Machine Learning Adaption in Predictive Maintenance Applications

The continuous advancement in research and development has evolved new intelligent solutions for decision making, especially with the growing capabilities of data collection mechanisms. This advancement enabled various industries to adapt new decision-making techniques, such as time segmentation, maintenance administration, and performance enhancement. Machine Learning (ML) algorithms have a tangible impact on decision-making techniques, along with the fast growth of cloud integrated solutions, and hardware solutions. Moreover, adapting robust management systems for maintenance work can decrease the unpredicted costs during equipment failures and shutdown periods. This article explores the development process of ML solutions in Predictive Maintenance and also highlights the major advantages and challenges of adapting ML solutions in industrial sectors.

Maintenance Management Methods 

In industrial sectors there are three main methods used in managing maintenance work. They include:

• Run-to-failure (R2F): This is the most traditional method where the maintenance work occurs upon the failure detection of processing equipment or instrument devices. R2F is considered the simplest method when managing maintenance work. This method is also the least efficient since the cost impact and shutdown period could increase negatively in comparison with planned maintenance work.
• Preventive maintenance (PvM): This method manages the maintenance work based on a planned schedule instead of acting upon failure. Although unpredicted failures could be eliminated by PvM, unneeded corrective actions are usually applied, which could result in ineffective utilization of resources along with cost escalation.
• Predictive Maintenance (PdM): This is the most sufficient method where maintenance work occurs according to a continuous monitoring using a healthiness check for processing equipment or instrumentation. PdM enables the maintenance team to have an advance prediction of failures and allows the team to take corrective action ahead of time.   

Machine Learning and Predictive Maintenance

ML adaption in PdM can mitigate several challenges associated with maintenance activities, especially for unpredicted failures. Thus it is worth exploring this kind of integration to optimize maintenance work and avoid severe consequences during unplanned downtime periods. The integration between ML and PdM falls under two classes:  Supervised and Unsupervised.

The Supervised class means there is available information in the system database for failure prediction, while the Unsupervised class there is no available information about maintenance requirements. The system database contains high-level information on processing equipment, and the system utilizes grouping and co-relation techniques to identify the characteristic groupings within the processing data and then predicts ways to understand it. The selection between both classes depends on the nature of the maintenance policy. As an example, the Supervised class can fit for particular applications where failure events can be predicted between two maintenance cycles, otherwise an alternative approach can be adapted, like the Unsupervised class.  

What is the driver behind PdM in ML?

Many industrial sectors are moving toward Industry Revolution (IR) 4.0 and the Internet of Things (IoT) is considered as the most important driver of IR 4.0. IoT enables the data exchange capability between different machines, and these techniques were not explored very-well in the past. IoT enabled the adaption of smart sensors, for which processing data can be viewed and analyzed in a more efficient way.   

Data collection provides a smart gateway for future prediction related to operating equipment healthiness and or instrument devices, and it is one of the essential functionalities for PdM. The ML algorithm builds a training model, based on historical information, and then develops a healthiness prediction for the machine, such as the likelihood of machine failure.  

How can ML be implemented?

PdM utilizes a data streaming mechanism from machine instrument devices (pressure, temperature, etc.) to determine the up-normal condition in machine behavior and then predict the possibility of defectiveness during a specific timeframe. ML modeling can be built according to the following phases: 

Data Collection

The first phase begins with data collection from the possible failing parts within the operating machine (such as bearings, rotors, etc.) by using smart sensors. The overall process could achieve better results with the help of a data set, which demonstrates the machine condition and behavior during its lifecycle and captures the potential failures. This approach can help data scientists in developing PdM models.       

Data Analysis

To achieve higher accuracy and better representation of data prediction, data streaming process is combined with machine processing settings, such as set points, configuration, and historical data. These details can be gathered from different sources, e.g., the enterprise management system.    

Data Modeling

Data streaming goes into a thorough analysis to determine dependencies, and also perform technical propositions associated with the possible indications of failure and create certain behaviors for the anticipated failure.

Data Prediction

Data modeling provides a primarily conception used to detect failures, as well as to build ML algorithms as the basis for predictive models. Data prediction contains various steps for evaluating failure detection accuracy before granting final approval for the prediction models.

What are the main advantages of Adapting ML in PdM?

Adapting ML solutions in PdM can provide significant benefits for industrial sectors, specifically oil and gas companies, which include the following:

• Enhancing operating equipment’s reliability and reducing cost expenses

ML solutions provide oil and gas companies smart tools that are used for estimating the potential breakdowns before occurring. This feature allows companies to prepare an effective plan for maintenance work and prioritize their focus on critical equipment, based on a high possibility of failure. As a result, companies can reduce maintenance expenses where unnecessary work can be eliminated by prioritizing planned maintenance based on a failure forecast. 

• Improving operating equipment efficiency

ML solutions maximize the utilization of operating equipment and enhance the operating unit productivity through ensuring more reliable and flexible operation. Data modeling gives a deep indication for several parts of operating equipment, which helps companies in maintaining the production performance and establishes a mechanism in improving the operating equipment lifespan. 

• Reducing environmental impact

ML solutions can reduce the environmental impact associated with leak detection including oil and gas. Adapting such solutions help companies to detect the potential leaks in petrochemical pipelines before occurring.

What are the main challenges?

Even though many industrial sectors, like oil and gas companies, are considered the most capable and successful players in applying ML solutions in PdM, there are several challenges and limitations that could prevent the implementation:

• Applying ML solutions to legacy operating equipment

Industrial sectors are using control systems, like Distributed Control System (DCS) and Supervisory Control and Data Acquisition (SCADA), for many years. The majority of legacy operating equipment are mostly linked to internal network infrastructure and not linked to TCP/IP networks. Even though there are some techniques that can be used to establish connectivity between old and new communication protocols, the full integration becomes a major challenge for industrial sectors. 

• Gathering an adequate volume of data streaming

To ensure an accurate and reliable estimate, it’s required to install smart sensors to collect the needed information about the triggered operating equipment to modeling the failure detection feature. Collecting a sufficient amount of information could take long period, and there is a possibility for a delay during the implantation lifecycle. 

• Anticipating disruption with weak network coverage

Industrial sectors need to maintain a strong network coverage specially for operating equipment or units located in remote areas where the network coverage could be disrupted, and this affects the reliability as well as accuracy of data streaming process. Network disruption could deteriorate the overall performance of ML algorithms and could make misleading assumptions associated with failure prediction.   

About the authors: Anwar R. Al-Odail is a Control Systems and Automation specialist at Saudi Aramco, and he holds a BS and ME degree in Systems & Control Engineering. Fahad A Al-Amer is the Control Systems and Automation Group Leader at Saudi Aramco.