Data & Analytics

Machine Learning Is Not Enough

An integrated advanced-analytics program is necessary to improve the performance and reliability of pumping systems.

machinelearning-hero.jpg
Combining site data with advanced analytics and expert technical support optimizes maintenance response efforts. Maintenance intervention, therefore, can be planned, minimizing disruption and costs.
Credit: Sulzer.

Incorporating machine learning or artificial intelligence (AI) in the monitoring of industry assets is becoming increasingly common, especially in industries where downtime causes profit loss. Statistical modeling alone, however, can be susceptible to producing false positives, cultural resistance, and ineffective actions. Building a more-robust and -flexible solution that can validate real-world anomalies and adapt to changing information-technology policies requires an end-to-end approach that combines technology knowhow, equipment expertise, and operating experience.

Pumping systems, especially those classified as “high energy,” deliver raw materials, such as oil, gas, and water, to processing plants before they are transformed into products that are used in everyday life. Analyzing asset health to optimize operation and maintenance actions provides the competitive advantage in production and transportation and, thus, the motivation for introducing machine learning.

Measuring Success

Advanced analytics solutions that aim to improve business performance can be deemed successful only if they deliver a measurable benefit to safety, the environment, or business finances. From the outset, the solution needs to be flexible and end-to-end capable in order to deliver the business’ objectives.

As such, these digital solutions are a value enabler, giving operations the opportunity to take the best action based on the provided intelligence. Pumping systems can vary greatly by design and cost. Running high-energy pumps, or those with high-risk fluids, to the point of failure can easily have implications far greater than $100,000 per occurrence.

Using a digital solution, however, to accurately predict prefailure conditions of pump-system components and then performing remedial actions is far from a simple process. In fact, creating a solution that achieves the corporation’s goals and offers the flexibility to adapt to future changes requires considerable system-design expertise and historical equipment knowledge. This is best explained using an actual installation where the operator was searching for an efficient asset-management solution.

Case Study: Setting the Criteria

A large solar-power installation, operated by a company that has sites on three continents, wanted to improve their predictive maintenance program while meeting a project rationale and corporate data policies. An advanced-analytics solution requires data security based on individual company policies as well as the ability to change and grow over time with these policies.

In this case, the operator was already performing data analysis and anomaly detection with a data science team. Engineering firm Sulzer’s role was to augment the approach and, by working closely with them, provide analytics that complement their existing work. Sulzer outlined the capabilities and benefits of its Blue Box software and how a collaboration would enable the operator to achieve its goals.

During onboarding, the process of uploading historical data began to train the machine-learning algorithms. Once complete, an open channel for new data points was created and monitoring began. Throughout this process, weekly meetings were held to discuss all the tasks, including data quality, and any potential issues with the pumping system would be highlighted. Interaction between the groups aimed to deliver improved data quality, fewer system issues, and mutual training gains.

Data Checking

As part of the process of receiving information from the data historian, a need arises to perform a data sanity check, which essentially is an audit of the operational data, highlighting any issues that need to be remediated. An understanding of not just the data but also the equipment and application is necessary to accomplish this.

For example, one of the data points received was motor current. Equipment expertise and operating experience made clear that this was bad data. Further investigation onsite revealed that the data being sent was not amperage but percentage yield instead; the data tags on the supervisory control and data acquisition (SCADA) system were incorrect.

Data quality control must be part of the analysis process, not only during setup but also continually throughout the period of the advanced analytics program. In the case of the Blue Box software suite, data quality control is a continual task every time an anomaly is detected. An advanced-analytics program must never suggest an action based on bad intelligence.

Anomaly Detection

Mathematical models learn from historical data to identify similar patterns. In this application, however, pump failures do not occur very often and most operators do not have the history of operation and maintenance readily available in a digitized format. For that reason, Sulzer leverages unsupervised machine-learning techniques, where the models are trained with the recent operational history of the pumps together with physical pump modeling.

Returning to the example, not long after the SCADA issue was resolved, Blue Box flagged four anomalies on a single pump over a couple of days, where the performance of the asset deviated from the healthy state. The onsite data-science team was also running, in the background, its own algorithms, and further investigation found a similar abnormal event for the same period.

By combining the two independent sets of results, validating the information and determining the presence of a prefailure condition within the motor was possible. Analysis of the motor power and shaft speed had identified a bearing that was in prefailure condition yet far below the alarm limits for vibration and temperature.

Acting on Information Received

Without an in-depth understanding of pumping systems and the lack of any other alarms, some operators may choose not to follow the decision support delivered from a predictive-maintenance system. The benefits of these systems are achieved only if the operator trusts the information being delivered.

This is the point, for Sulzer, where the equipment optimization specialists (EOS) come to the fore. Building on Blue Box as a common basis for engaging with the operator, these experts address a need for the technical ability and experience to complement the insights—such as on a detected anomaly—offering expert advice to the operator and enabling an action plan to be formulated. Collaboration between site personnel and the EOS allows the options to be defined, along with the required actions, timings, and benefits. From there, the information can be delivered to the decision makers, with supporting evidence, to allow them to make a well-informed choice.

A solution turns anomalies into insights and insights into actions. Digital solutions deliver decision support; however, no solution is of value without connecting to action. Ultimately, machine learning on its own has limited benefits; several other aspects must be combined to deliver an end-to-end solution.

Data quality must have a continual reality check. An expert understanding of how the physical equipment design is represented in a digital format is necessary to protect from false positives and react to true intelligence.

Data security is paramount; without meeting the corporate security policy, the project will never take off. Any changes to this policy must be accommodated, so the analytics solution needs to be flexible.

Sulzer’s Blue Box is a good example of how machine learning can be combined with extensive pump knowledge that comes from being an original equipment manufacturer to enable actions to be planned, keeping costs to a minimum and avoiding downtime wherever possible.