Maximizing the availability of production or energy systems is a major challenge for many companies. If machines in tightly synchronized production and logistics processes break down due to unforeseen damage or unscheduled maintenance work, companies quickly come under pressure with regard to the targets agreed with their business partners. Storing spare parts, which generally ties up too much capital for too long, has also long been considered an insufficient solution.

Harmonizing connections

The validated condition classes and target functions for the individual criteria are defined in advance.

© PSI FLS Fuzzy Logic & Neuro Systems

Against this backdrop, the focus is increasingly on the predictive maintenance approach. This promises to reconcile the often complex interrelationships and has gained further momentum with the new technologies and opportunities of the fourth industrial revolution.

A study by McKinsey even sees predictive maintenance as one of the most important fields of application for the Internet of Things (IoT) and anticipates potential savings of up to 630 billion US dollars by 2025. For example, 20 to 40 % lower maintenance costs for production facilities and medical products, 50 % less downtime and 10 to 40 % lower maintenance costs for airlines, for example, can be expected thanks to condition-based maintenance. In addition, investment requirements will fall by 3 to 5% due to the longer service life of products and systems. This is primarily based on the opportunities arising from the increasing networking of production systems and the systematic use of machine data supplied by sensors.

The market now offers tried-and-tested systems for this, such as the 'Qualicision' decision-making software from PSI FLS Fuzzy Logik & Neuro Systeme. Combined with artificial neural networks, it enables predictive or even automatic situational maintenance based on large amounts of data in the sense of big data.

Categorization of maintenance-relevant signals

In order to derive automatic rules or an automatic classification from the machine data supplied, the software differentiates between relevant machine data such as temperature, pressure, working hours, date of last maintenance, power consumption or criticality of the machine failure and between their negative, normal and positive effects on maintenance. For this purpose, the secured condition classes and target functions for the individual criteria are defined in advance. Together with the machine experts, the first step is to define the relevant criteria and the various classification clusters (corridors) for data evaluation.

For example, questions such as "In which value ranges should a gradation be defined according to the urgency of maintenance?" are taken into account. As a result, together with the machine data supplied, a qualified decision can be made with regard to a criterion or a combination of criteria as to when maintenance is necessary.

The software can be used to prioritize the criteria and define recommended decisions.

© PSI FLS Fuzzy Logic & Neuro Systems

The machine data is then classified according to categories such as 'Acute maintenance requirement', 'Medium-term maintenance requirement' and 'Long-term maintenance requirement' or according to further, definable gradations. Artificial neural networks learn the associated condition classes and then recognize them automatically.

In the second step, the software automatically qualifies the sensor data supplied by a maintenance-relevant object, such as a machine, as data records according to the defined and relevant maintenance criteria. The software can prioritize the criteria differently and give them a higher or lower weighting in the interactions. In this way, the maintenance-relevant signals can be categorized. Qualicision thus makes it possible to provide comprehensible information for decision-making when planning and controlling maintenance teams and even to automatically trigger the scheduling of an urgent fault clearance case or a scheduled maintenance activity - for example in a workforce management system - based on the recommended category.

Integrate maintenance processes

The industry has long since recognized the obvious benefits of predictive maintenance. The latest solutions, which link decision-making software with artificial neural networks, for example, make it possible to integrate maintenance processes into existing company processes without having to revise the basic production processes. As a result, they lead to better production planning, a longer machine service life and meet the goal of minimizing unplanned, expensive downtimes and even high reinvestments.

Advanced fuzzy logic and artificial neural networks

The Qualicision (Qualified Decision) decision-making software stands for qualified decision support in the optimization of complex business processes based on the specially developed, complementary extended fuzzy logic. Fuzziness arises in particular from the variety of data and the interactions between the possibilities for controlling complex business processes and the process targets in the form of key performance indicators (KPIs). Fuzzy logic allows, for example, the expression of a property such as 'a little', 'quite', 'strong' or 'very' to be combined symbolically and numerically (subsymbolically) to strengthen or weaken a predicate and thus to model the fuzziness of a linguistic expression with mathematical precision. Qualicision determines qualified selection decisions or the creation of rankings for selectable alternatives. The software evaluates the available decision alternatives with regard to the fulfillment of the process objectives and presents them as an impact matrix. This is examined by means of a conflict and compatibility analysis (CT analysis). The result is a balanced decision or a ranking of decision alternatives that reflects the priorities with regard to the decision-maker's process objectives. From a technical point of view, the CT analysis makes the so-called combinatorial variety of control options manageable with regard to the optimization of KPIs.

The Qualicision software is combined with artificial neural networks. Such networks are an algorithmic approach based on the biological process of activating and processing neurons in the brain in order to solve complex data processing problems. Convolutional neural networks (CNN) have established themselves in recent years, particularly in the field of image processing, thanks to their classification capabilities, which are superior even to those of humans. In the areas of sequence analysis, speech recognition and text classification, recurrent neural networks (RNN) are achieving top performance in current problems. Artificial neural networks are superior to many other machine learning methods, such as statistical regression, particularly due to their robustness against errors and data noise. The immense increase in the computing power of graphics cards in recent years has also made the complex training of many millions of pieces of data possible. The aim is that applications of artificial neural networks should now offer significantly more possibilities for explaining the solution path in a qualified manner. PSI FLS aims to achieve this goal by combining artificial neural networks with Qualicision aspects.

Qualicision is used as a decision support and optimization technology across all industries, for example for the optimization of production sequences in the automotive industry or in manufacturing companies in general, for the management of transport processes, the optimization of operational processes such as in bus and streetcar depots or for the maintenance management of electrical networks as well as the optimization of production control processes and forecasting methods.

Author:
Dr. Rudolf Felix is Managing Director at PSI FLS Fuzzy Logik & Neuro Systeme.