Machine Learning shapes how we live. It is continuously making an impact on a wide range of applications. What the machines have in plenty is data – data on how to squeeze the best out of those machines, how often to repair or replace which parts and many such minute details.
The cost of unscheduled downtime
As per industry research, reducing unscheduled downtime of a critical equipment by 1% itself would translate to more than 1 million dollars of annual savings. However, every industry, no matter how small or big, has specific preventive processes to avoid such situations. Again, all such measures being dependent on either humans or numerical thresholds are often proving to be inadequate in terms of analytical control over cost, quality, and productivity. Hence, there is a long-felt need for proactively resolving challenges in capturing device level data, delayed access to data, and insight derivations from data.
Predictive analytics for asset maintenance
Using a wide variety of techniques, such as, data mining, machine learning, statistical algorithms, predictive modelling, predictive analytics, conclusively forecasts future occurrences by evaluating both, historical and current statistics. The three fundamental components of every predictive analytics application are data quality and quantity of the historical data analysed, the numerous statistical techniques varying in functional complexity used to derive meaning and insights, and assumptions or conclusions that are drawn for future pattern modelling.
As the 4th industrial revolution transforms global industries with IIoT, there are an infinite number of such predictable insights to be extracted from voluminous amounts of structured or unstructured data generated by sensors at every pulse point of the connected factory. It is due to this evolution of predictive analytics into predictive maintenance that alongside tech giants, even traditional software companies are opting for predictive and prescriptive analytics, over their classically descriptive and diagnostic analytics, as revealed by Gartner in their 2018 Magic Quadrant for Data Science and Machine Learning Platforms report. Apart from reducing repair costs and warranty costs by preempting root causes of failure, predictive maintenance also leverages non-intrusive testing techniques, such as, thermodynamics, acoustics, vibration and infrared analysis to optimise and maintain asset health analytically. Even McKinsey’s research asserts that predictive maintenance complemented by Artificial Intelligence (AI) facilitates improved machine failure forecasts by analysing data from the advanced IoT sensors, maintenance logs, and external sources to increase asset productivity by up to 20%, and reduce overall maintenance costs by up to 10%.
How does the improvement happen?
Initially, the data residing in silos at multiple edge device sensors are analysed with historical alarm data and existing multivariate time-series data models. With real-time data acquisition, the mined data is analysed for hidden patterns using unsupervised feature learning and labelled as per known conditions, with semi-supervised machine learning. Furthermore, these models are evaluated and deployed to implement decisions and make predictions with a degree of accuracy that always has a scope for improvement. However, predictive analytics on its own can only share insights. It’s only by coupling with Machine Learning (ML) that predictive modelling enhances into insights and helps in decision making.
ML is a sub-section of computer science that enables computers to learn without being explicitly programmed. Evolving from the study of pattern recognition, ML explores the belief that algorithms can learn from and make predictions on data. With increasing intelligence of these algorithms, they don’t need rules or simple threshold settings and are capable of identifying abnormal and correlated sensor data patterns through behavioral analysis to accurately identify machine degradation or forecast faults before they occur.
The implementation challenges
PwC reports reveal that in the next 5 years, manufacturers are going to be 38% more likely to adopt machine learning and analytics for improving predictive maintenance, while also increasing the implementation of process visualisation and automation by 34%. Therefore, it is the manufacturing and MRO (Maintenance, Repair & Operations) division of the organisation that stands to reap the maximum benefits of predictive maintenance. However, these divisions do not have the right incentive to capitalise on predictive maintenance or are looking for Commercial Off-the-Shelf (COTS) solution, likely to be a fictional silver-bullet for their predictive maintenance problems. Being the gatekeepers of internal data, they are also worried about data security, when the data is taken to the cloud.
The future of predictive analytics
Predictive analytics is also further evolving into prescriptive analytics, wherein proactive decisions are made without human interaction. Prescriptive analytics is the next leap in AI use cases, and here, the output from ML and deep learning techniques are used to predict failures and then initiate proactive actions. Riding the tide of all such developments, the predictive maintenance market is expected to be around US$5 billion by 2021.
Therefore, to successfully leverage these leading-edge tech solutions for reducing equipment downtime, all that manufacturing CXOs need to do is smarten their manufacturing machinery with sensors and build internal analytics and AI capabilities. However, if they are constrained regarding time to market, cost and availability of resources, it’s worthwhile to collaborate with engineering service enablers who have a proven track record of engineering the frontiers of excellence in analytics and AI.