How Predictive Maintenance Can Quietly Transform Factory Operations

Predictive maintenance is one of the most funded uses of AI across all heavy industry sectors, from transportation to manufacturing and beyond. This is due to both its potential to improve budgeting and strategizing and reduce costs by providing an overview of the machinery that needs to be replaced.

Predictive maintenance in many factories is usually performed by a quality engineer who develops a set schedule for conducting maintenance on specific machines such as galvanizing furnaces, welders, or tube cutters in the case of a steel factory. However, these equipment checklists don’t always catch sources of error or wear and tear on the equipment. Furthermore, if a quality engineer conducts their maintenance purely on the basis of a time schedule, then this doesn't account for a time with elevated demand that may put a strain on a particular machine.

Predictive maintenance through machine learning enables factories to take a step further. Using a technique such as “random forests”, for example, which is a predictive model that consists of a large number of individual decision trees operating at once, factories can plan maintenance strategically instead of arbitrarily setting a schedule. The principle is that a large number of relatively uncorrelated models (trees), operating as an ensemble, can outperform any individual constituent models.

Predictive maintenance makes it easier to identify when a piece of equipment is failing systematically, significantly reducing the risk of scrap or a low-quality part going to a customer – both of which are far more damaging than the cost of replacing the equipment. The issue for many manufacturers is that when a piece of equipment is shut down for maintenance, the in-process inventory can become backed up, creating an adverse effect on other equipment and the workforce operating it. Furthermore, meeting customer delivery expectations without fully operating machinery becomes challenging.

Capital equipment is not only expensive, but it also takes a long time to assemble and be shipped. An industrial 3D printer could cost anywhere from $50,000 to $500,000 to replace, and a furnace could cost close to $1 million. So, predictive maintenance can ensure that a line doesn’t have to be shut down for three or four weeks while that piece of machinery gets delivered. Naturally, the monetary savings are critical, but the planning which predictive maintenance permits is possibly its most valuable aspect. A model which notifies you months in advance allows budgeting to take this information into account and make decisions in terms of other areas where expenditures are required. 

Predictive maintenance, as a concept, highlights how maintenance is not conducted as often as we’d like. It requires knowing the success criteria for a particular piece of equipment, which could be a category such as ‘the percentage of parts produced within an allowed variance’ (dimensional accuracy), and how this can change as the equipment wears over time. Typically, this would take between one to three months, during which time that data is used to build an algorithm using a random forest or another machine learning technique. Depending on how the accuracy changes, you can create a profile for the equipment to determine when it will wear to the point of being unusable, and set a maintenance schedule accordingly.

In a literal sense, the training data flows into the algorithm, and then the script rewrites it every month or so. The output of that script can be transmitted via an API to your ERP system, for instance, where your maintenance schedule is stored. An algorithm’s success is dependent on the amount of data you feed it – the more, the merrier. It’s vital that you use the most recent data possible because so many factors can influence a machine’s performance, making old data redundant.

Predictive maintenance intends to monitor the settings of the equipment rather than logging the outcome. However, it is possible to monitor the outcome and tie the final product quality to particular equipment settings. For example, in a steel factory, there may be a process where steel is bathed in acid to boil off any rust. If the equipment has degraded slightly, then the temperature could be marginally lower, and so rough spots may not be burned off. This tells a quality monitoring engineer that they may need to recalibrate the settings.

Factories that are operating with in-built sensors for their machinery can enable millions of data points, and therefore the effectiveness of the algorithm for machine learning is much higher. In a car-making plant, there might be a robot calibrated to exactly 60.3 degrees to deposit a piece of metal into a screw when it twists. If it were to veer from this degree, the screw could get jammed. The automaker software monitors the lubrication level of the bearing, the temperature of the coolant, and the angle of the robotic arm over an extended period of time to find a correlation between the angle of deposition and the other settings. 

The value of machine learning in monitoring settings is that it broadens the ability to determine what factors indicate when maintenance needs to occur. It's not so much that it unlocks a wholly foreign capability to humans, but instead that it can pick up on those trends a lot more quickly and reliably than a human ever could.

Predictive maintenance allows manufacturing employees and management to utilize analytics applications with intuitive dashboards. The ability to oversee machine health and its effect on end-product gives more confidence that downtimes can be avoided. 

Even with the successful addition of AI, predictive maintenance challenges remain in determining precisely what data should be used as the diagnostic data stream for a specific piece of equipment to avoid expensive and time-consuming iteration. However, those at the executive level can have peace of mind that low-quality products are unlikely to make it to customers with predictive maintenance in place. It’s not something that should require regular interaction from top management, but it will certainly make their jobs easier in the long run.