Across every industry, there is a growing trend to transition from the current paradigm of scheduled and reactive maintenance to a new model which relies on a predictive maintenance framework to improve asset life, increase visibility into asset operations, and deliver higher quality service to customers cost effectively.
From subsystem and component failures across an aircraft fleet, to faults in the distribution network of a power utility that cause widespread outages, to asset failures at on- and offshore oil production facilities or at construction or mining sites, the economic impact of unplanned downtime on assets is significant. Reduced readiness levels for military assets, reduced service reliability for utilities, lost production for oil and gas assets, and unplanned downtime for heavy construction or mining equipment operating in remote areas – these are some real-world examples where a robust predictive maintenance framework has the potential for remarkable positive impact.
The traditional solution to the maintenance problem
Classic predictive maintenance refers to online condition monitoring using rule-based systems developed by subject matter experts (SMEs). These SMEs rely on decades of experience to understand asset operations and use this knowledge to drive condition-based maintenance. Such human-based decision making mechanisms are powerful, yet simplistic. And most importantly:
- they are not codified, so the knowledge is lost with attrition of experienced talent
- they are not suited to learn from and evolve with emerging failure modes in the field, driven by new assets being deployed, or changes to the operating conditions of assets.
This truly is a big data problem. To effectively perform predictive maintenance, we need to evaluate ALL relevant data, focused not only on the monitored asset, but also on all other assets of the overall operating system that the monitored asset is operating in, as well as extraprise data that provide insights about ambient operating conditions, and relevant external parameters that affect the performance of the monitored asset. If you can solve the big data problem effectively, the system will automatically codify the knowledge of the SMEs into features that inform a self-learning predictive model that continuously improves from feedback from the field. This capability can then be used across the enterprise and transitioned to other analysts, and improve over time as more data and feedback is provided to the system.
Applying AI to Predictive Maintenance
The confluence of three important technology vectors has for the first time, enabled automation of complex decision-making processes across organizations with many silos of data to solve daunting business problems such as data-driven predictive maintenance. These vectors are the ubiquity and democratization of elastic cloud computing, the maturity of AI and machine learning algorithms such as deep learning, and the abundance of data from sensors known as the Internet of Things (IoT). Leveraging these four technology vectors, we formulate and solve this problem as an machine learning classification problem. We then use the entirety of data from enterprise and extraprise data sources, along with initial identified features, and feed them to machine learning or deep learning frameworks to train classifiers that predict failures. The resulting classifier has gone through all the historical failure data provided, and has combined features to learn optimal signatures for an impending failure. The resulting automation frees the asset operators from manual reviews of every single asset, and helps counter potential inspection biases by providing an objective maintenance recommendation based on live signal readings from assets. Combining these recommendations with effective decision-making support from AI-based solutions, we have the potential to significantly scale the knowledge of SMEs. With this approach, we can construct significantly stronger decision-making mechanisms, and apply this powerful framework in every industry sector globally.
Predictive Maintenance for all
At C3, we have not only demonstrated the feasibility of this concept, but also achieved superhuman performance by deploying such solutions at scale in production environments at large organizations, with closed-loop feedback from the field to enable continuous learning. Furthermore, we have applied these solutions to widely disparate situations. These include:
- Multiple subsystems of multiple aircraft classes using telemetry data and pilot and maintenance notes
- Electricity distribution networks and power generation assets (both conventional and renewable) with sparse failure labels
- Oil and gas assets such as compressors or pumps with high frequency measurements (>1 Hz) from more than 50 sensors
- Industrial refrigeration equipment such as chillers
- Construction and mining heavy machinery
- Healthcare application such as drug dependency and disease prediction
It is noteworthy that to solve these seemingly different problems, we use a single solution with industry/asset specific modules that build on a base set of machine learning techniques. Specific use case implementations require configuration of the features for the machine learning model and the choice of a specific classifier (e.g., XGBoost vs Deep Neural Network) to achieve desired results.
Distinction from Fault Detection
It’s important to emphasize that predictive maintenance is different from identifying an already-occurred fault (a.k.a., fault detection). Fault detection techniques learn normal sensor readings, and flag significant deviations from these norms as faults. Although C3 also provides fault detection solutions—typically using unsupervised machine learning techniques—in this post we are focused on predicting failures or faults which are expected to occur in the future.
Let’s consider an example from the oil and gas industry, where an oil well pump system starts to show signs of an impending failure. Such failures can cause up to six months of lost or reduced production before staff can fix the issue or replace the damaged part. A failure-detection capability can reduce this downtime by about two months, but still leaves behind up to four months of impaired production. We have shown with our oil and gas customers that machine learning-based predictive maintenance can significantly further reduce periods of lost production and downtime, and thereby save organizations half a million dollar per well per year, totaling hundreds of millions of dollars per year for some customers.
Superhuman Performance, Super-fast
Today, AI and machine learning have enabled us to augment or replace decades of SME capability. We can now codify SME knowledge as machine learning features, and build additional features by leveraging the elastic computing resources available from Infrastructure as a Service (IaaS) providers such as AWS and Azure. Within a few months of work, we can surpass the performance of SMEs, while in the process we preserve institutional expertise before experienced, senior SMEs retire or leave the organization. Most recently, we have shown using deep learning that we are no longer dependent on SME input on features for machine learning classifiers and we can use raw sensor readings with no need for feature engineering. This saves time and yields better performance, in a very short period of time.
By applying these concepts to an array of domains and various asset classes in each domain, C3 has already helped many organizations extract significant operational and economic value. In future blog posts, we intend to dive deep into specific use cases, describing challenges we faced and learnings from those projects.