In an Expert Focus article for WaterBriefing, Dr Atai Winkler, Principal Consultant at PAM Analytics, describes his work with Collaborate Water to develop a Predictive Asset Management System using data science and visualisation to improve asset resilience and environmental performance.

Background

Dr. Atai Winkler: The work described in this paper received a UKRI (UK Research and Innovation) award in September 2022. To quote from UKRI: ‘the government funding competition is available for highly innovative ideas that are significantly ahead of others currently available and propose an innovative use of existing processes and services, helping to achieve net zero and demonstrating that new technology can enable affordable, adoptable and investable innovations and catalyse further innovation’.

Benefits of Applying Data Science to Asset Management

The importance of maintaining assets in good condition to optimise their performance, achieve high levels of asset availability and longer than expected lifetimes, and minimise asset management costs is well documented. Now, additionally, utilities and similar industries are under unrelenting pressure and scrutiny from governments, regulators and the public to improve their environmental performance, and the safety and reliability of their assets.

Furthermore, the energy used by heavy duty assets depends on their condition. Assets in poor condition are at greater risk of failure and use much more energy than expected. Increased risks of failure can have severe environmental and financial consequences. This paper describes how the system developed by PAM Analytics and Collaborate Water for the UKRI award uses data science and visualisation to optimise asset performance and reduce the risk of asset failure.

The System

The system has three complementary components: data science, visualisation and secure cloud hosting. The data science component has data preparation, exploratory data analysis, deterioration curves, predictive analytics and simulation elements, and the visualisation component applies powerful viewing, drill down and reporting technology to the data supplied and to the outputs of the data science component. The solution is hosted by Collaborate Water in a cloud environment using their data science platform Dimension^TM. Cloud hosting allows secure, scalable hosting of the data and models, and the visualisations to be accessed.

Figure 1 shows the structure of the data science component PAM (PAM Analytics’ Predictive Asset Management system). Table 1 shows the role and output of each module in PAM and Table 2 shows example input data for PAM.

It is unlikely that all the data in the table will be available. Similarly, other data may be available. Thus, the data in the table can be regarded as a very good indication of the type of data required for applying data science to asset management.

Time to Failure Transformations Module

The exact preparatory work carried out in the module depends on the data supplied but the objective of the work is to make the data fit for purpose.

All the work order data that PAM Analytics has analysed has the problem of nested and overlapping work orders as shown in Figures 2 and 3. To the best of our knowledge this problem is not addressed in other systems but it must be addressed before analytics can be applied to work order data because these types of work order can lead to negative intervals between consecutive work orders and failure models (distributions) cannot handle negative times. The result of deduping such work orders is one work order that captures their true nature and does not coincide with other work orders.

Deduping reduces the number of work order records significantly. Overlapping and nested work orders have a number of possible causes including recording individual tasks rather than whole tasks as work orders, repeated attempts to complete the work were required, the work took much longer than expected and the work actually carried out was not recorded correctly.

The Need for Bespoke Models and Software

An integral part of PAM is the bespoke software PAM Analytics developed for deduping nested and overlapping work orders. This illustrates the drawbacks of using software that does not have powerful procedures and programming flexibility and so does not allow data to be prepared and modelled in ways that consider the particular features of the data-modelling data is a creative task, not a prescriptive process.

Asset Key Performance Indicators Module

This module presents key relationships and performance indicators in the data supplied and in the data after they have been prepared for analysis and modelling. It only uses empirical methods (empirical methods do not use models and therefore do not make any assumptions about the data whereas parametric methods use models with their implicit assumptions and assumed or calculated parameter values). The results and conclusions from the module help guide the first stage of model development.

Asset Deterioration Curves Module - Tactical Asset Management Optimisation

PAM uses Kaplan Meier (KM) analysis to produce asset deterioration curves. It has two significant advantages over distribution-based methods for producing deterioration curves.

• It is non-parametric and therefore does not make any assumptions about the data, for example that they follow a particular distribution (for example Weibull).
• The effects of different types and levels of factors, for example different types and levels of maintenance, on asset deterioration can be studied.

Failure distributions, for example Weibull, model the risk of asset failure as a function of time, for example asset age. KM analysis of failure data shows that maintenance and failure factors that include measures of asset age are much better predictors of the risk of failure than asset age alone. Therefore, when failure data are modelled using failure distributions and the assets have different maintenance and failure histories, the results have large errors. KM analysis overcomes this problem by analysing each factor individually so providing insight of how each factor affects the risk of failure.

Asset Survival Model Module

PAM uses the Cox proportional hazards model as the predictive asset survival model. It shows how the assets’ maintenance and failure history and other data contribute to the risk of failure. It is useful to note that the model is the same model as that used to analyse and model Covid-19 infection and mortality data. The analogy between human condition (survival/death) and treatment, and asset condition (working/failure) and maintenance is clear.

Predicted Maintenance Interventions Module - Operational Asset Management Optimisation

This module uses the asset survival model to optimise the performance of each asset at the operational level by identifying assets at greatest risk of imminent failure and therefore in greatest need of preventive maintenance to reduce their risks of failure. Thus, the module changes the asset management policy from reactive fail-and-fix to proactive predict-and-prevent and minimises the cost of operational asset management.

Figure 4 shows the modelled cumulative hazard from a recent project when the pumps are ranked in descending order of cumulative hazard (cumulative hazard is the hazard (risk of failure) accumulated up to time . It can be thought of as the expected number of failures if failures were repeatable.)

Risk of Failure Metrics

The basic measure of the risk of asset failure is quantified as a probability. It can be enhanced by the maintenance and replacement cost of the assets, the criticality of the assets or the consequence costs of failure. The enhanced measures do not result in the same asset order when the assets are ranked by the enhanced measures. For example, an asset with a low risk of failure and high consequence score may have a higher risk score than when the assets’ risks of failure are weighted by their maintenance and replacement costs. Graphs similar to Figure 4 can be plotted for each enhanced risk measure.

Asset Survival Simulations Module - Strategic Asset Management Optimisation

This module applies simulation and queueing theory to the asset survival model to optimise asset management at the strategic level with respect to the total asset management cost or asset availability subject to operational constraints, for example the organisation’s maintenance capacity and risk tolerance. The output of the module is the risk (probability) of failure of each asset at each simulation time, which can then be used to calculate the enhanced risk measures described above.

The input data are the number of failures in the previous 12 months an asset can have before it is disposed of, and the maximum survival probabilities for reactive, corrective and preventive maintenance. Table 3 shows example values for four levels of risk (it is possible to have more levels). When the survival probability of an asset first falls below the maximum survival probability of a maintenance type, the asset is deemed to require that level of maintenance.

Asset Availability

A key performance metric is asset availability calculated at each simulation time. It is the number of assets in use at time as a percentage of the number of assets in use at the start of the simulation. At each time it considers the number of assets that are out of service (because they failed earlier in the simulation and have not yet been repaired) and the number of assets that have been disposed of.

Figure 5 shows how asset availability after 5 years depends on the monthly preventive maintenance capacity and Figure 6 shows the profile of the cumulative asset maintenance cost up to 5 years. The monthly preventive, corrective and reactive maintenance capacities in Figure 6 are the same (80/month - they are set independently and so can have different values). The figures do not consider the costs of replacement assets and the consequent costs of asset failure because they were not available.

Base Case Simulation

The base case simulation is used when assets do not have any maintenance and so are allowed to fail. The only parameter required for this simulation is the survival probability that defines when assets are deemed to fail. When the survival probability of an asset falls below this probability, it is taken out of service so reducing asset availability. Figure 7 shows the asset availability after five years and Figure 8 shows the cumulative reactive maintenance cost up to 5 years for different probabilities that define failure.

Asset Maintenance Route Optimisation

The system also has route optimisation functionality for efficient route planning for sites with high risks of failure and therefore in need of immediate attention. It uses the calculated risks of failure of the assets and their locations to calculate the route for visiting sites with the highest risks of failure that minimises the total travel cost and time. The optimisation then identifies nearby sites with lower but similar risks of failure which should be visited at the same time. Figure 9 shows an example of the visualisation for route optimisation.

Conclusion

This article has shown how data science and visualisation can help water companies and other asset-rich companies gain insight and understanding into the performance of their assets and so improve the operational performance, safety and reliability of the assets. These benefits will manifest themselves as improved environmental and financial performance of the organisation.

Dr Atai Winkler, Principal Consultant, PAM Analytics

w: pamanalytics.com

e: This e-mail address is being protected from spambots. You need JavaScript enabled to view it

m: +44 (0)7817 263016