The basic idea behind predictive maintenance is to collect relevant data in real time, evaluate it using suitable models as part of a data analysis and predict when maintenance is due in good time - in other words, to answer the question "What will happen when?" based on the current data situation. The data analysis used for this is known as "predictive analytics". It can be carried out with the help of supervised machine learning, a branch of artificial intelligence. Similar applications have long been widely used in the IT world to predict future developments, such as customer behavior, stock market prices or trend forecasts. Mathematically, the whole thing is a regression task with several predictors - i.e. special variables that describe a machine state, for example. In most cases, this is where the first practical problem becomes clear: data from which suitable predictor or forecast variables can be formed does not exist at all in most machine landscapes.

Create a data integration layer

The machine landscape of a manufacturing company with series production has usually evolved over a long period of time and is therefore extremely diverse. Machines and subsystems from different manufacturers, some of which do not even support a common communication standard, have been combined to form systems. Each individual machine requires extensive expert knowledge. Most system components have their own individual automation technology. In some cases, remote maintenance access is available so that external help can be obtained from a machine manufacturer or service company in the event of problems. In some cases, there are also cloud connections, for example for frequency converter monitoring - all in all, a very heterogeneous technology landscape without open data interfaces.

Before successful predictive maintenance applications can be implemented in such an environment, an open horizontal data integration level is required. This should correspond to the state of the art in the IT world and, if possible, only support three protocols via Ethernet LAN with MQTT, HTTP(S)-based REST API and, if necessary, OPC UA. All subsystems that do not directly support Ethernet and at least one of the three protocols are connected as a data retrofit with the help of a suitable adapter (the market now offers a very large selection here).

Before integrating the data of the individual machines, the application-related question of "What machines do we actually have, what do we want to achieve with a predictive maintenance concept and what data and information are likely to be required for this?" should be clarified as far as possible. An on-site analysis of the current situation and expert discussions between experienced data specialists, machine operators and automation technicians are required for the overall data-related picture. In each case, the goal is a so-called real-time feature vector with context-related cross-sectional data that fits the respective task. The individual data elements of this feature vector can be used to determine which additional sensors are required.