Many maintenance, repair, and operations (MRO) teams are understandably wary of glitzy new AI solutions that don’t deliver results. So, how can you tell which solutions add value, and which are simply for show?
The difference is in the training. Superior AI tools have been painstakingly trained on vast amounts of high-quality data. That’s what gives solutions like Azima DLI the depth of knowledge needed to identify patterns, diagnose problems, and make accurate predictions for your predictive maintenance program.
Azima’s diagnostic engine uses billions of data points to isolate faults in virtually every kind of standard machine and component operating today. The result is a precise analysis that guides maintenance strategies, in turn extending asset lifespan, improving performance, and lowering costs.
In one oil refinery, for example, implementing Azima transformed maintenance operations. The number of unplanned machine failures dropped to nearly zero (where beforehand, the plant had experienced near-daily failures). Mean time between failure (MTBF) increased from 22 months to 52 months because faults were detected early on. And costs plummeted, in large part because maintenance crews no longer worked overtime.
That’s the kind of success many Azima clients have experienced, thanks to the effective use of AI and big data.
Big Data and Artificial Intelligence
Artificial intelligence is most effective when it is trained on a high volume of data. This makes intuitive sense when we remember that AI is designed to mimic human intelligence.
The human brain is continually collecting data, contextualizing it, and using it to draw conclusions. Experienced technicians spend years building up a foundation of institutional knowledge — that’s what allows them to pinpoint machine faults and diagnose new defects in the early stages. With the right training, AI models can draw similarly reliable conclusions from datasets far larger than what human technicians typically handle.
The cliché “garbage in, garbage out” is all too true when it comes to training AI. In order to produce a reliable, trustworthy AI model, data scientists need to invest extensive time and resources in training the model on carefully collected and curated data. (In that sense, it’s not so different from the process of educating a person. The more time and effort we pour into “teaching” AI how to do something, the better the result will be.)
The process of training an AI model can be broken down into roughly three steps:
- Data collection – The first stage of training is data aggregation, or acquiring a large amount of data and creating defined data sets for training, testing, and evaluating the AI model. An AI-driven diagnostic engine needs to have a huge volume of information about industrial assets. The model should be able to recognize healthy and unhealthy asset function.
- Data preparation – At this stage, the data collection is carefully structured, processed, and cleaned. The cleaning process reduces the “noise,” or irrelevant information that the model has to sift through. The AI’s model ability to predict asset failures depends on its capacity to spot patterns in the data. This requires carefully curated, consistent, and well-prepared data.
- Model training – A successful AI model is able to make predictions and draw accurate conclusions from new data. This requires an intensive training period, in which the model runs a series of tests, makes predictions, and then measures its results against the correct answer. It takes multiple rounds of training for the model to “learn” how to read and analyze data correctly (after all, the model is catching up to human experts who have spent years learning how to interpret data).
A successful AI model is trained on round after round of carefully chosen and prepared data, until that model can autonomously analyze new data and come up with accurate predictions. That’s the kind of training that went into Azima’s diagnostic engine, which uses billions of data points to identify machine faults.
Azima’s Competitive Advantage
The engineers at Azima have spent about four decades building up the data resources and insights that drive their diagnostic engine. Let’s take a closer look at how Azima’s AI model was developed.
Building an Automated Diagnostic System
The roots of Azima’s data-based AI go back to the 1980s, when Azima was a marine engineering firm based out of Bainbridge Island, Washington. In the 1980s, the company was contracted by the U.S. Navy to create a condition assessment system for Naval assets.
Azima’s engineers developed a system of vibration measurement technology that allowed technicians to establish machine baselines and detect early indications of asset faults. This system proved so effective that the Navy awarded Azima a contract to work on its aircraft carriers. Over time, Azima’s senior analysts built an automated diagnostics program based on vibration measurement technology. This system also laid the foundation for today’s AI-driven diagnostic engine.
By the late 1980s, Azima was leveraging technological improvements like new, portable vibration monitors and computers to scale its automated diagnostic program and meet the needs of industrial facilities. The team’s developers created a secure means of sending data to the cloud for analysis.
At first, the analysis was carried out by expert technicians. Eventually, the engineers perfected the automated diagnostic system so that it could accurately identify the earliest signs of machine and component faults on its own.
A Data Collection Rich in Institutional Knowledge
As Azima scaled and expanded into industrial settings, the team’s engineers were also collecting and saving vast amounts of data.
By the 1990s, Azima was saving machine measurements in a standardized database — the same database format that they still use and maintain today. Only now, the Azima databases contain machine tests on just about every kind of commonly used industrial machinery.
That means precise vibration measurements on compressors, pumps, and gearboxes in a wide range of conditions, from optimal running condition to failing. There’s data showing exactly what a compressor’s measurements look like when it’s developing a minor defect, and what it looks like when it’s near failure.
That’s not all. Azima’s databases also contain the assessments that were carried out by the technicians who worked on all those machines. The result: the databases are a repository of decades of expert human analysis and insight. Altogether, Azima’s tagged data resources comprise ideal training material for an AI model: decades worth of data points on machine measurements, patterns of fault progressions, and human insights into those machines.
Training the AI Model
The Azima diagnostic engine is trained on Azima’s vast reams of data until it can accurately diagnose hundreds of distinct faults, in just about every kind of machine and component.
Azima uses a rules-based system to train its AI model. The model learns how to interpret a complex set of vibration measurements to assess the health of motors, pumps, gearboxes, and other industrial components. During the training process, the AI model is tested using data from Azima’s storehouses, to see whether it can weigh the data and apply the rules to reach the correct diagnosis. The process is iterative: the model is corrected until it is able to match the original diagnosis created by human experts. Once the model is successfully trained, it can apply the rules-based system at a massive scale.
How do those rules work? Azima’s system uses a range of vibration measurements to create a matrix of relevant data and extract features that it compares against a baseline for each individual machine. The rules don’t simply flag high vibration levels: they examine the whole machine, comparing sensor readings to see precisely where the vibration peaks are located, how they relate to all other vibration peaks across the entire machine train, and if any exceed their baselines in various categories. The result is a system capable of spotting minor defects in a component part, long before the asset gets close to failure.
What Sets Azima Apart?
Multiple factors set Azima apart from other AI tools in the market today.
Data resources. Azima’s database of machine information is unmatched. The AI model is trained on over a trillion data points. The model has already seen exactly what machine failure looks like; it can easily spot a loose coupling or a faulty bearing and decipher between imbalance, misalignment, and soft foot. Additionally, it can identify hard-to-decipher faults like piston rates, engine misfires, and broken motor rotor bars. That extensive training means that Azima’s predictive maintenance AI is ready to work right out of the box, without a lengthy on-the-job training period.
Generations of expertise. Azima’s database includes decades of expert technical input. This gives Azima’s AI tool a depth of institutional knowledge that other tools can’t match.
A complex rules system. Some Azima competitors rely on a simple formula for predictive maintenance: once the vibration levels exceed the ISO recommendations, the tool issues an alert. Azima’s system is much more complex. It doesn’t simply spot major anomalies; it drills down and finds the precise location of the problem. Then it sifts through the data and issues a prioritized, plain language diagnosis of the problem, along with a set of repair action recommendations for crews to follow. That’s prescriptive maintenance at its most reliable.
The Real-World Impact of Azima AI
Azima’s diagnostic engine has a broad range of use cases across many sectors. Here are a few examples of the tool’s ability to increase asset performance, reduce costs, and improve safety in the workplace:
- In one mining operation, Azima’s diagnostic engine identified two severe faults that would have risked lives and damaged productivity if left unchecked. After just one day of collecting data on a longwall mine, the Azima team discovered serious damage to a conveyor which, if not caught in time, could have led to fires or explosions, not to mention a significant loss of productivity.
- Azima’s diagnostic tools revealed a potentially catastrophic electrical fault at a petrochemical plant. Left undetected, the plant’s capacitor would have failed, incurring a penalty of $30,000 a month. Plant personnel could potentially have been seriously injured as well. Significant costs would likely have resulted from collateral damage to nearby equipment and from extended repairs if the fault had not been detected early on.
- At one military facility, transitioning to Azima’s automated predictive maintenance tool saved tens of thousands of dollars every month — and also resulted in greater diagnostic accuracy than the old approach of manually combing through data.
- The Azima remote condition monitoring team discovered a serious fault in the gearbox of an offshore wind turbine. The team flagged the issue quickly enough to have it taken down and repaired before it could create a catastrophic failure.
Final Thoughts
Predictive maintenance is the surest way to stay ahead of asset failure, improve overall performance, and reduce maintenance costs. But for many organizations, factors like labor shortages and budget restrictions can make predictive maintenance seem difficult.
That’s why Azima’s AI tool is such a game-changer. By harnessing decades worth of data on industrial maintenance, Azima has produced a diagnostic tool that operates right out of the box, no matter what your workplace looks like. Azima’s wealth of institutional knowledge is all packed into its AI model, so that every user can benefit from those decades of expertise.
Ready to learn more about what Azima can do for your operation? Our team is ready to answer your questions – reach out today to get started.