Configuring the IIoT roadmap: Machine Learning for the digitalisation journey

 

Lower computing and IoT sensor costs are paving the way for Machine Learning-based IIoT for Maintenance. Though the pace of automating Machine Learning makes it increasingly simple to deploy AI-driven PdM at scale, organizational and technological challenges are preventing more widespread adoption.
While Industry 4.0 continues to generate analyst and media attention, many plants are still struggling with the realities of implementation. The potential ROI from increased uptime and new revenue streams are compelling, yet in most cases, there is no single roadmap for how to realize the benefits of IIoT.

This article reviews the subject of Machine Learning for Asset Maintenance and the role it plays in Industry 4.0 integration.

Falling IT infrastructure costs: An enabler of Industry 4.0

Industry 4.0 (or Industrie 4.0) started as a German government-backed initiative that supported decentralized “smart” manufacturing. Revolutions do not occur in a vacuum and in the case of Industry 4.0, there were several factors:

  1. The cost of computing power and data storage has fallen dramatically. For example, in 1964 the cost to store a terabyte of data would cost the equivalent of $3.5 billion dollars. Today, the cost to store the data is approximately $25. Data production has consequently soared. The global datasphere is projected to grow from 33 zettabytes in 2018 to 175 zettabytes in 2025. Similar trends are prevalent in cloud computing, connectivity and other IT/OT infrastructure that forms the basis of the Smart Factory.
  2. IoT sensors have become dramatically cheaper. According to a 2014 Goldman Sachs research report, the cost of IoT sensors fell from $1.30 in 2004 to 60 cents in 2014; a 2019 Microsoft study found that this price had dropped further to 44 cents per sensor by 2018. Why is this important? Sensors have been termed the “nervous system” of IIoT. At the most basic level, IIoT software requires real-time access to factory asset sensor data to provide intelligent recommendations that relate to production, asset maintenance and supply chain management.

The application of Machine Learning to asset maintenance

Today, Condition Monitoring systems use SCADA data to monitor asset performance. Manual thresholds are set based on human-made rules, and when sensor data breach thresholds an alert is triggered signalling potential machine fault. Based on bandwidth constraints, a number of critical factory assets are monitored this way.

With Machine Learning, an algorithm is trained to detect abnormal – and correlated patterns of abnormal – sensor data. Based on this behavioural analysis, the algorithm identifies machine degradation or fault before they occur. Machine Learning does not require rules or simplistic threshold setting, because it is looking at behavioural patterns. Vast amounts of data can be analysed in real time without the need for human involvement.

What is Automated Machine Learning?

Only a few years ago, research in the field of Artificial Intelligence was conducted at an academic level. In the last couple of years, there has been significant investments in Machine Learning R&D. Today, a plethora of Machine Learning and Deep Learning algorithms can be applied to Smart Factory applications.

Given vast amounts of data that are relatively inexpensive to analyse, the key today is to automatically select the optimal Machine Learning algorithm. There are dozens of Machine Learning and Deep Learning algorithms and pre-processing and data cleansing methods. Furthermore, the hyperparameters of the selected method need to be tuned and the number of trees (or the number of layers and nodes at each layer) needs to be calibrated as well.

How well Machine Learning and Deep Learning is applied has a significant impact on the performance of the model. With Automated Machine Learning, the algorithm is trained to select the right model to use, how to use it without the input of a data scientist and without delay. Unlike manual Machine Learning, the algorithm is trained to constantly check that the model being used is the best possible model. If conditions change, the algorithm automatically replaces the model with a better one.

At the plant level, this type of solution does not ask engineers to become citizen data scientists, a fluctuating role that risks the rise of shadow solutions and workarounds. Instead, engineers see the end results of the Automated Machine Learning process: an alert delivered via dashboard, SMS or email warning of upcoming machine failure.

What’s preventing widespread adoption?

In our estimate, less than 5% of sensor data is ever analysed by PdM tools. Given the reduction of cost in computing and sensor costs, cloud-based Machine Learning should easily become a dominate solution in the PdM category.

However, one of the most significant constraints is tactical: plants often lack access to the sensor data that their own machines generated. In some cases, third-party vendors are the gatekeepers of internal data and lack the incentive to help factories migrate to Machine Learning based Predictive Maintenance.

Another common limitation is connectivity. Despite being a major enabler of many IIoT initiatives, connected sensors have yet to become the norm across industrial sectors and regions. The quality of connectivity available is in itself in flux, and in the near future plants will need to decide whether to stream data using Wi-Fi or cellular networks such as 4G or 5G.

At present, however, plants that have not yet invested in connected sensors cannot take full advantage of Machine Learning-based Predictive Maintenance. A lack of access to real-time, high-quality Big Data makes it impossible to train Machine Learning algorithms to recognize healthy and unhealthy asset behaviours. Plants that rely primarily on handheld data collection are unable to adopt AI-driven Predictive Maintenance solutions.

Summary and conclusion

Many of the enablers for Machine Learning are in place: inexpensive, cloud-based computational capabilities and advanced Machine Learning algorithms that can be applied to sensor data. In the short and medium term, plant executives will need to work with O&M stakeholders to define their Machine Learning strategy and then find ways to overcome internal obstacles that are preventing data access and analysis.

As more industrial plants use sensor-generated data to predict evolving machine asset failure, it will create momentum in the Machine Learning Predictive Maintenance category. Cloud-based solutions are now ubiquitous in the corporate IT environment and it is inevitable that this technology is applied to Big Data.

It is not a question of whether Machine Learning for Predictive Maintenance will be applied. Deployment is a function of the maturity of each industrial plant in terms of technology readiness and its digitalisation roadmap.

B&P2024