al 2020:5481)
"The application of human-centric AI (HAI) in internet of things (IoT) systems, so that IoT systems cannot only learn from users but also provide easy-to-understand explanations about decisions or estimations is a new research field [12]." (Trakadas et al 2020:5481)
ASPETTI MOTIVAZIONALI MOLTO IMPORTANTI. (note on p.5481)
"At the same time, introducing AI will lead to a more productive and safer working space, relieving human workers from routine procedures and employing intelligent machines and robots to perform heavy tasks, thus allowing human workers to focus on creativity, reasoning and decision making [20,21]." (Trakadas et al 2020:5481)
MOTIVATION (note on p.5481)
"to provide an IIoT-based system that increases performance and safety in the manufacturing domain." (Trakadas et al 2020:5481)
MAIN GOAL (note on p.5481)
"AI solutions are implemented in dispersed and isolated components of manufacturing IT systems." (Trakadas et al 2020:5482)
"Current Industry 4.0 reference architectures do not properly integrate the needed building blocks such as new deployment paradigms (e.g., edge-based learning to reduce bandwidth load on the enterprise network)" (Trakadas et al 2020:5482)
"scalable data-processing pipelines and information models" (Trakadas et al 2020:5482)
"AI-enabled digital twins used for monitoring and optimizing business intelligence [24-26]" (Trakadas et al 2020:5482)
IMPORTANT MOTIVATIONS AND CONTEXT DESCRIPTION (note on p.5482)
"availability of big data has been one of the most important enablers for the recent wave of AI innovations [27]" (Trakadas et al 2020:5482)
"Moreover, every phase of AI algorithm design requires high-level skills (model selection, training, hyperparameter optimization). In the agile manufacturing of the future, these costs must be amortized over low-volume batches (even lot-size-one)." (Trakadas et al 2020:5482)
IMPORTANT MOTIVATIONS AND CONTEXT DESCRIPTION (note on p.5482)
"AI technologies should not only be used for data analytics in support of business intelligence, but also for automated decision making on manufacturing process parameters and configurations." (Trakadas et al 2020:5482)
"AI algorithms are often black-box models (e.g., deep learning), while the inner workings of an algorithm fetched from a remote repository are not fully understood or the decisions of one algorithm create a conflict with other algorithmic decisions" (Trakadas et al 2020:5482)
"Proper secure federation mechanisms and AI-based cyberattack risk analysis are crucial cross-cutting concerns in AI-based manufacturing systems." (Trakadas et al 2020:5482)
"the performance of modern AI techniques requires large volumes of high-quality data which are often not available inside a single enterpris" (Trakadas et al 2020:5482)
MOTIVA BENE LA QUESTIONE DI LUCIANO, PERCHE' SI HA BISOGNO DI FEDERATED LEARNING (note on p.5482)
"AI techniques will be used to extend and improve the levels of communication and collaboration between computer systems and human workers." (Trakadas et al 2020:5482)
"new intelligent design and decision-making tools must be developed to promote human agency and oversight, simplifying the understanding and usage of AI results and considering multiple collaboration schemes depending on the situation. Human-AI will work in tandem in any phase of the product construction process, from design, over intelligent manufacturing execution monitoring to predictive maintenance." (Trakadas et al 2020:5482)
OBIETTIVI (note on p.5482)
"collaborative and intelligent factory of the future." (Trakadas et al 2020:5484)
"ration with other manufacturing sites, while security is aboration and for cross-cutting concern. In the following, we describe the functionality of each component as shown inis a cross-cutting Figure 2 concern" (Trakadas et al 2020:5484)
LUCIANO (note on p.5484)
"the factory-wide datasets so that the components of the upperyersistoperform layers can make decisions based on the outcome of AI algorithms running on top of such processedrlayerscanmake data streams or batch datasets. Raw (non-labeled) data generated by manufacturing devices ared data streams or processed alo batchdatasets" (Trakadas et al 2020:5484)
"re functionality to deploy AI algorithms closer to the sensor (edgeer to the sensor computing) and to detect shifts in the dataset statistics, indicating a need to retrain algorithms. Inetrainalgorithms" (Trakadas et al 2020:5484)
"AI-enabled data pipelines orchestrator component enables the creation and deployment ofnddeploymentof data processing pipelines, with two major objectivemajorobjectives:" (Trakadas et al 2020:5484)
"(i) the component should allow the possibilitythepossibilityto to set up pipelines consisting of typical data processing tasks (feature conversion, feature reduction,rereduction,data data anonymization and fusion, data cleaning, labeling and annotation, etc.) and AI models (used inodels(usedin the the services of the upper latheupperlayers)" (Trakadas et al 2020:5484)
"e factory; (ii) thei)thecomponent component should allow the deployment of the pipelines and the orchestration of the differentcomponents" (Trakadas et al 2020:5484)
"automate the deployment process on distributed infrastructure (edge device, edge cloud, public cloud) and to orchestrate the dierent modules and exact frameworks needed to run the processes." (Trakadas et al 2020:5485)
IMPORTANTE: RELATIVO AL AI PIPELINES (DESIGN AND DEPLOYMENT) (note on p.5485)
"Edge-based learning is required for latency-sensitive situations and/or when upstream bandwidth is insucient, e.g., audio and video from an augmented reality (AR) headset or processing light detection and ranging (LIDAR) data on a mobile robot. The component will support novel neural network architectures that can be trained without requiring large amounts of labelled data and that are resource-ecient [36]." (Trakadas et al 2020:5485)
IMPORTANTE LUCIANO (note on p.5485)
"threat intelligence manager takes advantage of the collected and curated datasets and applies AI algorithms for executing threat analysis in order not only to predict potential cybersecurity incidents but most importantly to manage and mitigate such incidents in a timely manner." (Trakadas et al 2020:5485)
"4.2. Functional and Business Intelligence" (Trakadas et al 2020:5485)
"mirror the state of the production process in digital twins, including the logic that determines the transition to other production steps or states." (Trakadas et al 2020:5485)
"This layer provides innovative tools that will facilitate intuitive and ecient collaboration between humans, machines and AI systems allowing them to take advantage of each other 's strengths for more eective cooperative and intuitive task execution and decision making" (Trakadas et al 2020:5486)
"Federated Learning component aims at solving the problem of data collection for feeding or training AI models, while assuring the ownership and confidentiality of the data. In manufacturing, most (if not all) data and information are confidential because they relate directly to details of the production process, product characteristics, volumes, etc." (Trakadas et al 2020:5486)
"The inter-manufacturing knowledge exchange serves as an interface for knowledge exchange across manufacturing sites or distinct manufacturing processes." (Trakadas et al 2020:5486)
"ince there is a need to control what information is exposed and exchanged, rather than allowing open access to the local knowledge repository, this component contains a query engine to handle external requests. Such query engines also enable the realization of a federated query-processing mechanism over multiple sites." (Trakadas et al 2020:5486)
FEDERATED QUERY-RPOCESSING METCHANISM OVER MULTIPLE SITES (note on p.5486)
"4.5. Security and Authorization" (Trakadas et al 2020:5487)
"security and authorization requirements on information and data sharing." (Trakadas et al 2020:5487)
"Cybersecurity for Artificial Intelligence (AI)" (Trakadas et al 2020:5487)
"Artificial intelligence attacks, i.e., attacks on the AI algorithm, can take two forms: input attacks and poisoning attacks. The former consists in manipulating the input to the AI system during the operation phase so that it delivers the wrong results. Input attacks are relatively easy to launch and succeed since they do not require a manipulated AI system. Poisoning attacks, on the other hand, have to do with the corruption of the process used to build the AI model. In this case, inaccurate or mislabeled data are provided to the model during the training phase to manipulate the learning process. This type of attack can also be launched against federated learning; in this case, manipulated data or an algorithm of a member of the federation can result in the corruption of the global model." (Trakadas et al 2020:5487)
*IMPORTANT! RELATED TO AML WORK!!!
IT APPLIES ALSO IN THE CASE OF FEDERATED LEARNING (note on p.5487)*
"challenges regarding the implementation of the innovative AI-based components of the proposed system" (Trakadas et al 2020:5487)
"AI-Driven Modelling of Manufacturing Assets" (Trakadas et al 2020:5487)
"Digital twins are virtual, high-fidelity models of the current state and internal behavior of physical assets on the shop floor [25]." (Trakadas et al 2020:5487)
"there is a lack of information models and process libraries that allow users to replicate and scale their digital twins." (Trakadas et al 2020:5487)
"AutoML" (Trakadas et al 2020:5488)
DA VEDERE AUTOML (note on p.5488)
"For the IT resource aspect, processing all raw data on a public cloud infrastructure is an unscalable solution for many manufacturing companies, either because there is too much sensor data to upload, the latency to the cloud is prohibitive or because the sensor data is too sensitive and the company does not want to expose this. Therefore, edge computing has been proposed and several reference architectures for edge computing in Industry 4.0 have been proposed [26" (Trakadas et al 2020:5488)
LACK OF RESOURCES FOR AI IN INDUSTRY (note on p.5488)
"Techniques to make deep neural networks available for the latter form of edge computing start to find their way into production as more user-friendly tools become available. For instance, TensorFlow Lite allows converting a trained model for deployment on microcontrollers or embedded Graphics Processor Units (GPUs)." (Trakadas et al 2020:5488)
TensorFlow Lite: Come diceva Luciano (note on p.5488)
"5.2. Multi-Channel, Context-Aware Interaction on the Shop Floor" (Trakadas et al 2020:5488)
Relativo alla parte di NLI di Mariani (note on p.5488)
"The latest developments [48,49] allow humans to convey information with AI systems through multiple channels by integrating advanced human-machine interfaces (gestures, facial expressions). These interfaces, which can be obtained by using 2-D and/or 3-D cameras and other sensors, such as gyroscopes or accelerometers, oer information related to the context and the situation that is relevant to the interaction." (Trakadas et al 2020:5488)
Gesture, facial expressions interaction etc (note on p.5488)
"voice understanding" (Trakadas et al 2020:5488)
"experts in simulations and modelling are not available in SMEs and large enterprises, and that data-driven digital twins are usually trained on data collected at system level. These limitations make it economically costly to develop novel digital twins, which is often needed in modern manufacturing with agile reconfigurations." (Trakadas et al 2020:5488)
IMPORTANT - LACK OF EXPERTS (note on p.5488)
"Data-driven digital twins are built using machine learning." (Trakadas et al 2020:5488)
"DL comes with its own problems: it requires domain knowledge to select the appropriate machine learning pipeline and it is very resource hungry in terms of computing and storage resources [36,43,44]" (Trakadas et al 2020:5488)
PROBLEMS ON DEEP LEARNING (note on p.5488)
"Intelligent Decision Support" (Trakadas et al 2020:5489)
"increase the eciency in the decision-making process" (Trakadas et al 2020:5489)
"urrent approaches fail to learn from decisions taken by humans." (Trakadas et al 2020:5489)
"ontinuous learning functionalities" (Trakadas et al 2020:5489)
"Threat Intelligence Manager (TIM)" (Trakadas et al 2020:5489)
"In the case of malicious activity, various AI algorithms, such as naïve Bayes, random forests and support vector machine (SVM) have been proposed [55]." (Trakadas et al 2020:5489)
"By exploiting advanced AI methods, the threat intelligence manager (TIM) component intends to model the dynamic interactions of Industry 4.0 subsystems and discover known and unknown attacks, while surpassing existing signatureand anomaly-based methods." (Trakadas et al 2020:5489)
"Federated AI across Manufacturing Sites" (Trakadas et al 2020:5489)
"In this context, federated query processing [59] is an active research field dealing with techniques for proper delegation of the execution of parts of queries to specific sources" (Trakadas et al 2020:5490)
IMPORTANT! FEDERATED QUERY PROCESSING (note on p.5490)
"To improve the scalability of such federations, aggregation techniques could be used, where one or more independent aggregators would continuously crawl sources, and maintain data summaries [60]." (Trakadas et al 2020:5490)
"Joining datasets held by dierent actors can address this issue. Oft" (Trakadas et al 2020:5490)
THAT'S THE GOAL TO ADDRESS THE ISSUE THAT SINGLE PARTY DOES NOT HAVE SUFFICIENTLY LARGE DATA SETS FOR TRAINING (note on p.5490)
"federated learning provides a solution enabling machine learning over distributed and decentralized datasets." (Trakadas et al 2020:5490)
"federated learning opens up new business models (AI as a Service, AIaaS) to analyze data provided by a customer" (Trakadas et al 2020:5490)
"federated learning represents a solution to this problem enabling both the service provider and its customer to achieve their objectives while preserving the business assets." (Trakadas et al 2020:5490)
"t is possible to generate in an automatic manner intelligent services in parts of the building blocks or in the process as a whole." (Trakadas et al 2020:5490)
"handling and labeling of data of very dierent types." (Trakadas et al 2020:5491)
"dierent system components can be exploited for optimizing manufacturing logistics processes, and facilitating zero-defect manufacturing" (Trakadas et al 2020:5491)
"The current AMR production logistics is supposed to be carried out by the system illustrated in Figure 3a" (Trakadas et al 2020:5491)
"t is assumed that there are two issues that need to be addressed in an ecient manner: (a) unpredictable delivery times and downtime, and (b) vulnerability to network attacks." (Trakadas et al 2020:5491)
"le and machines at high levels of safety and tenance services for their AMRs" (Trakadas et al 2020:5492)
"reduced by avoiding heavy vehicles like forklifts or tuggers in fast-moving intralogistics areasorevencompromise during pe site safety." (Trakadas et al 2020:5492)
"neously, AMR downtime is undesirable andAMRstothenetwork should be reduced as much as possible. entcybersecurityriskduetodynamicsoftwarevulnerabilitiesthataffect This scenario assumes that a company has implemented a flmunicationprotocols." (Trakadas et al 2020:5492)
"Advances leveraging the proposed platform" (Trakadas et al 2020:5492)
"learning from patterns of business process models and anomaly detection techniques" (Trakadas et al 2020:5492)
"prediction of AMR downtime can be done using multivariate statistical models on data of key AMR system parameters" (Trakadas et al 2020:5492)
"Edge-based (i.e., on-robot) learning can be used to reduce the amount of data uploaded via the customer network to AMR supplier 's cloud back-end and assure the confidentiality of production activity related data" (Trakadas et al 2020:5492)
"run AI observers that collect data about typical movement patterns on the floor (e.g., edge-based learning on surveillance camera data) and about how fleet behaviors relate to production goals (e.g., from Enterprise Resource Planning—ERP and Manufacturing Execution System—MES systems)" (Trakadas et al 2020:5492)
"As unauthorized access to machines and data might compromise the physical integrity of human workers, communication between AMR, the platform and AMR supplier 's cloud back-end can be secured by signcryption schemes provided by the system." (Trakadas et al 2020:5493)
"The number and the diversity of characteristics of the orders do not allow for common process standardization." (Trakadas et al 2020:5494)
DIVERSITY (note on p.5494)
"The data and information collected across the departments through manual and automatic procedures are solely processed by each department, without considering factory-wide optimization metrics." (Trakadas et al 2020:5494)
FACTORY-WIDE OPTIMIZATION METRICS (note on p.5494)
"The deficiencies are observed, in the majority of the cases, during the quality control of the final product, thus leaving no space for corrective actions." (Trakadas et al 2020:5494)
NEEDS FOR EARLY CORRECTIVE ACTIONS (note on p.5494)
"letting personnel only make the final decisions." (Trakadas et al 2020:5494)
LACK OF PERSONNEL KNOWLEDGE AND TRAINING ON THE USE OF INNOVATIVE TOOLS (note on p.5494)
"The combination of data analytics and In testing and validation procedure, the goal is to define a set of KPIs per potential use case andhat can be used for validate the performance of our proposed AI-barootcause analyses" (Trakadas et al 2020:5494)
WHAT-IF ANALYSIS, PROJECTIONS, AND ROOT CAUSE ANALYSIS (note on p.5494)
"manufacturers will be capable of realizing agile production processes and improve the quality of products and processes" (Trakadas et al 2020:5495)
BUSINESS IMPACTS (note on p.5495)
"more competitive in the market and thus increase their market share." (Trakadas et al 2020:5495)
"The federated intelligence layer introduced in our approach enables new business models (e.g., high-quality AIaaS), as well as the collaboration of dierent industries towards the creation of digital twins" (Trakadas et al 2020:5495)
"rchitecture that facilitates the collaboration between manufacturing machinery, AI and humans" (Trakadas et al 2020:5495)
FACILITATING ARCHITECTURE!!! (note on p.5495)
"Last but not least, companies with expertise in AI for manufacturing can create significantly higher revenues by being capable of integrating their components with IoT and IT systems from dierent vendors/creators." (Trakadas et al 2020:5495)
OPEN COLLABORATION AMONG COMPONENTS FROM DIFFERENT VENDORS/CREATIOS (note on p.5495)
"components for timely data collection" (Trakadas et al 2020:5496)
"processing and curation" (Trakadas et al 2020:5496)
"elying on the dynamic instantiation of data pipelines" (Trakadas et al 2020:5496)
"while addressing security, privacy and confidentiality concerns" (Trakadas et al 2020:5496)
"cross the physical and virtual entities" (Trakadas et al 2020:5496)
"The collected data and information models are transformed into AI-enabled functional intelligence, leading to business knowledge, actionable insights and informed decisions, while being capable of recognizing complex events and process deviations that cannot be captured easily and in a timely way through human judgment" (Trakadas et al 2020:5496)
MAIN GOALS/BENEFITS (note on p.5496)
"25. Tao, F.; Qi, Q.; Wang, L.; Nee, A.Y.C. Digital Twins and Cyber-Physical Systems toward Smart Manufacturing and Industry 4.0. Engineering 2019, 5, 653-661. [CrossRef]" (Trakadas et al 2020:5497)