Bearing useful news09 January 2023

Image: Delyk

A new approach to condition monitoring relies on computers making their own connections on bearing health from large datasets. By Toby Clark

Sensors and condition monitoring systems (CMS) for machinery are cheaper and more capable than ever, so they can produce impressive amounts of data on vibration, temperature, acoustic emissions and other parameters – but this has to be interpreted correctly to provide useful information.

An alarm can show when a reading is out of range and failure is likely. Better yet, an experienced technician can tell when a small clue (a slightly elevated temperature, say, or a particular frequency of vibration) is likely to develop into something serious. But this requires a high level of training, and constant vigilance.

Now bearing manufacturers are looking at the data produced by thousands of installations across the world, to see if they can automate the diagnosis of potential failures – and even detect problems long before a human could. They are using artificial intelligence (AI) techniques, and firms such as Petasense, Swim, and Falkonry have sprung up with expertise in the area of AI called automated machine learning (AML or AutoML).

‘Big Data’ is the idea that you can combine data sets from a multitude of sources and detect previously unseen correlations which can help to predict future behaviour. Machine learning systems are designed to look for patterns in the data they are given, and ‘trained’ to emphasise significant patterns. If a particular combination of parameters results in a failure, the system should look for that combination, or for factors which could be precursors to that state: for instance, a sudden increase in the rate of temperature rise.

Firms such as US software supplier Falkonry talk about ‘time series AI’: this is a similar concept, which emphasises that the data coming in is changing over time, and that trends are more important than single measurements. They also use the term ‘multivariate’ – in other words, a number of inputs (‘variables’) are analysed at once, and the way they interact is significant.

Corporate moves

In 2019, bearings giant SKF bought an Israeli AI company called Presenso to make use of its proprietary AML software, which SKF called a ‘production-ready analytics solution’. Presenso said at the time that its AI capability could find “anomalies that were previously undetectable” and that it would replace “labour-intensive data analytics tasks with advanced machine learning algorithms that perform these tasks automatically.” Since the purchase, SKF’s AI division has released its own branded ‘industrial analytics solution’ as Enlight AI, and stated its ambition of monitoring 40 million bearings by 2025.

Each Enlight IMX-1 sensor is “a data-collector and radio combined into one compact battery-operated device” which communicates with a line-powered network gateway either directly (if in range) or via a relay of nearby sensors. The sensors typically have a four-year battery life, and the whole network communicates with SKF’s software and (for commissioning) with a smartphone app. The system monitors the usual parameters of temperature and vibration, and can process them with techniques such as ‘acceleration enveloping’ which detects impact phenomena to give early warning of defects in bearings and gears.

The firm describes the goal of the process as “identifying irregular data patterns that indicate upcoming machine failure” and that “automated machine learning sifts through the data and chooses the optimal algorithms for analysing the specific data stream.”

SKF says that Enlight AI complements existing distributed control system (DCS) and supervisory control and data acquisition (SCADA) systems by alerting technicians that the machine is in danger of shutdown “long before manual threshold conditions have been breached”. It adds that although the information is derived from a completely automated process of machine learning, the alerts are delivered via “a user-friendly interface in a format that technicians can understand and act on.”

One of the advantages of machine learning should be that its performance is constantly improving, as the library of algorithms grows and the mass of data that it can build on is increased. SKF says: “By shifting many repetitive machine learning tasks from data scientists to algorithms, the speed and accuracy of asset failure alerts are increased.”

The most sophisticated AML systems can autonomously relate failures or other events to conditions within the components of a machine, so they become self-training. SKF’s Enlight AI system creates models which are “auto-selected out of a pool of dozens of different modelling algorithms and later, auto-calibrated and optimised.” Once they have been deployed, auto-validation processes verify the accuracy of the models continuously: this is said to train a model in “minutes or seconds as opposed to hours and even days using manual methods.”

Digital twins

A related idea is the ‘digital twin’: this is a virtual copy of a machine or system, designed to mirror the behaviour of the physical item, using simulation techniques to help predict when things might go wrong, or to simulate operational changes to try to improve performance. The key to this concept is that the ‘twinning’ goes both ways: there is constant feedback of information from the actual device to the digital twin, updating its parameters and improving its accuracy all the time.

Schaeffler (manufacturer of FAG and INA bearings) has its own condition monitoring system called OPTIME, pictured above right, which uses machine learning to give “an advance warning period of several weeks and specific recommendations for action.” It does this via a smartphone or desktop app, which can monitor hundreds of machines.

To help with traceability (and to make digital twinning easier) Schaeffler has introduced its Bearing Data Service, which stores precise information on each individual product. All Schaeffler spindle bearings with a bore from 25mm up to an outside diameter of 320mm are marked with a unique data matrix code (DMC) – a two-dimensional barcode, similar to a QR code, which contains the serial number and links to a database of the specifications, dimensions and tolerances of that particular bearing (pictured, above left).

Much of the impetus for predicting bearing failures is coming from the wind turbine industry, as gearbox bearing failure is a critical and expensive problem. In Japan, NTN offers a CMS whimsically known as ‘Wind Doctor’ which collects sensor data and aggregates it on a cloud-based server; by sharing data between operators, the quality of analysis and prediction is improving every year.

Toby Clark

Related Companies
Schaeffler (UK) Ltd
SKF (UK) Ltd

This material is protected by MA Business copyright
See Terms and Conditions.
One-off usage is permitted but bulk copying is not.
For multiple copies contact the sales team.