Predictive maintenance - from data acquisition and forecasting to smooth operation

Where's the best place for a heart attack? An somewhat cynical question from the department of opportunities and risks. If there's a medical emergency that we hope won't happen, it's best to occur in front of a hospital, right? You'd think so, because every second counts in an emergency. But there is a better place. Where? In the pit lane area of a race track. Every box of a reasonably equipped facility is prepared with an emergency telephone and even at hobby events a battle-tested emergency physician is literally ready to jump and can reach every point of the race track within 90-120 seconds, the pit lane is of course correspondingly faster. Well, from this point on a discussion can unfold excellently, whether there aren't places somewhere, where expert help is faster on the spot. But it was only an example to open the view for the perhaps not so obvious.

But how do we move from a medical scenario to predictive maintenance and possibly to DATATRONiQ? Human beings, like machines and industrial plants, are concerned with complex systems, with vulnerable points, in the broadest sense with disturbances or breakdowns, and all too often they are directly or indirectly concerned with human lives. And despite all the optimisation of the reaction speed, both considerations must not lose sight of the fact that the aim must be to counteract incidents and emergencies well in advance of their occurrence. Experience has shown that prevention is better than cure.

Routine maintenance according to a stubborn plan

With regard to technical equipment, this is reflected in service and maintenance intervals for complete systems and components. Regularly renewed filters, lubricants and wearing parts ensure reliable operation and a long service life of the entire system. But when is the right time for maintenance or replacement of a component? In many cases, preventive maintenance, i.e. precautionary maintenance at fixed intervals, was and is the focus. The design of these intervals takes into account the results of tests under laboratory conditions as well as more or less standardized empirical values that determine, among other things, what is to be checked during maintenance and when assemblies or individual parts are to be replaced.

However, it is difficult to design everything in such a way that regular maintenance intervals are suitable for all components of a system in every area of application. The components are too diverse for this and, above all, the stresses and strains in the respective individual application. In addition, there are unforeseeable fluctuations in stress due to complex operating and environmental conditions such as operating range, temperature, dust, length of the respective application, air humidity, etc. What can make sense for a stationary industrial plant can be completely different for a wind farm or a container ship. The fluctuation margins of the general conditions become clear when one considers, for example, the different loads on a harvesting machine in the summer of 2017 and 2018. While the 2017 harvest was marked by vigorous wet grain on wet soils, frequent interruptions and the need for rapid readiness for use, the summer of 2018 posed a challenge with heat and dust. And the load on the machinery was just as varied.

Precise planning of preventive maintenance is correspondingly difficult. However, it is always more economically and humanly attractive to carry out a planned replacement than having to repair it spontaneously on public holidays or in the seasonal business while the system is running at night under poor night-time conditions.

Ideal image - continuous expert appraisal

How does a perfect predictive maintenance look like in contrast to a strictly planned maintenance? Like this, for example: Remove the actuator every 10 minutes, disassemble it and subject it to a professional inspection, i.e. not only testing, measuring, gauging, but also let the appearance and the sure instinct speak for themselves, or even smell it, and all this without interrupting operation. This may be ideal in theory, but it is simply nonsense.

In human medicine as well as in technology, indirect condition monitoring has always been used: why not listen to the little patient first before opening him up? An impressive example of this procedure is the ageing film "Das Boot", in which machinist Johann (the "Ghost") monitors the propulsion of his submarine via stethoscope or ear trumpet and counteracts the slightest deviations of the valve concert with oil can and grease press. However, such sensitive technicians are an expensive scarce commodity. And they also have the decisive disadvantage that they have to sleep.

Predictive Maintenance - continuous expert appraisal

Today, the stethoscope can be replaced cost-effectively by sensor technology, which measures vibrations and structure-borne sound with high resolution and extremely high sampling rates. However, the expert analysis of the data obtained requires highly developed analytical procedures. This is illustrated by the example of video surveillance.

Getting the images of a camera on a screen at the entrance of a power plant is no longer a great art. You can now place a person in front of this screen who can stare at the screen day and night and take action when a suspicious person approaches. This is expensive and error-prone, as it is both manual and tedious work. The first generation of video analysis software provided for the detection of movements in the camera image - there is something that should be looked at more closely by a human being. The following generation of video analysis software could already tell people from stray dogs or foliage moving in the wind. This means fewer alarms and less alarm verification by security personnel. The person detected can be classified by face recognition - is it the plant manager who comes back to the plant at 10 p.m. or is it an external person? Of course, the analysis software does not look at the images, it only analyzes countless pixels in the data stream every day, recognizes deviations from a digital target image and classifies them. Yes, we are drowning in measurement data, and the higher the resolution of the camera image, the more true this is. With the help of sophisticated analysis tools, however, it is possible to automatically elicit exactly what we want to know from the terabyte-wise digital data: is there someone we should not trust?

As part of the condition monitoring of special mechatronic assemblies, a high-resolution acoustic image is created using appropriate sensor technology, for example by measuring structure-borne noise. Even as good as new, high-precision mounted components generate vibrations that may not be audible to the human ear but can be easily measured by direct scanning, for example by measuring directly on the machine element, and can thus be used as a setpoint basis. Here, too, the analysis does not monitor anything according to human standards, but rather analyzes the resulting digital data stream. Detecting anomalies in the digital sound path, possibly long before it becomes perceptible to the human ear, is the first step. Learning algorithms such as those of DATATRONiQ's data-driven condition monitoring can be trained in such a way that the second valuable step is also successful - classification. If the type of deviation in the measured values is clear, well-founded decisions are possible with regard to measures because it is a difference whether a bearing is slowly approaching the wear limit or the running surface of a linear unit, which has a direct influence on the quality of a production. If the analysis tools are upgraded with practical expert knowledge, it is not only possible to predict that damage is imminent, precise statements can be made as to what type of damage is imminent at what point. The more data is collected and the higher the resolution of the data, the more precise the derived findings and decision bases become. Thus, the crushing mass of data loses its horror and can be automatically and profitably compressed.

The variety of measured variables to be recorded knows hardly any limits. Pressure sensors provide data for process monitoring and simultaneously support the detection of leaks in hydraulic and pneumatic systems. Piezoelectric force and acceleration sensors enable the analysis of mechanical loads under dynamic loading. Optical and thermographic processes also offer considerable potential for machine and, in particular, tool monitoring. Indirect variables in particular, such as electrical drive currents, provide information about a load behavior that changes slowly or abruptly, since the drive current is directly proportional to the torque that is requested from a drive. Ambient conditions such as temperatures, humidity, etc. complete the overall picture since the load case can only be described completely in this way. Many of the quantities mentioned are already being recorded and only have to be made accessible, or they can be additionally recorded with little effort.

The classification procedures trained in the system can be used to derive further details on the actual causes of an incipient fault from the measured values. Is the problem in the bearings of a drive or in the associated driven mechanics? Similarly, intelligently analyzed temperature curves can also indicate a lack of lubricant or excessive wear of bearings, guides and power transmission elements before failure occurs.

Regardless of the measurement method used, data volume, data quality and adaptive analyses are decisive for the quality of predictive maintenance forecasts. The more precise the measured values, the more often the measured values are recorded and the better the analysis procedures for error classification are trained, the more precise the basis for deciding whether a component is to be maintained or replaced or not yet. The pure mass of data is therefore less of a challenge. It is about targeted selection and recording of high-quality data streams, and the analysis of these data streams is about combining the expertise required for the respective application with first-class tools such as those from DATATRONiQ in order to reach economically sensible decisions quickly. And if less time has to be spent on data analysis, more time remains for the continuous further development and optimization of plants.