Jan-2020

Nature of errors in machinery diagnostics

A reliability theory has been developing for almost 70 years, and all the time scientists, engineers and practitioners attempt to point out an instance of a machine’s failure to be able to prevent its breakdown.

Andrey Kostyukov, Alexey Kostyukov and Sergey Boichenko
DYNAMICS Scientific Production Center USA

Article Summary

Since that time, many kinds of instruments have been developed, and all of them - vibration pens, portable analysers, protection, and condition monitoring systems - have been pursuing that goal. Nowadays, some AI systems utilise the statistical analysis of big data to solve that problem; however, their developers meet either poor and insufficient data or wrong data which are either almost or completely not related with the equipment’s lifespan. So, it is inevitable that AI assesses the probability of failure seldom more than 50%. To be able to provide precise diagnosis of machinery health, the algorithms of AI should include the physics-based rules of degradation. Hence, there are three topics will be considered in this paper: what nature of errors of degradation process’ recognition is, what a failure is, and what it means to recognise a defect.

The nature of errors of degradation process’ recognition is not so explicit as it seems to be at first glance. First of all, the process of degradation could be defined as a period of machinery lifespan since the destructive forces had emerged and began to influence the weakest part of the machine, until defect has evolved and collapsed machinery operation. Second, within the process of degradation, machinery undergoes presumably three stages which could be specified as non-linear wear, exponential wear, and critical wear, as shown in Figure 1. And ultimately, there are two distinct errors – an error of static recognition and an error of dynamic recognition (Kostyukov & Kostyukov, 2009). Nature of the error of static recognition appears when a failure cannot be detected because either the wrong non-destructive testing (NDT) method is utilised, or proper the NDT-method is used via a wrong way. The nature of the error of dynamic recognition appears when a failure cannot be detected and prevented because an interval of monitoring equates or exceeds an interval of defect evolution from emergence to failure.

The next essential question to reveal the nature is what a failure is. Depending on whether the goal is to recognise defects, or to protect a machine from damage only, there are different instruments which need to be used. A protection system must identify the finite stage of degradation and protect the machine from damage. Because the finite stage of degradation of rotating equipment is constantly short, the predominate requirements for protection systems are to measure direct failure-related parameters, such as vibration displacement, and subsequently shut down a machine as fast as possible if the trigger of the measured feature has snapped. If the goal is to recognise defects, at least the exponential stage of degradation should be found out. It is well-known that the velocity of vibration displays most defects of rotating equipment at the exponential and finite stages of their evolution. For several decades, scientists and practitioners have developed methods and algorithms, notions and norms to be able to detect particular defects of rotating equipment and prevent failures. These reasons explain why most portable devices and condition monitoring systems measure the velocity of vibration, in some cases computing coherent parameters. But the prime challenge of defect recognition is the diagnostic skills of vibration experts. Moreover, the challenge of timely data collection isn’t usually resolved, and many of users employ portable devices for those activities.

However, what if we could convert the meaning of “recognise” to “predict” or even to “prevent” the defect? When a defect happens, an exponential wear does not appear simultaneously. Therefore, monitoring systems must identify the emergent non-linear wear of machine’s weakest part. To be able to do this, a monitoring system must be comprised of physics-based rules of defect recognition and features which are strongly corresponded with the most frequent defects of most of the machine’s parts. If so, this combination could be named as physics-based AI because the aforementioned features can be chosen only by using statistical analysis of the relevant data. Also, as this kind of systems provide virtual diagnostics of machinery health, they should be named as state (health) monitoring systems or diagnostic systems.

It’s essential that only real-time diagnostic systems, which have low errors of static and dynamic recognition of defects and which are able to identify destructive forces influencing on emergence of defects can shift operation reliability and maintenance efficacy tremendously towards paramount safety and uptime, in fact, disrupting the existing mechanism of conventional relationships among operators, the maintenance team, and management. This shift becomes possible only when we understand the nature of the errors of the degradation process’ recognition.