Exit Menu

Zero breakdown, a predictive maintenance for manufacturing systems


Project description


Horizon 2020: Strategies and Predictive Maintenance models wrapped around physical systems for Zero-unexpected-Breakdowns and increased operating life of factories 

Nowadays, the efficiency and sustainability of the manufacturing processes of high-tech products depend on the introduction of technologies towards digital, virtual and resource-efficient factories. In particular, the development of intelligent maintenance systems for increased reliability of production systems is considered to be crucial for securing competitive advantage for manufacturing companies. This trend is perfectly identified by “European Factories of the Future Research Association” (EFFRA) in the “Multi-Annual Roadmap for Factories of the Future” for 2014-2020 under the framework of the penetration of design and management manufacturing strategies and processes in the field of data collection, operation and planning, and from real-time to long term optimisation approaches.

At present, maintenance in general and predictive maintenance strategies in particular are facing significant challenges in dealing with the evolution of the equipment, instrumentation and manufacturing processes they support.

The success of those adaptive and responsive maintenance strategies highly depends on real-time and operation- synchronous information from the production system, the production process and the individual product, which should enrich and extend more traditional techniques and models. Nevertheless, it can be stated, that the traditional maintenance concept for fully automated production systems is no longer a viable vision, as it has been shown, that the conventional automation is not able to deal with the ever-rising complexity of modern production systems. Especially, a high reactivity, agility and adaptability that is required by modern production systems, can only be reached by human operators with their immense cognitive capabilities, which enable them to react to unpredictable situations, to plan their further actions, to learn and to gain experience and to communicate with others. Thus, new concepts are required, that apply these cognitive principles to support autonomy in the planning processes and control systems of maintenance operations of production systems.

Networked machine simulators are beneficial for failure analysis of production lines because of their modularity and flexibility in process engineering whose feedback loop is closed over communication networks. To enhance the process reliability, the main aim is to develop a general framework for investigating the fault detection and fault- tolerant control (FTC) problems of distributed simulation networks (DSNs). Specifically, we will focus our attention on following three measurable objectives: 1) design and deployment of a DSN that exhibits typical network-induced phenomena (e.g., communication delay, packet dropout and signal quantization); 2) design of network-based fault-tolerant control schemes consisting of a network performance recovery method and a network aggregation strategy; 3) design of a fault detection and isolation approach that ensures real-time failure analysis; and 4) testing and validation of the network-based DSN algorithms against sensor/actuator faults (hydraulic, electrical or mechanical components).

System Engineering Research Group (SERG) concerns with the creation of machine and production line breakdown taxonomy. The existing event (e.g. failure) signatures, patterns or models will be discussed analysed and clearly defined with the end-users by SERG team.

SERG will use Event-Clustering and EventTracking algorithms as Failure detection and prediction solution for decision making on the existing monitoring and control systems of the pilots. All the existing event (e.g. failure) signatures, patterns or models will be analysed. The outcome is the definition of system inputs, outputs and control parameters, i.e. the information and data management architecture for fault analysis. The second step is the exploration and the development of clustering and classification techniques for event detection employing a highly efficient image analysis technique using pyramidal type grouping, namely copasetic clustering. This algorithm will be highly effective to both the event detection and clustering. Pattern classification allows the identification of situations that are responsible for a certain event. SERG will use the efficient classifier developed in house, called Naïve Bayes classifier, on a subspace of carefully selected features. Also a range of recurrent neural network methods developed at SERG can be used to classify temporal data. The output will be a bespoke monitoring and control system designed for fault detection and prognostics.

Additionally, SERG will conduct time series analysis for event prediction. Since the corresponding data traffic time series might be rather short, useful models will be learnt from short time series using a combination of statistical time series model- ling, relevant domain knowledge and intelligent search techniques. Next, SERG will apply developed dynamic Bayesian networks to the modelling of process traffic data, which allows the testing of the effect of changes in a dynamic process by manipulating key components of the system. Finally, SERG will analyse the outlier for distinguishing real and false events. To enable successful separations of different types of outliers, we will employ knowledge-based method for distinguishing between measurement errors and phenomena of interest. We will also propose a cautious approach focusing on abnormality instead by modelling seemingly noisy data and error processes to make the distinction. As a final step, performance assessment will be conducted against variety of synthetic and real-world process data.