Space data with earthly applications
Durability that’s out of this world: Proekspert data scientists work to make European Space Agency equipment last longer in space.
In this article, Tanel Peet, a Proekspert data scientist, talks about Proekspert’s work with the European Space Agency.
Longevity in Space
Proekspert’s project’s main goal was to test artificial intelligence algorithms to help monitor the condition of reaction wheels, all toward the goal of potentially increasing the lifetime of mission critical assets in space.
What is a reaction wheel?
A reaction wheel is a fly wheel which keeps a spacecraft stable or helps it maneuver by changing attitude when required. At least three are needed to keep spacecraft stable and maneuver and, in practice, four wheels are on board each spacecraft.
Reaction wheels are ripe for study, because they’re a mission critical part and are constantly rotating – up to 4,000 revolutions per minute in the case of the spacecraft Proekspert studied. Some reaction wheels have been in orbit for over 20 years, and since they’re always working, and because there is some data available, they’re an aspect of space missions of keen interest to data scientists who wish to explore how their tools can enhance engineering quality.
Engineering in space
Generally speaking, the engineering quality of space tech is excellent. However, when missions successfully last twice their planned lifetimes, problems may start to appear.
Our data scientists started to analyze erratic behavior of ball bearings in reaction wheels. One type of such behavior observed on a mission is a phenomenon called cage instability. The wheel suddenly begins to vibrate – so much that it can be heard – and then suddenly returns to its normal stable humming sound. No higher education is needed to understand that something is wrong.
This vibration has quite a few negative effects. First, the scientific or commercial tasks of the space mission may be disrupted. Second, additional energy is required to maintain the desired wheel speed. Most troublesome is that the extra energy is turned into heat which depletes oil. And lack of oil causes additional cage instability. It’s a downward spiral that will potentially end in with the failure of the reaction wheel – if you don’t intervene.
Just as in health care, there are two ways to deal with the health of reaction wheels: treatment and prevention.
Treatment – condition monitoring
Treatment requires understanding the symptoms – the condition of the reaction wheels. Spacecraft constantly send telemetry data to the ground station, where operators monitor the condition. For a single spacecraft there can be tens of thousands data points flowing in, each generating hundreds of values per hour. Human operators cannot manually cope with so much data, and therefore clever software solutions are helpful.
For Proekspert data scientists used case histories from two space missions made available to study. Based on the reaction wheel cage instability examples, our team developed a machine learning tool for detecting weird behavior and potentially interesting symptoms from over 15 million data points.
In order to use the algorithm, a set of telemetry data from a healthy reaction wheel is needed. The algorithm then learns how a reaction wheel is expected to behave, after which it can let engineers know when something out of the ordinary is happening. From there, the engineers can take over and do the interesting work of making a diagnosis and finding a treatment.
Prevention – quality assurance
Prevention involves performing thorough health checks before reaction wheels are launched to space. Just as with telemetry data, a lot of data are produced in the process. The numbers are turned into plots which are manually viewed and compared by engineers. This process can be compared to looking at x-ray and CT scan images for detecting lung cancer.
A challenge for our data scientists was that we did not possess the data signature of a faulty reaction wheel, due to good engineering quality. But we did have around 100 signatures from healthy reaction wheels from micro-vibration measurements.
Instead of directly classifying faults, a method was developed which allowed the automatic comparison of the difference between two measurements. For this, a deep learning solution was used, for which a huge number of measurement signatures were synthesized. The solution has the potential to automate a large part of health checks performed on the ground.
While numerous other techniques were tested throughout the project, these two were the most promising ones. Since finding unexpected events from the telemetry data is already an active subject of study inside the ESA, we are concentrating on maturing technologies for the quality assurance part. By introducing some additional sensors, this technology could be used on future space missions to perform in situ health checks.
While these technologies are useful for the space industry, we encounter similar problems here on earth. There are many industries that need to keep rotating machinery working in remote locations, because losses mount when equipment fails. Proekspert is planning to collaborate with our partners to bring what we’ve learned in space back to earth.
Receive our weeky newsletter! Inspiring ideas that are worth your time