Artificial intelligence (AI) is a broad and poorly-defined term that involves the simulation of human intelligence. This simulation can include allowing computers and robots to understand and generate natural language, perceive their environment, reason, learn and solve problems. A general AI, which has not yet been created, could do all of this in a wide range of circumstances. The very broad definition means that any condition monitoring system could be considered to use AI, since sensors give it the ability to perceive its environment and some algorithms allow it to process this information to determine when maintenance may be required.
However, a conventional condition monitoring system is explicitly programmed to give alerts when threshold values are reached. When people say that AI is being used, they are not normally referring to basic sensor systems that have been programmed with a set of logical instructions. What they really mean is that it uses machine learning (ML), which includes any technique allowing machines to learn to recognise patterns, make predictions or perform tasks without being explicitly programmed. ML includes relatively simple methods such as linear regression that allow inferences and extrapolations to be made from data, and optimisation algorithms that use a deterministic search to adjust variables in order to minimise or maximise some function of those variables. Machine learning also includes artificial neural networks (ANNs) that can learn much more complex patterns that really wouldn’t be feasible to program explicitly. Recognising handwriting or individuals from photos are examples of tasks that ANNs perform well.
Condition monitoring involves spotting patterns in sensor data that indicate developing issues with machines or structures. This can be as simple as a single sensor reading, such as a temperature exceeding a predefined limit. Often, vibration signals are decomposed into underlying frequencies using fast Fourier transforms (FFTs) and limits are then set on the amplitude for each significant frequency. The limits may be set by recording a baseline condition for a machine that is operating optimally and then adding some tolerance. Since the baseline values are learnt by observing the machine in operation, technically this could be considered as a simple example of machine learning, but this would barely qualify in most people’s mind. Where the capabilities of ML become more interesting are when slight changes in the signals from multiple sensors produces a signature that can be recognised by a learning algorithm. This might enable earlier predictive maintenance, or more powerful diagnostics of the root cause of faults.
“Our position at ABB is that one should use the right tool for the job,” explains Ryan Conger, technical sales lead for APM at ABB Energy Industries. “We use a number of algorithms for different methods. For unsupervised approaches, we primarily use a clustering algorithm. Since process and machine data are heavily correlated in a plant or refinery, we pre-process the sensor data to extract the process’ causal elements, dramatically increasing the signal to noise ratio. For supervised methods, we’re employing ANNs where useful, using regression approaches in other instances but, in all cases, are anchoring to the physics of the machine or process.”
FINDING A HOME IN PHARMA
Dr Ing Peter Kytka, global sales specialist at Bruel & Kjaer Vibro, asks: “Why should pharma consider condition monitoring?” It’s an interesting one. “There are primary production assets in the pharmaceutical manufacturing, which are tablet presses or mixers. However, similarly important to getting these primary production assets are secondar production assets which are fans, or pumps for pumping clean water and for the clean room having clean air,” reasons Dr Kytka. “So those secondary production assets are quite important to monitor and keep to these assets reliable.
He adds that the IPF curve is an important element of explaining the value of condition monitoring. “It shows on the vertical axis the asset health over time. ‘I’ relates to the installation of the asset, while ‘P’ represents a potential failure starting, such as a rolling element bearing supporting your fan. There are different techniques, which, as the asset is degrading, can predict asset health. These techniques are ultrasound, vibration analysis, oil analysis, infrared and, if you move further down, your machine starts to ‘cry’. That means it gets very loud and, at the end, it gets hot before it gets very dangerous which is what the ‘F’ refers to (a functional failure).
The use of AI also avoids the need to set arbitrary limits, and instead can intelligently detect more meaningful changes in the sensor information coming from assets being monitored. “Over a long time there are a lot of spikes, but a general increase in the trend may not be there,” reasons Dr Kytka. “So, if you set it wrong, you have a tremendous number of false alarms by the threshold-based alarm approach. And mostly only two level changes can be detected, hence you are unable to detect various types of changes. You can monitor a whole rolling element bearing, but you can also monitor the components of the bearing itself.”
For typical vibration, Dr Kytka points to the ISO 10816 standard, which can be used when analysing velocity for standard machinery. “If it was a pump or a fan on a rigid foundation, there is a standard that gives you the alarm limits,” he reasons. “But one of the main assets from a reliability point of view is rolling element bearings – the patterns are quite complex and there are no standards how to set the values. In very simple terms, the basic information about the change in the data distribution over time is used and applied in our AI and ML knowledge – and it can be used to remove completely the limits. After that we don’t have any more limits, the intelligence automatically detects the change points.”
Conger maintains that leveraging AI algorithms provides two main benefits: breadth of coverage and speed of deployment. “This scenario allows you to: fix the clogged lube oil filter before it damages the bearing; adjust the control logic for the scrubber before liquid destroys the impeller and find the dry gas seal issue before there’s a release of the process fluid. You can also notify the team that there’s a high temperature in the switchgear over a month before it would start a fire. In short, the maintenance team can look around the corner for what’s coming next.”
BOX: HOW AI CAN BE USED FOR CONDITION MONITORING
Data collection: sensors gather data from machinery, capturing parameters such as temperature, vibration, pressure, or sound. This data is crucial for detecting anomalies or patterns indicating potential issues. Feature extraction: machine learning algorithms help extract relevant features from the collected data. These features could represent trends, frequencies, or other patterns that are indicative of the machine’s condition. Model training: models – such as regression models, decision trees or neural networks – are trained using historical data. These models learn patterns of normal behaviour and variations associated with different conditions. Anomaly detection: once the model is trained, it can detect anomalies or deviations from normal patterns. Any deviation might signal a potential issue or impending failure. Predictive maintenance: by analysing patterns and anomalies, machine learning models can predict when equipment is likely to fail. This enables proactive maintenance, reducing downtime and preventing catastrophic failures. Real-time monitoring: machine learning models can continuously monitor equipment in real-time. They analyse incoming data and raise alerts if any deviation from normal behaviour is detected, allowing for immediate action. Continuous improvement: as new data becomes available, machine learning models can be retrained to improve accuracy and adapt to changing conditions, ensuring better predictive capabilities.