You are Home   »   News   »   View Article

Using AI for asset management on data and images

Thursday, October 6, 2022

AI is being used for asset management in analysing data and analysing images. Two experts explained what is happening and how to do it, on an AVEVA webinar

AI is used in asset management in two different ways - to better analyse data, and to better analyse images.

Engineering and asset management software company AVEVA is aiming to 'infuse' its existing and new software products with this AI, said Mike Reed, senior manager of AVEVA's AI centre of excellence, based in Chicago. He was speaking at a webinar on Apr 13, 'Implementing AI for Enhanced Asset Management.'

AI works particularly well for software running on the cloud, because it is easier to build tools which pull data out of multiple cloud software products,' he said.


AI for asset data

AVEVA sees the various forms of AI on asset data in terms of ever-expanding circles, where each stage adds more to the previous one, Mr Reed said.

Stage 1 is stand-alone automated analytics, data tools which run through operational data, looking for differences to normal behaviour, and which can analyse images. They can also learn patterns of normal behaviour and so spot if something is different to the norm.

Stage 2, known as 'condition-based rules,' is where the results of the Stage 1 analytics trigger some specific step to be followed. This can be known as 'dumb AI'.

Stage 3 is guided analytics, sometimes known as 'light touch AI", where you are analysing streams of data, finding relationships between data which you did not know about before.

Stage 4 is what Mr Reed calls 'advanced analytics' - trained AI models built around open platforms, covering machines, systems or processes, together with background knowledge such as engineering, operations and historical data. Also, known patterns of failure and known anomalies.

A form of this is predictive analytics, which is done by combining some level of AI with historical data, so you compare what seems to be happening now with what has happened before.

For example, if you are about to rush out of the office at 5.33pm to catch a train, it could tell you that most of the other times you left at 5.33pm, you did actually miss the train. Or it could tell you that according to its modelling, you need to leave the office at 5.30pm to do all the necessary steps involved in getting to the train.

It is possible to get much better insights from data with AI techniques, compared to straight analytics or rules-based techniques, he said.


Working with data

Predictive analytics can get more meaningful insights from data, Mr Reed said.

Basic analytics techniques might be limited to showing you graphs of data - but there's no means of detecting from that whether what you are seeing is normal or not.

And basic analytics is limited in what it can do with alerting. A system can be configured to sound an alert if a certain sensor reading goes above a set point, such as a temperature being too high. But this set point needs to be set wide enough to handle normal changes in behaviour, seasonal variations and product variations, without an alarm going off in normal conditions.

But if the set points have too wide a range, the system will not detect any problems. Or, you will find that by the time the system operations have gone outside the set point, you have gone 'past the point of no return' - something is damaged, and you need to take reactive steps, he said.

With a predictive analytics system, in contrast, you can overlay what is happening from where you think you should be, also using data from other sensors.

The point of concern can be the point where the signal diverts from what it would be expected to be, he said.

An analogy could be the monitoring systems some of us have on our own body, such as for blood pressure or cholesterol. If they start deviating from what we expect, that can be an early warning indicator of something going wrong to the body.


How much maintenance

A big challenge with preventative maintenance is working out the right amount of it to do, he said. AI should be able to help with this.

If you do too little maintenance, the equipment will ultimately break, which can be very expensive. You are always in 'fire-fighting' mode, fixing problems. But 'over maintenance' means doing work which doesn't need to be done.

Before we had computers, there were basically three modes of maintenance - preventative maintenance, where you did maintenance tasks at the interval suggested by the manufacturer; rules-based maintenance, where maintenance is done according to certain logical rules; and condition-based maintenance, where maintenance is done based on some condition monitoring, he said.

With AI, you can automate the analysis of the condition-based rules, and then develop the 'predictive' mode, comparing where something is operating to where it should be operating, to get a better idea of what maintenance is needed.

The system can work out the best amount of maintenance to achieve the multiple goals of preventing failures and reducing equipment downtime; reducing costs; improving safety; extending equipment life; and optimising the asset strategy.


Across the company

If you are implementing predictive analytics 'at scale' such as across a company, there are ways that models can be re-used to save time.

For example, you can create templates of models for one piece of equipment which can be re-used on other equipment of the same type. The model can include what is known about the behaviours of that equipment, he said.

The templates can be used to compare one piece of equipment with another, if it is configured similarly and on the same site.

AVEVA has its own 'template' models of equipment from its work in multiple industries, including power, oil and gas, manufacturing and chemicals.

It is possible to take data out of multiple predictive analytics systems to give guidance to the person in charge of running the whole plant. For example, it can tell them which items are going out of range, what they need to address, and what information they need to pass onto others.

The analysis can also try to identify the remaining useful life of a piece of equipment, or how long until something becomes inoperable, where maintenance work is essential rather than a choice.

There can be analysis of the start-up and shutdown patterns, to see if they would work better at a different time scale.

'This is all integrated in the software and between the software,' he said.

Over the whole company, many customers use 'enterprise-wide solutions, with a central predictive analytics server connected to multiple historians on different facilities,' Mr Reed said.


Getting started

The best way to get started might be to run AI tools on the data in your historian, since it has a large amount of historical data ready to use. 'The predictive analytics platform can sit on that and talk to it,' Mr Reed said. 'You build the models based on sensor data that you have available.'

Systems are typically deployed in 1-3 months, Mr Reed said. The deployment phase can itself have a value, if you discover system anomalies during this time.

A side benefit is that you focus on making sure the sensors work. Otherwise, 'if sensors are not required to operate equipment, and if something goes wrong, it typically gets put in a queue to be worked on and gets forgotten about.'

'Since we're leveraging those [sensor data] in the models, we're going to fix them if they are not working.'


Vision AI

Another way to use AI for asset management is to analyse imagery. AVEVA has a software tool 'Vision AI Assistant' for this. It can use any images, including from phones, webcams or drones.

There are two primary modes of operation - anomaly detection, where you are looking for something unusual, and 'discrete state classifier', where you are looking for something specific, such as whether an image shows apples or oranges.

An example of using anomaly detection was a system to detect problems with a chain in a conveyor on a factory.

This chain is very critical to operation of the whole plant, driving a conveyor which moves parts from one station to another. The chain is nearly a mile long and takes 22 minutes to go completely around, and the company had 10 such chains.

If a chain damage is not identified straight away and the chain snaps, it needs to be rethreaded, a time-consuming task. The costs of breakage was so high the company was employing people to personally monitor the chain - but it was a difficult job to keep paying attention to, particularly on the night shift.

AVEVA's Vision AI system could identify the difference between a normal chain link, and one with a broken metal plate. If the chain ever got snagged, and had a broken link, a snap was likely.

For training, the system was shown images of what 'normal' looks like, so it can identify an abnormal chain link. Any anomaly is scored, with a higher score for a bigger anomaly.

The system makes an estimate of the deviation (level of the anomaly) according to an algorithm. This can be a large deviation in one sensor, or multiple sensors showing a small deviation at the same time.

That is analogous to waking up in the morning and feeling lots of aches, pains, and having a sick feeling and sweating. Individually they may be minor, but getting them all at the same time indicates a problem.

The other way to use Vision AI is a discrete state classifier. An example is monitoring flares in upstream oil and gas - whether they are lit and for how long.

EPA regulations require operators to submit data about how long flares are active. This data is also useful commercially. 'If the flare is running 24/7, you're pumping more than you can handle,' said John Leighton, Sr. Presales Consultant with AVEVA.

The system was trained to work out if a flare was on or off from the image. It can also look for black smoke and determine if a pilot light is on or off. It can measure the size of the flare, which is likely to correlate with the gas flow rate to the flare.

The data was gathered using a static video camera.

These discrete state classifiers can be part of a data processing 'pipeline' model, where raw images go in, and the insights come out.

As part of the pipeline, you can tell the AI which part of the camera's frame it should be looking at, so it does not get a false positive from the sun.

A more complex example was a system to inspect welding of a component in a manufacturing process. The inspection needed to be done accurately and quickly.

The system was programmed to look at two specific areas of the image of the component, detecting burr detection and a bimetal bend. It could be programmed to detect anything which is easy to 'see' visually, such as the height of a biscuit or the colour of a soda drink.

Typically, anomaly detection may need 1000 images for training, and a discrete state classifier uses 100 images for training, Mr Reed said.

The AI Vision could be used in asset integrity management, in determining whether a certain component was running properly or not. It could be used with thermal camera images of temperature patterns, and used to identify hot spots.

Getting good data quality is normally easy, for example for the chain imagery a camera is taking a clean image of the chain every time its beam is broken, indicating the chain has moved to a new link.



Associated Companies
» AVEVA Solutions Ltd

CREATE A FREE MEMBERSHIP

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

DIGITAL ENERGY JOURNAL

Latest Edition Mar-Apr 2024
Apr 2024

Download latest and back issues

COMPANIES SUPPORTING ONE OR MORE DIGITAL ENERGY JOURNAL EVENTS INCLUDE

Learn more about supporting Digital Energy Journal