Over the years, asset-intensive industries like energy, utilities and manufacturing have witnessed several service failures and safety incidents around machines and structures - without warning. Consultants estimate that typically manufacturers miss 5% to 20% of their productive capacity due to downtime¹. Even though organizations understand the cost implications of downtime, fewer than 24% of operators use a predictive maintenance approach based on data and analytics to handle asset failure proactively². Shifting from calendar-based maintenance to data-driven asset management can help anticipate problems based on asset health condition, enabling optimal usage of asset life and increased reliability.

Rigid physical laws govern assets that make inherent patterns, which are repetitive in nature and easy to decipher using machine learning algorithms. However, the patterns and failure behaviors are specific to asset type. Hence, a single predictive model cannot solve the issues of multiple devices used in these industries, which will own completely different characteristics and would demand exclusive treatment. The proliferation of failure detection systems for thousands of assets in large organizations has imposed an increasing burden in terms of cost and effort. Lack of availability of a generic framework is thus a major shortcoming in the current state-of-the-art failure prediction systems in large enterprises.

This paper will judiciously provide the business with an expansive guideline to select an appropriate analytical framework to predict asset failure based on the component type, failure modes and asset failure information.

Asset types and failure modes

It is critical to understand the similarities and distinctions of failure patterns and their sources across asset types. Firstly, we would begin with asset classification followed by analysis of key failure causing parameters across different asset types. Further, we drill down to component level failure modes to comprehend the pattern of the root cause. Lastly, we would arrive at a logical methodology to choose an appropriate analytical framework to build a model for asset failure prediction.

Asset-specific failure parameters

To create a robust technique applicable to multiple asset failure scenarios, it is critical to understand the component failure modes in different asset types and corresponding parameters viz. pressure, temperature, moisture etc. Figure 2 highlights some of the key parameters causing major failures in transformers and pumps.

Asset-specific component-level failure modes

To understand the impact of failure parameters, we performed a detailed sample analysis of the different types of component failures in Transformers, Centrifugal Pumps and Progressive Cavity Pump (PCP) systems. It was noted that while causes of failure is a function of operational parameters in case of static assets, it is a function of time in case of rotary assets. Hence, a failure-time distribution model is more relevant than a cross-sectional event prediction model for the rotary class. Thus, mechanical behavior plays a significant role in defining the model framework. Table 1 enunciates the systematic procedure of the analysis performed.

Framework for analytical model selection

According to Jerald F. Lawless, two sets of choices in analytical models are whether to use discrete or continuous-time models, and whether to use parametric or nonparametric assumptions³.

The model should effectively capture features of the lifetime distribution that are seen from empirical data. Large samples are often required to substantiate the superiority of one model over another in terms of goodness of fit. Censoring schemes are extensively used for comparison of models and their usability. This leads to use models that are computationally useful, and to an extensive use of Weibull, log-logistic, and log-normal models. As the number and complexity of fixed covariates increases, the emphasis on distributional shape reduces, the primary focus being on location and dispersion aspects.

Nonparametric and semiparametric methods are more robust with respect to assumptions than fully parametric methods. It allows more information to pass from the current set of data to the model at the current state, to be able to predict any future data. Several models may provide a description of the observed data. However, use case objective and actionable insights help explain characteristics of data much better.

Thus, the generic framework for asset failure modeling takes into consideration the asset type and failures as a function of time or operational parameters. Environmental parameters have a snowballing impact over time, which helps determine the probability of occurrence of failure by analyzing how particular circumstances or characteristics increase or decrease the probability of survival. However, on the other hand, numerous other operational parameters could trigger asset failure. For these scenarios, causal modeling is required for the identification of lead indicators instigating failures. Figure 3 illuminates the comprehensive approach to choose an analytical framework in case of asset failure modeling.

Real-life application of analytical framework

A leading Oil & Gas major based out of Australia enabled PCP pump failure prediction by leveraging this analytical framework.

The challenges the organization faced included inconsistent view of assets and their operational status. Also, manual intervention was slowing down preventive maintenance and leading to inaccuracy. The analytical framework enabled the prediction of the expected time for pump failure through the insights gathered by identifying influential parameters with survival analysis.

Table 2 showcases additional issues that were addressed using the analytical framework.

Use of Bayesian methods as future trend

With the advent of sophisticated computational capability and integration of analytics in key business processes, Bayesian models are going to be the need of the hour for the future of reliability analysis. They have a very important advantage of using information from different studies, prior information based on previous experience and engineering knowledge alongside the appealing Machine Learning (ML) characteristics of versatility and statistical properties. Advancements in the implementation of Bayesian paradigm and computer hardware have opened the doors of a new horizon in front of analysts to marry the engineer’s knowledge with statistician’s theory.

References

1. https://www.isa.org/standards-and-publications/isa-publications/intech-magazine/2006/January/channel-chat-when-true-cost-of-downtime-is-unknown-bad-decisions-ensue/

2. Study by Kimberlite, an international oil and gas market research and analytics company (http://phdsoft.com/category/phdc4d/)

3. Lawless, Jerald F (2011). “Statistical Models and Methods for Lifetime Data”, John Wiley & Sons, ISBN 978-0471372158