The disproportionately increasing amount of data generated during the production process today is one of the main drivers of global megatrends such as Smart Industry, Industry 4.0 and the Industrial Internet of Things (IIoT). Inexpensive and space-saving sensor technology is installed virtually everywhere where potentially interesting measured variables can be assumed - enormous storage capacities ensure that no bit or byte is lost. But what happens to all the measurement data? How is it already helping to improve the production process and the machines involved?

One area of application that has already been successfully implemented by many innovative companies is predictive maintenance. Mathematical models are used to derive predictions about the condition of the system based on the machine and production data that is continuously read in (health monitoring). These are then generally used to optimally plan service intervals, avoid or minimize production downtime and maximize production turnover.

The predictive maintenance workflow can be divided into four steps - reading in the data, pre-processing, prediction based on predictive models and integration into the IT system.

© MathWorks

The technical implementation of predictive maintenance can be divided into four broad steps. In the first step, the data is read in. This can be done directly from the industrial controller installed on the machine - for example via OPC UA or MQTT - or via an intermediate step using a database or cloud. In the second step, the data is pre-processed. This may involve smoothing noisy sensor signals, interpolating missing measuring points or transforming the data into the frequency domain. As pre-processing is often time-critical and reduces the amount of data to be transmitted, this step is often carried out directly on the embedded controller or the real-time PLC. The third step involves the actual prediction based on prediction models, for example using machine learning. In practice, the predictive maintenance workflow is integrated directly into the production plant's existing IT system.

Reading and pre-processing

In order to be able to make predictions about the condition of the system and the production process based on measured machine, product and energy data, the data must first be read in and pre-processed before it is fed into the actual prediction algorithm. In many applications, this is simplified by the fact that the measurement data is already stored in a database, on a central server or in the cloud. In this case, the data only needs to be retrieved via an appropriate interface. In some cases, however, it is necessary to read in measured values directly from the real-time system - i.e. a PLC, an industrial PC, a servo drive or an embedded controller. Universal communication protocols such as OPC Unified Architecture (OPC UA) or Message Queue Telemetry Transport (MQTT) are generally used for this.

Pre-processing the data is an essential part of any successful predictive maintenance implementation. For example, noisy signals are smoothed using filters or missing measured values are supplemented by interpolation. Pre-processing is usually an extensive and time-critical step that can also significantly reduce the amount of data required for transmission to the predictive algorithm. For this reason, in most cases preprocessing still takes place directly on the real-time system. In practice, the corresponding algorithms are developed on the basis of models and then automatically converted into real-time-capable source code.

Machine learning as a basis

The predictive model is trained intuitively using the 'Classification Learner App' in Matlab.

© MathWorks

With increasing computing power, statistical methods for modeling predictions such as machine learning or deep learning are also becoming increasingly relevant in practice. These algorithms make it possible to model the processes taking place at the plant even if the analytical models are not known in detail. Platforms for data analysis, such as Matlab, offer an extensive range of machine learning algorithms that can be compared by the user in conjunction with the recorded measurement data and then implemented in the production system. The predictive model is trained intuitively via the 'Classification Learner App'. After the measurement data set has been divided into an identification and a validation part, various machine learning algorithms, such as support vector machines, nearest neighbor classifiers or decision trees, are applied in the interactive user interface.

For example, Mondi in Gronau, a manufacturer of packaging films, can save 50,000 to 80,000 euros in rejects per month on each of its film production machines by using machine learning algorithms in combination with recorded sensor and quality data.

Integration into existing IT systems

The algorithms and models for predictive maintenance are usually divided into a time-critical part, which runs on real-time capable embedded systems, and a non-time-critical part, which is integrated into the central IT infrastructure.

© MathWorks

In order to anchor predictive maintenance in the production process, the predictive models and algorithms are integrated into the existing IT system in practice. This can mean - as in the case of pre-processing - that the algorithms are converted into real-time-capable C or C++ source code and implemented on the industrial control system. In the case of non-time-critical statistical models, these run as part of the existing process software on the central production server.

The Stiwa Group, an automation company headquartered in Austria, was able to optimize its own production facilities and reduce the cycle time for its machining processes by up to 30% by integrating optimization algorithms developed in Matlab into its 'AMS ZPoint-CI' plant control software.

New business models

The benefits of using predictive maintenance for plant and machine operators are obvious. The optimized planning of service intervals, the reduction of downtimes and the maximization of production represent a directly measurable added value, which in many cases is reflected in return-on-investment times of just a few months. However, predictive maintenance also has clear advantages for plant and machine manufacturers. For example, European manufacturers in particular, who can hardly set themselves apart from the competition from East and West due to price advantages, can score points in international competition with more reliable machines. In addition, numerous machine manufacturers already offer service contracts for their customers that guarantee a minimum throughput and are based on predictive data models. As a result, they have opened up an additional business segment alongside the actual sale of machines and systems through continuous service sales.

Author:
Philipp Wallner is Industry Manager for Industrial Automation & Machinery at Mathworks.