Turning the Dream into Reality
Published by : Industrial Automation
Philipp H F Wallner dwells upon How to realise the dream of implementing successful predictive maintenance applications.
Predictive maintenance has been heralded as a solution to manufacturers’ and engineers’ woes thanks to its potential to give equipment users the ability to anticipate imminent malfunctions, proactively arrange repairs, reduce disruption to operations on the factory floor, and, most crucially, safeguard against the failure of equipment, which can have a disastrous impact on the whole business. Based on these use cases it is understandable why this technology is perceived to have huge value. According to Deloitte, predictive maintenance can reduce overall maintenance costs by between 5% and 10%.
But the key word here is perceived value. While predictive maintenance has the capability to be truly transformative, when the time comes to implement the technology on existing equipment the process is not always straightforward and this fact is reflected in the actual number of businesses that have implemented predictive maintenance in operation, which is very few. So why is the rate of adoption so slow? Industry commentators have highlighted four factors that seem to be the key stumbling blocks that need to be conquered by equipment builders and operators to effectively work alongside the data science community and realise the dream of successful predictive maintenance solutions. These factors include:
1. Encouraging teamwork and knowledge sharing to benefit from existing domain know- how during the algorithm design process: It can be difficult for businesses to cultivate a collaborative environment where powerful algorithms for predictive maintenance based on statistics methods are designed that integrate the domain knowledge and expertise of both data scientists and domain experts. Furthermore, how can domain experts and data scientists work together to make sure that the key elements of each effective predictive maintenance application are fully leveraged? How can they be sure to include both data analytics methods and domain knowledge?
The best predictive maintenance applications include both of these components:
statistics-based data analytics methods, like machine learning, in addition to the domain expertise about equipment that the R&D engineers possess (very often already incorporated into existing simulation models). If predictive maintenance is approached with a singular data analytics mindset, users will not capture all of the useful information retained by the operations and engineering teams that build the equipment and are responsible for their ongoing upkeep.
2. Determining how to train algorithms without access to sufficient failure data: Training an algorithm on data from the field is a fundamental part of machine learning. Those creating the algorithm must include ‘good’ data from everyday production on top of a variety of failure data taken from the numerous error scenarios that can happen while the equipment is operating. However, if the goal is to never allow the equipment to break in
the first place, where can the failure data be obtained? This is turning out to be a progressively important problem to solve for businesses utilising predictive maintenance for their industrial systems. What’s more, it is irrespective of use cases and can range from air compressors to wind turbines. To overcome this issue, simulation models can be brought in to produce artificial failure data, so the algorithms have something to be trained on when there isn’t any, or not enough, measured failure data from the factory floor.
3. Taking the algorithms from the design stage to real-world operation: After the training and design of the predictive maintenance algorithm has been carried out on the desktop, the next step is deployment onto the equipment. The difficulty level of this process directly correlates with the condition of the existing IT and OT infrastructure. Whereas some algorithms are applied on a real-time hardware platform – for example, on an
industrial PC, an embedded controller, or a PLC – there will be some that are in the cloud or will be merged with the current non-real-time infrastructure (for example, an edge device running on Linux or Windows). At a growing rate, businesses are taking the option of implementing an efficient way of using toolchains that facilitate automatic generation of C, C++ or IEC 61131-3 code, .NET components, or standalone executables.
For example, a manufacturer of packaging and paper products installed predictive maintenance software into its manufacturing line as a way to lower the amount of waste produced and reduce machine downtime in its plastics manufacturing facilities. (See tint panel for details)
4. Proving the potential return on investment (RoI) of predictive maintenance solutions: When any organisation kicks off a predicative maintenance project, the most important question it has to be able to answer at the outset is, how can I prove the RoI of this investment? Otherwise, it will struggle to get that initial budget going.
In the absence of an answer to this question, all the time and energy spent on developing a detailed predictive maintenance plan and solution will quickly run aground. Identifying a concrete business case and developing an approach for how to monetise predictive maintenance will prove vital when trying to persuade your corporation’s management team to rationalise the investment in executing a predictive maintenance project.
The most obvious quantifiable benefit for equipment operators will be the reduction in equipment failure during operation. While this often justifies the investment for operators, for equipment builders, building a case is more difficult. However, there are a number of ideas that have been proposed by some of our clients, which have significantly contributed to building a solid case for implementing predictive maintenance. They can be summarised as:
1. Linking service fees to predictive maintenance of the equipment used by the operators— equipment builders’ customers.
2. Taking advantage of IP protection to sell the deployed predictive maintenance algorithm itself.
3. Moving to a new business model based on usage (for example, selling elevator usage hours rather than entire elevators, or cubic meters of compressed air rather than compressors). It will be only a matter of time before the C-suite – armed with the information of these possibilities – jumps on board with realising predictive maintenance in all its glory.
Philipp H F Wallner, Industry Manager, Industrial Automation & Machinery, MathWorks, and is responsible for driving the business development of this industry segment that comprises energy production, automation components, and production machines. Prior to joining MathWorks, Philipp worked in the machine builder industry, where he held different engineering and management positions. He has a MS in electrical engineering from Graz University of Technology and an executive MBA in project and process management from Salzburg Management and Business School.