How Ultrasound and AI become the dominant technology
Ultrasound and AI are a perfect combination
Ultrasound is a commonly used technology to detect and predict failures. Space agencies such as NASA use ultrasound to check equipment before they shoot it into the space. Ultrasound is widely acknowledged as the first line of defence for maintenance, as it can detect failures as early as no other technology.
In industrial environments, ultrasound is also used for decades, but with limited reach and success, as the analysis of ultrasound data is a very manual, subjective and therefore a very time-consuming process. With the combination of ultrasound and Artificial Intelligence, this time-consuming process is something of the past and the maintenance crew can just get a simple dashboard with “green”, “yellow” and “red” as implemented within our solution.
The basis concept: Artificial Intelligence predicts a health status
The basis for the health status is the ultrasound file, which is saved during a measurement. In the case of ultrasound, this is often a format which is also used to store high quality music. The Artificial Intelligence then has basically two views on this measurement: a.) The course of a measured value over time. This view is also called "Time Domain" and a visual representation typically looks like this:
To be able to evaluate such measurements by artificial intelligence (AI), these mountains, valleys, distances and their changes over time are described mathematically using statistical factors.
b.) The second view of our Artificial Intelligence focuses on the analysis of the frequency components of the signal, also called "frequency domain". The color differences represent the strength of the signal in different frequency ranges.
Each measurement is additionally broken down into individual sections, whereby in total, a few hundred different factors are extracted per measurement. This is practically the unique “fingerprint” of assets, similar to the "Shazam" app for recognizing music songs. This approach works for all Industries.
How does Artificial Intelligence learn what makes a good and what makes a bad asset?
The answer is twofold:
a.) The characteristics of damaged bearings, for example, have already been described in detail in several research projects. These results can be leveraged and translated into machine learning logic
b.) When it comes to estimating the remaining service life of systems, the only thing that helps is data, data and data.
AI is quite a buzzword this time - whats special about Senzoro AI?
The AI itself consists of "transparent" components where processes take place that we as humans can still understand. But it also contains "black boxes", which lead to the correct result, but are no longer comprehensible to humans in detail. It is therefore not recommended to rely exclusively on "black boxes", when dealing with such an important topic as the prediction of asset failures. We therefore derive the healts status based on various indicators in order to make a final decision if maintenance is needed or not. We use several indicators and in order to not reveal all the “secret sauce”, here are two of them:
(1) Classification of measurements by decision trees (for the data scientists among us: "Random Forrests"), which are automatically derived by machine learning based on the huge data set we have.
(2) Image processing of visual representations of the measurement. This can be the course of the signal over time as shown above, or different representations of the frequency spectrum. The great advantage of this indicator is the possibility to access those neural networks which have revolutionized image processing in other areas as well. In a sense, it is an attempt to solve the challenges of predictive maintenance using the most powerful algorithms in the world.
The evaluation of the ultrasound data with Artificial Intelligence is a completely automated process, without any human intervention. However, due to the novelty of the approach, results are still controlled by our experts in individual cases, especially when stakes are high.
So how do we approach the "Holy Grail” of estimating the Remaining Useful Lifetime (RUL) of an asset?
The "standard” data science approach of observing many identical components repeatedly again over their life cycle has led to very good results for aircraft turbines. Unfortunately, the industrial environment simply lacks the ability to observe many assets over their entire life cycle again and again. To derive a statement that the asset will last another 5 years, 4 months and 24 days is a widely desired goal, but an unrealistic expectation.
With a lot of data, however, one can derive a similarly useful statement, namely: that a failure is not to be expected within the next 12 months.
The basic idea behind it: When we started our measurement records, we took the unique fingerprints of several hundred assets. For some of them, we could measure how “failure” looks like in the ultrasound data, but the majority of them will run without problems for another several years. These huge sets of fingerprints and durations of “trouble-free operating” allow us to derive a health status with one measurement only, that the asset will most likely last for at least another “X” months.
Every company constantly needs to evaluate which technological innovations it should introduce rather early and which it should introduce rather late, or not at all. That is why I would like to finally answer the question:
What am I missing as a company if I don't jump on this ultrasound "train" or only jump on it later?
With the concept of "virtual assets", which is possible with ultrasound and Artificial Intelligence, it is very easy to find an entry point into predictive maintenance. However, the sustainable competitive advantage of a company will be achieved by enriching these general data with measurements of its very own asset in order to make the estimation of the Remaining Useful Lifetime (RUL) more accurate. Companies that stick to the tradition maintenance concepts therefore miss the chance to digitize the exact fingerprints of a "breakdown" or "new condition" of their equipment. These opportunities are really irretrievably lost, since the degree of wear and tear of assets cannot be artificially accelerated over the years and the next "chance" sometimes only comes back in many years or even decades. Before any exchange of spare parts, companies should therefore ask themselves the question:
Am I missing a unique opportunity to digitize the health status of my assets?