Johannesburg, 01 Jun 2020
If you own a business that relies heavily on machinery, and you’re not doing any predictive maintenance yet, the chances are that you’re suffering from downtime losses even while reading this.
Carington Phahlamohlaka, Data Scientist at Altron Bytes Managed Solutions, says: “It’s a real challenge to predict when you need to service your equipment and it’s difficult to weigh the risks of lost productive time against those of a potential breakdown.”
This challenge is traditionally addressed in one of two ways: either reactively, by fixing the already existing failures, or proactively, where past experience is used to anticipate potential breakdowns. Unfortunately, neither of these approaches is effective enough.
If you don’t predict precisely when a machine or a piece of equipment is going to break down, the resulting downtime may be longer than anticipated, as you don’t only need to replace a failed part, but you may also need to order it and ship it, and sometimes from overseas. That results in reduced productivity and downtime costs will naturally soar in terms of penalties or loss in profit, not to mention fault assessment errors.
Organisations often face the challenge of ensuring maximum availability of operating critical physical systems, while simultaneously minimising the cost of maintenance and repairs. The early identification of potential issues helps organisations deploy limited resources more cost-effectively and maximise their equipment uptime.
Phahlamohlaka says: “The decisions are now data-driven, whereas historically they relied on people’s judgment. If you consider that one business could have 5 000 devices for one client, there simply isn’t the manpower to go through those devices and analyse their various components for potential failures. This is where complexity comes into the equation.”
Essentially, if all devices operate under controlled circumstances, then each device component converges towards the average behaviour across the base. However, in uncontrolled conditions such as different locations, weather conditions, human interactions, etc, each machine emulates a unique behavioural pattern per component. Owing to these uncontrolled conditions, a business waits for an alarm signifying that a device has failed, then logs a fault and a technician is dispatched.
“Data science allows for the collection and processing of device dependent usage statistics in order to determine how busy a device has been, depending on its location and the time of day, over a selected period. This allows for the identification of behavioural patterns per device, down to component level, from which component lifespan can be inferred.”
He cites the example of a card machine. “If you know that a specific model will perform a certain number of card reads before a certain component needs replacing, but the retailer is in a busy area and it’s a busy time of year, such as December, we can plan around that. The service and maintenance data for all of those devices in the field is available, and can be used to identify the optimum moment to replace a particular component before it breaks without doing so too early.”
The complexity of data science lies in interpreting data that is very much machine dependent, together with historic maintenance and service intervals, to recommend the niche area between maximisation of uptime, and minimisation of false positives and true negatives. The intelligence that is brought in considers the usage statistics, how the machine has been used and how individual components in the machine have been used. An intelligent algorithm is used to mimic that pattern and predict when certain events will occur, such as when a component will break down. The predictive actions will drive improved service delivery, reduced costs and improved efficiencies.
Artificial intelligence, as a subset of computational intelligence, is the umbrella within which machine learning resides, which is a trendy term around predictive work. Hence, predictive maintenance employs algorithms from this space. Among these algorithms are biologically inspired, population-inspired, and brain-inspired algorithms. “We use biologically and population-inspired algorithms to associate devices so that those with less statistics can be improved by similar devices, by characteristics, with more statistical data. Furthermore, we use brain-inspired algorithms in collaboration with other numerical methods to build predictive maintenance schedules, following a cognitive process.”
He explains, these algorithms emulate what a human being would do, if they had the time, they just do it faster and with more accuracy. “A human would be able to figure out the issue eventually, given sufficient time, but there are so many devices, each of which has multiple components, which might have sub modules, making it virtually impossible for a human to analyse everything. The algorithm is able to emulate the human thinking process, hence cognitively.
According to a US study, about 70% of projects in data science fail primarily because of a lack of understanding, a shortage of the right kinds of skills and people being employed in the wrong positions. If you understand AI properly, it becomes clear that it can never replace a human. Instead, it augments the human and makes their job easier. AI is like a cellphone: you need it, it makes things easier, but it cannot possibly replace you.”
Share