The Path Forward with Artificial Intelligence in Enterprise Maintenance

Artificial Intelligence is a trending technological topic that has been steadily on the rise for the last two years.  Numerous companies, including some of the tech giants are heavily investing in AI. According to McKinsey & Co., Baidu and Google spent between $20bn and $30bn on AI in 2016, for both R&D activities and acquisitions of AI-related companies.

This surge of interest on AI is largely due to recent advances in this field. In particular, in March 2016 and May 2017, Google’s AlphaGo AI engine beat two of the most prominent grandmasters in the centuries-old Chinese game, Go.

While this may sound similar to earlier achievements (e.g., In 1997, when IBM’s Deep Blue computer bested chess champion Garry Kasparov), it is considered a breakthrough; AlphaGo employed human-like reasoning over a few optimal moves based on deep learning, instead of an exhaustive analysis of all possible moves as in Deep Blue’s case.

This breakthrough has been empowered by advances in Big Data and neural networks, as well as by the evolution of computers’ computational abilities (as per the famous Moore’s law). As a result, there are increasing possibilities to apply AI in various domains. Industries such as utilities, oil and gas, and manufacturing are looking to leverage the power of the emerging “deep learning paradigm”, towards intelligent reasoning and greater automation.

In this context, we must consider the application of new advances in AI in the field of industrial maintenance, as evident in recent due diligence reports and the emergence of AI-based enterprises and products for the Industrial internet.

The evolution of Artificial Intelligence in enterprise maintenance

In principle, AI refers to machines that can act like humans in a given context, through:

• Perceiving the environment;
• Learning how to perform tasks;
• Reasoning over their context; and
• Making optimal decisions.

Based on this definition, AI-based systems have been around for several decades and have been applied in maintenance, repair and field service engineering operations. In fact, we can break down the history of AI-based systems in maintenance into three phases:

Phase 1: Single method systems

This early phase involved the individual deployment of different forms of AI-based systems such as fuzzy logic systems, expert systems and more.

For example, fuzzy logic systems have been used to prioritize different maintenance approaches and policies. The merit of such systems stems from their ability to evaluate alternative policies against multiple criteria and indicators such as number of failures, life of specific components and Overall Equipment Efficiency (OEE).

Similarly, expert systems have been used to assist professionals in the inspection and maintenance of offshore structures. They provide quantitative and qualitative knowledge about defects in the structures, along with relevant recommendations for dealing with these defects. Expert systems complement the knowledge of the professional with a pool of databases and reasoning, which help users assess the severity of the defects and decide the best way to resolve them.

Phase 2: Multi-model systems

This phase of applying AI in industrial maintenance involved the development of maintenance systems that combined more than one AI technique in an integrated system. For example, several research works have demonstrated that AI techniques can be combined with Operations Research (OR) and constraint programming techniques to increase the credibility of maintenance systems.

During this phase, the scope of AI-based maintenance projects expanded beyond failure modes detection and assessment to maintenance projects budgeting and selecting optimal maintenance methods.

Phase 3: Today’s Deep Learning era

In recent years, AI systems have begun to employ deep learning as a means of leveraging the vast amounts of maintenance data that are generated by wireless sensors networks and internet of things deployments. Deep learning is based on complex neural networks, which mimic the operation of the human brain. Common maintenance use cases based on deep learning include:

• Predicting and avoiding machines failures. According to McKinsey, this could increase asset productivity up to 20% and reduce maintenance costs up to 10%.
• Using robots as digital companions of human workers. For instance, aircraft maintenance innovator MRO.AIR is already developing such a product for aircraft maintenance engineers.

While these systems employ deep learning, they also combine additional machine learning techniques and modalities, which optimizes results.

Today’s AI systems can be classified into two categories:

1. Weak AI systems: Are machine/deep learning systems that require prior training based on appropriate data sets.
2. Strong AI systems: Can identify new situations without prior training.

Currently, most AI systems fall in the realm of weak AI, which highlights the importance of proper datasets. In the future, we expect that stronger AI systems will emerge to cope with more complex and unpredictable situations.

Adopting Artificial Intelligence successfully

The deployment of deep learning, industrial robots and other forms of state-of-the-art AI in maintenance tasks is promising, though still in its infancy. AI is not a panacea and it won’t resolve all our maintenance challenges any time soon.

Nevertheless, maintenance enterprises cannot afford to ignore AI. Artificial Intelligence is disrupting industrial processes and is one of the main technological trends of the digitization of industry.

Rather, enterprises need to take careful steps to adopt AI, by maintaining a balance between their practical business needs and the pace of innovation.

These recommendations can help maintenance enterprises adopt AI successfully:

• Put business objectives first: The AI adoption journey should be driven by pragmatic business objectives and tangible maintenance needs. If not, any AI deployment is bound to fail. It’s therefore important that you accompany any AI idea with thorough cost-benefit considerations and a justification of its ROI (Return-On-Investment). Likewise, use cases should be prioritized based on their business value.
  

• Leverage maintenance data: In most cases machine learning for AI (including deep learning) is not possible without the proper datasets. Therefore, enterprises should make sure that they collect, process and consolidate the datasets that will empower their AI use cases.
  
  Data management can be a significant part of an AI deployment for industrial maintenance, given that maintenance data tend to be fragmented, disaggregated and non-interoperable, as they reside in different platforms like asset management, condition monitoring and ERP systems and databases.
  

• Assemble the right team: No AI system can be effective without the proper domain knowledge. Standard algorithms for detecting and assessing maintenance patterns or even failure modes will render sub-optimal results.
  
  It’s the domain knowledge from field experts that will let AI specialists tune their algorithms and optimize them for the problem at hand, while avoiding common AI pitfalls such as “overfitting” algorithms to training datasets.To ensure that domain knowledge is available, an enterprise needs to assemble a multi-disciplinary team that will comprise both AI experts and maintenance experts. The formation of such a team is challenging, given the talent gap in maintenance, AI and Industry 4.0 technologies.
  

• Create new maintenance processes: The successful deployment of AI in maintenance is primarily about identifying and deploying the right process, rather than solely implementing an advanced IT system.
  
  In most cases, AI will not be operating in isolation: employees need to understand where, when, and how AI will be deployed and what elements they are responsible for.
  

• Create new maintenance processes: The successful deployment of AI in maintenance is primarily about identifying and deploying the right process, rather than solely implementing an advanced IT system.
  
  To this end, adopters should first define small pilots to prove the business value of the selected maintenance use cases (e.g., predictive maintenance, use of robots). Following the successful conclusion of pilots, the pilot systems can move to production and scale.
  
  For example, a pilot maintenance deployment could involve maintenance a specific part or machine. As part of the scale up process, a successful deployment can be replicated to more parts or machines of the same type..
  
  Testing and simulation are also useful tools that can be used during the pilot stage to evaluate the solution without significant upfront investments. Overall, a staged approach to AI deployment reduces risks and minimizes the adverse effects of potential failures.
  

• Stay agile: There is no “one-size-fits” all AI solution and no silver bullet for success. AI deployers may need to test, validate, deploy and redeploy a host of different algorithms to identify failure modes or achieve human-like reasoning for the maintenance task at hand. Therefore, they’ll need to be agile and responsive to feedback from the field.

Artificial Intelligence isn’t new, but recent exponential advances mean that enterprise maintenance organizations and those in other related industries must pay attention. It’s time for them to explore which parts of their maintenance processes could benefit from AI and its human-like reasoning capabilities. Combining machine intelligence with human intelligence will increase automation and asset productivity – it’s only a matter of time.