Z-Fact0r | Zero-defect manufacturing strategies towards on-line production management for European factories
Comments The technologies used were Artificial Neural Networks (ANN) and deep learning algorithms.
Periodic Reporting for period 1 - Z-Fact0r (Zero-defect manufacturing strategies towards on-line production management for European factories) Result description , an overall system architecture was designed including the main software functionalities, data information flows, middleware and other external sources, in such a way that all future efforts are coherently
Techniques for product reworking Result comments In Z-Fact0r, we exploited AM-based technologies as a tool for repairing of components in a production line.
Semantic/information interoperability Taxon title Semantic/information interoperability
Control technologies Taxon description technologies, system modelling approaches and distributed intelligence architectures.
Data processing Comments The technologies used were Artificial Neural Networks (ANN) and deep learning algorithms.
- Image-based feature extraction: Convolutional Neural Networks (CNN) and Variational Auto Encoders (VAE) used as feature extractors. CNNs will define the appropriate features that have been used for the classification between healthy and defected parts, and VAEs will be utilized to distinguish that can be used for image generation.
- Feature selection: An efficient filter feature selection (FS) method was developed for selecting informative and non- redundant feature subsets. In addition to enhanced accuracy rates and dimensionality reduction, the method have reasonably low computational demands. A robust and computationally efficient evaluation criterion with respect to patterns was defined allowing us to assess the redundancy between the features. The proposed FS technique was performed on a forward selection basis handling simultaneously both the discrimination power and the complementary characteristics between the extracted features. To decide on the number of retained features, a termination condition was finally introduced, thus avoiding the trial-and-error procedure usually employed in the common FS techniques of the literature.
- Classification: The selected features were input in a virtual classification module. The role of this module is to provide a decision on the workpiece condition. The technologies used were Artificial Neural Networks (ANN) and deep learning algorithms.
Nowadays, it is familiar that within the Industry 4.0, the ICT and the CPS, as parts of the industrial processes, are implemented and merged. For data collection, sensors are being used, imbedded within the AI in order to make smooth communication among humans and machines. Thus, Z-Fact0r is a pioneer with several advances in predictive maintenance, IoT sensors on shop floors and endless communication between the various components of the system, creating effective and many efficient applications for Industry 4.0.
Within the Z-Fact0r, the proposed (higher level) DSS, with the support of the knowledge base and the online inspection module (1st level decision support at single stage), produce, verify and validate decisions aligned with the quality control policies, production targets, desired product specifications and maintenance management requirements. Key functional characteristics of the envisioned DSS incorporates among others, techniques for monitoring and predicting product quality, action prioritization, root cause analysis, and mitigation planning algorithms (at product and workstation level). Moving beyond existing solutions that focus only on specific aspects of the production procedure, or that are restrained to diagnosis, the proposed DSS system incorporates autonomous, hierarchical decision support, based on process analytical technologies and newly developed suitably adjusted knowledge-based systems, and combines product monitoring models and data analytics from heterogeneous sources. The envisioned DSS takes into account a wide set of multiple factors and criteria, such as data uncertainty, lack of information and information quality, involvement of multiple actors, and real-time response. Thanks to the 5 intertwined zero-defect strategies (i.e. Z-PREDICT, Z-PREVENT, Z-DETECT, Z-REPAIR and Z-MANAGE) the overall solution presents a significant contribution to a spectacular improvement in the overall performance and reliability of the targeted multi-stage manufacturing systems and in the production agility (response to continuous adjustments in production targets).
DATAPIXEL provides the information associated to the defect detection in the manufacturing parts selected. This information is used as an input for developing the defect detection algorithms of Z-Fact0r solution. Based on this input, a data conditioning methodology has been developed to extract information concerning to the defect position and type. This information will be used as baseline for the model validation, via comparison with the respective simulation results.
The procedure that has been used is the following:
Z-Fact0r hybrid framework, obtained by applying a software and hardware integration strategy, is installed on the industrial end users shop floors. This architecture exploits features from Relational Databases and Triplestore while using the blackboard architectural pattern which ensures efficient and accurate communication of data transfer among software applications and devices.