Digital Energy Journal- ML on equipment - spotting failure signatures

Join | Existing Member Login

You are Home » News » View Article

ML on equipment - spotting failure signatures

Friday, November 22, 2019

The best way to use machine learning to improve equipment performance is by spotting a 'signature' of an emerging problem, perhaps months before it actually emerges. We talked to Mike Brooks, Asset Performance Management (APM) consultant at AspenTech.

One of the most useful ways that machine learning can help improve equipment reliability is if it can spot the specific 'signatures' in sensor data which indicate a problem emerging, says Mike Brooks, Asset Performance Management (APM) consultant at AspenTech.

AspenTech provides software for asset optimisation, working in the chemicals, manufacturing, oil and gas (upstream / downstream), power, engineering, and wastewater sectors. It is based in Massachusetts.

Equipment problems often follow a certain course of action, which you might be able to spot with the help of machine learning, Mr Brooks says.

One example, often referred to as the silent killer of compressors is liquid carry-over when micro water droplets travel into a compressor, causing pitting and deposits that lead to vibration and catastrophic bearing failure.

This can arise from a complex sequence of events, starting with a higher ambient temperature in summer, meaning that gas is not cooled to the desired temperature in a cooling system, so it has a higher volume and so a higher velocity through a separator, which can shear droplets from the surface. The water droplets then cause damage to the blade, or imbalances in the bearings.

All of this leaves a trail of sensor data which sophisticated machine learning algorithms can detect - seeing indications of a problem emerging months before it actually does.

So you are not just identifying a change (or 'anomaly') to a normal operating pattern, you are seeing a specific series of events identifying a failure about to occur, he says.

The same failure patterns can often be seen on similar equipment, such as pumps, and even static process equipment such as heat exchangers and boilers.

The AspenTech software takes data from (for example) 160 different sensors on a compressor, to make a continuous assessment of how close operations are to normal and to detect when specific failure patterns emerge.

By learning the signatures leading to failure rather than only detecting anomaly conditions starting to happen, Mtell provides much earlier warning of problems, which gives companies much more time to arrange maintenance work, in a safe coordinated manner including spare part deliveries.

The system is also designed to spot deteriorating process operating conditions which might impact product qualities and yields.

ARC of Boston, USA has conducted studies that have shown that 85 per cent of all equipment failures are caused by errant process conditions, rather than just wear and tear of equipment, Mr Brooks says.

Bearing this in mind, some operators of complex equipment now say they want to eliminate maintenance completely - just by establishing when they have disorderly process behaviour and adjusting accordingly to avoid the deterioration.

Other companies

Many companies have aimed to develop services which don't do anything more than spot 'anomalies', something different from normal conditions, by analysing sensor data, Mr Brooks says.

But these can just end up dumping lots of unhelpful work on company experts. For example, the software spots 10 'anomalies', an expert then needs to figure out if they are a genuine problem or not - and if so, work out what to do about them.

Also, many other equipment monitoring companies limit themselves to vibration analysis. By the time the equipment is showing high vibration it may be too late, you already have damage.

The AspenTech software, by contrast, looks for the root cause conditions manifested in small changes that happen over time.

Many companies build engineering and statistical models of 'normal' operation of the equipment, so they can be compared with current operations. These models can take months to build and expect pristine data and operating conditions.

But the only time equipment normally works according to the engineering and statistical model is when it is new, Mr Brooks says. All equipment will wear and 'drift' over time, and this is normal. The AspenTech software can track slow drift in the sensor readings over time, without causing any alarm, unless it sees a signature of a fault.

Emphasise what it does

Although the AspenTech software uses machine learning, Mr Brooks would prefer the software to be recognised simply for what it can do - provide precise early warning of a problem, so you have plenty of time to fix it.

The purpose of the software can simply be described as 'to make machines smart' - to tell you when they are about to break.

'We say we are [helping customers] understand the problem and providing the kit to solve the problem. We are going to stop machines breaking,' he says. 'We get machines that can run better, and longer and maintenance cost goes down.'

And of course, machine learning technology itself does not fix the problem - you also need a work process and a methodology.

Mr Brooks also believes software should be known for what it does (help improve the reliability of the equipment), not the embedded technology alone; it's machine learning and it needs data context and domain knowledge to achieve the results.

Mr Brooks notes that in the past a lot of business intelligence and data warehousing projects failed because organisations did not know what problems they wanted to solve before proclaiming business intelligence or data warehousing as the answer.

'I see the same things happening today. Machine Learning alone cannot solve the problem. It needs guide rails to ensure it provides correct, meaningful answers. It needs domain expertise and meaningful contextual data.'

In energy and manufacturing, the overall problem people want to solve is usually improving operational excellence, or achieving small improvements. For example, many firms see the deployment of digitalization, IIoT, Industry 4.0, big data and machine learning as a business initiative where really what they are really trying to do is improve operational excellence and those other things are the means to do it.

In his earlier role as CEO of Mtell (a machine learning company acquired by AspenTech) Mr Brooks insisted that the software should be usable by 'Joe Normal', who may be an expert in compressors, but is not an expert in machine learning. 'I wanted to do this so you don't need a data scientist,' he says. 'Small companies can't afford to hire their own data scientists.'

Also, the software should not require anybody to change their standard work processes in order to benefit from it, Mr Brooks believes. They should be able to do it with their current skills and the way they work right now.

Rather than exposing customers to the workings of the machine learning, you just see a nice user interface. The customer never needs to see the model underneath it. 'We don't talk to people about our algorithms,' he says, 'they do not need to know, we handle all that under the covers.'

Prescriptive maintenance

Mr Brooks tries to avoid the term 'predictive maintenance,' saying that most companies that sell it offer no prediction at all, or just basic prediction, which makes it meaningless.

So instead, it uses the term 'prescriptive maintenance,' meaning a software system which identifies issues and then tells you what operational changes or maintenance tasks need to be done to solve any emerging problem. This notification can also be passed into work management and procurement systems.

Building and updating the model

The pattern identification for detecting normal and failure conditions can be made in a few hours or for very complex equipment in a few days - this is a profound change for software in this space that can take months to develop and deploy - and end users can do it themselves. The company wrote down its 'best practise' for implementing the system, and it has been through four iterations so far.

Machine learning needs domain knowledgeable people to provide 'guide rails' - working out what data would be needed to feed it, which data relationships (that the machine identifies) make real world sense, Mr Brooks says.

The set-up work involves gathering data for example for the past 2-5 years, also talking to engineers, and looking at work orders (for example in ERP software) to see what maintenance was done. Then it can be possible to see patterns in the sensor data which led to failures.

The work, to conduct a test on several assets at the Saras refinery in Sardina, Italy was done in about 2 weeks. Another job on a petrotechnical plant in Southeast Asia took 2 people 3 weeks, covering 4 pieces of equipment, where the competition took months, Mr Brooks says.

The system can work on much more than rotational equipment, which is a limitation of most other similar services, he says. There is no reason why the system can't be used on more complicated equipment as long as it has access to sensor data. Consider aeroplanes. The engines may be carefully monitored by their manufacturers - but the item which fails most often is the toilet.

The technology could be applied in the shipping industry, although the company has not worked in shipping yet - but has done similar work on locomotive engines. There are big engines which fail. In this sector it might welcome a partner interested in applying and implementing the technology.

Updating the model

If the system notices that something is happening which it hasn't seen before, the next task is to classify it manually.

If it is identified as a new 'failure pattern', then it is a manual task to build an 'Agent' to look to see if it happens again. These 'agents' can then be used in other similar equipment, such as your other boiler feedwater pumps.

If the change is identified as a 'process change,' such as the refinery taking a different type of crude oil, it is a simple task to just tell the computer, and the baseline pattern is automatically recalculated. After process changes, there may need to be recalibration in what the 'failure modes' look like.

Background

Mr. Brooks has a background in control, automation and industrial IT systems. He is a former president and CEO of a company called Mtell, which was purchased by AspenTech.

Before that he was a venture executive with the venture capital group of Chevron. He began his career as an engineer at Esso and Chevron.

Mtell was the first company to use machine learning in asset management, beginning 10 years ago Mr Brooks claims.

The CTO had worked with machine learning in neuroscience and brought forward the idea of applying the technique in the maintenance arena.

He has been involved in the 'O&M Foundation', a project with a mission to connect operations and maintenance systems.

AspenTech had been a leader in developing dynamic models for complex equipment, and so the machine learning technology was complementary. It provides products for mining, energy, chemicals, food and beverage, as well as the oil and gas sector.

Associated Companies
» AspenTech

CREATE A FREE MEMBERSHIP

To attend our free events, receive our newsletter, and receive the free colour Digital Energy Journal.