Technology > Alternatives
Prognostic Techniques Demystified:
Like many other technical and managerial disciplines, reliability management and predictive maintenance (PdM) have been subject to a myriad of mathematical and statistical techniques. Practitioners often encounter these techniques as buzzwords of seemingly universal applicability: Weibull fitting, stochastic processes, symbolic regression, Bayesian networks, Monte Carlo simulation, expert systems and neural networks, to name but a few. Beyond their broad exposure to commercial software and consulting offerings, reliability managers need a basic, pragmatic, and practice-oriented understanding of these techniques to judge their applicability, benefits, and limitations in their field of work.
How is Prognostics different from other asset management tools?
Cassantec helps to demystify prognostic techniques. We do so through our transparent and pragmatic solution configuration process, followed by an ongoing and dedicated working-level collaboration. Demystification is driven by the use of prognostic information in daily operations, and by the understanding of the nature and origin of such information in condition and process data. Demystification is also supported through access to our customer community of dedicated prognostic practitioners.
Are there alternatives to our prognostic approach? Contact Us
There certainly are alternative approaches, which we have used ourselves for different products and applications in the past: The main four legacy approaches that have been used for prognostic application in predictive maintenance (PdM) and reliability management are classic linear regression approaches, specialized non-linear regression approaches such as Weibull fitting and neural networks, and expert systems.
Cassantec's prognostic solution is in fact unique, resting on a novel combination of state-of-the-art technologies. Below, we discuss traditional alternatives which we have used ourselves for different products and applications in the past. We point out their respective limitations that disqualify them as serious competitors for our technology.
Basic predictive offerings extending classic condition monitoring often extrapolate asset conditions using linear regression. Yet, since condition parameter uptake is typically exponential rather than linear in nature, the underlying assumption of linearity rarely applies. Symbolic regression approaches with linear, polynomial, exponential, and other elements allow dropping linearity assumptions, but are too sophisticated and too data-intensive for PdM applications in practice. With few specialized exceptions, linear regression approaches are typically not appropriate for the computation of an asset's remaining useful life (RUL).
More advanced predictive offerings use Weibull fitting to approximate an exponential RUL curve. Weibull fitting is based on functional assumptions and empirical data on asset failure and ageing effects. Clearly, failure data is hard to obtain for an asset that is not run to failure. Crash tests are quite expensive. Ageing effects, on the other hand, are barely relevant for assets that are being maintained or even rebuilt during maintenance. In fact, according to syndicated statistics, only 11% of industrial asset malfunctions are age-related. Asset condition data such as vibration amplitudes or wear particle counts in turn is highly relevant for RUL computation, but cannot be formally translated to Weibull parameters. Hence, while RUL may be Weibull distributed, classical Weibull fitting is inappropriate for condition-based RUL computation.
Comprehensive support for reliability managers and component engineers is provided by diagnostic expert systems. Expert systems capture deterministic knowledge about industrial assets in a complex set of causal rules. Based on these rules, a logic inference engine generates advice for reliability managers and component engineers. Expert systems perform well in a complex but certain world, which excludes predictions or even approximations of an uncertain future. Since the laws of predicate logic and stochastic calculus do not blend together, probabilistic expert systems typically do not work in practice. Hence, expert systems are inappropriate for RUL computation.
Note, that for applications in asset condition diagnostics, simple rule-based systems without inference engines are often positioned as full expert systems. The resulting shortness of logical inference leads to an apparent parrot effect, which has been criticized by condition experts, who first supply their own diagnostic rules, and then receive highly plausible diagnostic insights with limited value added.
Predictive systems based on neural networks, finally, make very limited functional assumptions on an asset's RUL curve. Neural networks allow a diagnostic classification of asset conditions based on large sets of empirical condition data, yet they do not support explicit causal reasoning or stochastic inference about the asset's condition. Most practice applications are in fully automated environments where a magic black box is acceptable. Neural networks are inappropriate for RUL computation when interfacing with (human) reliability managers or component engineers.
In summary, the discussed limitations of the listed alternatives disqualify them as serious competitors for Cassantec's prognostic technology in practice.