Facing increased global competition, the manufacturing industry depends on high-level solutions to ensure excellent machine functionality. Current analyses estimate that system downtimes and component breakdowns lead to an energy waste of 33% in the production and a significant loss of profits. At the same time, the complexity of production plants is steadily rising due to increasing product variances, product complexity, and pressure for production efficiency. Production systems must therefore evolve rapidly and operate optimally, which creates challenges for larger industries and especially for small and medium-sized enterprises (SMEs).
To meet these challenges, the European research and innovation project IMPROVE joins forces from 13 leading players in the field of academia, industry, and software development from Europe and beyond. Together, they have developed novel data-based solutions to enhance machine reliability and efficiency. Innovative tools in the fields of simulation & optimization, condition monitoring, alarm management, and quality prediction provide manufacturers with a human machine interface (HMI) and decision support system (DSS) to ensure best possible user support.Read More
In IMPROVE project, MInD-NET is leader of "Condition Monitoring and Diagnosis" work package, which deals with the analysis and the interpretation of the data received using machine learning techniques. In particular, a two-step analysis is applied to data received from a number of sensors/transducers. In the first step, the temporal data are segmented and converted into a sequence. In the second step a mathematical model is constructed to represent the behavior of system. This approach is known as "Syntactic Pattern Recognition" and shown to be successful in many similar applications.
Comparing the system behavior recorded at any time interval to the normal/regular behavior, the non-normal behaviours are picked up (Anomaly Detection), which mainly contributes to another work package titled as "Model Learning". Further, comparing the system behaviour with those originating from different models, the problem of the system is identified (Anomaly Identification). By calculating the deviation from the normal mode, a future anomaly is can be predicted, which enables overcoming a critical anomaly without interrupting production (Predictive Maintenance).
For these purposes, MInD-NET developed novel techniques to detect switching points where the current behaviour of the system is changed and represent different behaviours using Variable Order Markov Models (VOMMs). Each VOMM is shown with a Probabilistic Suffix Tree (PST), enabling the use of many handy graph theory related tools. In order to match and cluster various behaviours of the target system, a matching function is developed to be used on comparing the PSTs generated. These mentioned algorithms and results are published in international conferences and journals.
MInD-NET also contributed to "Demonstrator and Prototype" work package by building a lab-demonstrator using programmable LEGO® Technic™ kits with various electric motors and sensors. This demonstrator is used as a semi-synthetic data generator to verify/validate developed algorithms before testing them with the data collected from an operational environment.