In a low-margin, high-volume world, money spent on preventative maintenance may be seen, by some, as wasted, as there is no immediate benefit. Even long-term benefits are difficult to attribute to some earlier spend, since the benefit is typically either no breakdowns, or something not happening at all.
“In most cases manufacturers operate a combination of preventative and reactive maintenance, although since reactive maintenance is the priority, preventative maintenance may not happen when a plant becomes less reliable, meaning the situation becomes even worse,” says James Davey, service manager at Boulting Technology. “Traditionally, equipment maintenance is performed on a regular basis to prevent system failure. Despite these checks, equipment can still fail and cause lengthy plant downtime while repairs take place.”
Benefits of predictive maintenance
Davey believes that predictive maintenance is a much more effective method, combining sensors and machine learning to predict when equipment might fail, allowing plant managers to plan for downtime and ensure they have the correct parts for the job.
“Predictive maintenance brings with it a wealth of benefits in addition to reducing downtime,” he adds. “Its ability to detect potential faults before they happen can increase productivity and improve the quality of products as well. It can also reduce replacement equipment costs, increase safety and strengthen overall output.
“Through analysis of the data and application of clever machine-learning algorithms, the plant manager can build a comprehensive picture of maintenance requirements, resulting in a predictive maintenance solution.”
Risk-based maintenance
The ability to spot problems before they happen and put solutions in place, reducing the potential for unplanned downtime, is a vital tool for plant managers. One way of addressing this is through a risk management plan that can provide plant managers with an all-encompassing maintenance solution.
One such service is being pioneered by Boulting Technology. The new programme consists of a comprehensive survey that assesses control systems across a facility. A survey, conducted by Boulting’s engineers, provides plant managers with a report detailing the potential risks posed by equipment, such as control system failure, critical obsolescence and other scenarios that could cause unplanned downtime.
The initial online survey assesses areas such as obsolete parts, equipment lifecycle, and efficiency.
The corresponding report uses a traffic light system to make recommendations. The multi-stage recommendations, which are based on clients’ key parameters, provide a hierarchy of risk. This helps to characterise and focus on the high risk critical systems in the first instance, enabling the plant manager to implement an appropriate and cost-effective action plan.
“By proactively managing risk, plant managers can plan maintenance schedules around the equipment they currently have and so avoid costly breakdown and downtime,” Davey continues. “Control system maintenance is particularly important as these systems are often integral to the whole facility. PLC and SCADA systems are the heartbeat of a plant, so if they fail, the whole plant may become vulnerable.
“Initial feedback from our clients has been overwhelmingly positive, particularly from the low-margin, high-volume industries, for which unplanned downtime can be particularly devastating.”
Maintenance for servitisation
Predictive maintenance is becoming ever more important in manufacturing, as customers now expect maximised product uptime, increased performance and many other service-related metrics. This shift is known as servitisation or selling outcomes instead of strictly
selling products.
Industry 4.0 technologies such as sensors, the Internet of Things and machine learning enable the pre-emptive repair of equipment before it ever fails, and manufacturing is moving from a reactive, break-fix service model to one aimed at maximising product uptime.
For example, Rolls-Royce started by investing money in a data warehouse, not even knowing what to do with that data at the time. But, through time series and motion studies on active aircrafts to move to predictive maintenance, they began to fix parts between flights, eventually realising they could start selling engine uptime instead of just engines. Their messaging shifted from “you need an aircraft engine,” to “you need an aircraft engine that is functioning when you need it to function,” and they’re now positioned as a service organisation.
ThyssenKrupp had a similar problem: the company wanted to figure out how to disrupt the market, but they struggled due to their aging technicians and their expiring workforce knowledge. Their digital solution? Slap sensors on parts, give technicians augmented reality (AR) glasses and move from simply selling elevators to selling elevator uptime. Now, legacy elevators from competing manufacturers are running into the same problems ThyssenKrupp once faced, and they’re starting to put ThyssenKrupps sensors into their own elevators.
Anomaly detection
Another strategy gaining traction among manufacturers along the Industrial Internet of Things (IIoT) path is anomaly detection. This allows companies to load their data securely onto the cloud, detect equipment anomalies, predict failures before they occur, and validate against failures, both known and unknown, thereby confirming pro-active steps that should be taken in advance to avoid unplanned downtime and unscheduled maintenance.
One such system is being trialled by Progress, the provider of application development and deployment technologies, the Progress DataRPM self-service anomaly detection and prediction option for the IIoT market.
The flood of data coming from sensors on industrial equipment gives asset-intensive organisations a tremendous opportunity to prevent failures and optimise output. However, industrial organisations globally are struggling to make sense of their data and to detect anomalies and prevent failures that otherwise often go undetected until costly failures have already occurred.
By utilising the anomaly detection and prediction capabilities, asset-intensive companies can unlock the power of the IIoT
to capture and analyse their own industrial sensor data privately and securely to dramatically reduce downtime and increase overall equipment effectiveness.
“With billions of interconnected devices pumping out untold volumes of data, there is a huge demand for ways to gather valuable insights from the data,” adds Dmitri Tcherevik, chief technology officer at Progress. “But with limited budgets and lengthy deployment cycles for many machine-learning applications, the true value of data is often left untapped or underutilised. That is why we now offer an R&D self-service option for those organisations looking to start on their IIoT journey more quickly and easily than previously possible. R&D teams can use our self-service cognitive cloud-based application to immediately start detecting and predicting anomalies across their industrial data for fast time-to-insights and more accurate ROA [return on assets] calculations.”
The Progress DataRPM application uses cognitive techniques and advanced machine learning and meta learning-based algorithms to identify and predict anomalies, often before they occur in the production environment. Meta-learning, a subset of machine learning, is a set of algorithms that teach computers how to self-learn in difficult IIoT big data environments.
Anomalies in appliance manufacturing
Appliances and electronics have evolved, and customer expectations have evolved right alongside them. People expect their gadgets and appliances to run smoothly, with zero downtime or breakdown. With more competition in the space, manufacturers are being challenged to improve production quality to meet rising expectations.
One Fortune 50 industrial appliance manufacturer needed to identify the indicators of post-purchase failures for industrial washing machines. These washing machines, which were often running 24/7 with large loads at commercial establishments, had uptime service-level agreements (SLAs) that needed to be met.
Unfortunately, manual analysis of machine and sensor data is not truly scalable in the modern digital landscape. Predicting the possibility of such failures with a human approach is not feasible: there are too many possible factors that can be unique to each machine based on operating environment and other variables.
As such, the client was unable to prevent machine failure, which impacted the company’s ability to meet established SLAs. With upward of 75 sensors producing data, engineers simply didn’t have the ability to analyse all the data and predict failures.
Manual alert rules were sporadic and limited in scope as they were only able to pick up some of the sensor data at any given moment. On top of that, it took months for the company’s data scientists to build models and deploy them to production and they often only resulted
in high false positives. This increased maintenance costs, which only multiplied the existing set of challenges for the client.
The company needed an automated predictive maintenance solution that could learn from data and failure events and then adapt the predictive model on a continuous basis. This would help boost sales, increase customer satisfaction and improve the overall experience for customers.
To accurately predict and prevent future failures, the client used Progress DataRPM Cognitive Anomaly Detection and Prediction (CADP) to sample sensor data and generate an automated approach for identifying anomalous system behaviour. This analysis helped in predicting system failures with better overall accuracy, precision and recall than any of the previous manual models that were used by the client’s data science team.
This automated approach to predictive maintenance came as a boon to the client, which not only diagnosed discrepancies in manufacturing operations but also grouped the combination of sensor states. This resulted in a major shift of the sensor data sample that was used to train data models accurately, from just 15 minutes to months of sensor data that the machines could then use to learn patterns and anomalies. With this data, the client was able to build an ensemble of predictive models that delivered highly accurate predictions.
Using an automated predictive maintenance approach, the client was able to analyse all the data from 75 sensors, resulting in a highly accurate prediction model. The CADP solution analysed 300 million records together to group these states using automated segmentation algorithms into clusters.
These tagged clusters were then combined with the remaining sensor readings to characterise these clusters using CADP’s Influencing Factor Analysis. A further automated analysis triggered the product to either reject or accept the clusters as ‘failure states’ for the system.
Finally, the model was created automatically to predict future failure states trained on existing tagged data using CADP’s Prediction Algorithm. Using this predictive model, incoming readings were ingested for analysing potentially failure states.
As a result, the processing time to predict failures was drastically reduced from upward of six months to just less than two days. The client was thus not only able to accurately predict the indicators of appliance failures, but also predict the possibility of future failure with respect to service and was able to recommend suggestive actions.