In recent years we have seen an increase in the application of Machine Learning / Artificial Intelligence to predictive maintenance in industrial settings. And the consensus seems to be that AI is here to stay and that there is much value to be created.
However, as with any radically new technology, it’s adoption has been slow and many examples of practical AI applications have failed to create real impact and actual bottom-line value.
Today we will introduce some of the challenges in creating tangible value with AI in industrial applications and how to overcome them. We will also present our approach to predictive maintenance, its benefits and the advantages that we offer to our customers. Let’s get to it!
There are two types of domain knowledge needed to develop ML-based solutions successfully. On the one hand, there is the technical and operational knowledge which is about the ins-and-outs of the machines/processes and on the other hand there is the AI ML knowledge required to obtain the right modeling approaches.
Jungle is mainly composed of electrical engineers that speak machine who are also true data science professionals. This is a rare and powerful combination to have, but a crucial one in delivering meaningful AI-based solutions.
But be aware! There is also a pitfall involved here that can be easily missed: human bias and pre-existing beliefs. Even experts can inject too much prior knowledge and preclude better solutions from being created. Later on, we explain how we have avoided falling into this pitfall.
Most companies are now sitting on multiple years of untapped operational data. However, ML models were not envisioned as a direct application when these systems were built. Hence, they were not also set up while having failure/maintenance annotations as an essential ingredient. As a result, maintenance logs don’t have a standardised format or descriptions as their primary users were the maintenance teams themselves. Unfortunately, annotated examples are crucial to train AI models to identify failures.
Most common AI strategies overcome this drawback by requiring experts to annotate periods in which the assets were behaving unexpectedly. This, however, may quickly become a Herculean task since end-users would be necessary to comb several years of highly non-linear and correlated multivariate time-series data. Furthermore, this task assumes that the end-users could identify precisely the periods in which the assets are working unexpectedly. If this was quickly done, would AI models be required in the first place?
Furthermore, most of the available data is of regular operation. If we focus on the data leading to failures, we will end up using only a small subset of the available data. It is wasteful to neglect more than 90% of the available data.
Despite the many years and experience of the operational teams, it’s still difficult (if not impossible) for them to know if their processes and machinery are operating as they should. Nowadays, machines are so complex that we cannot create a mental model that encompasses all the dynamics and interactions relevant to their operation.
The electromechanical machines in industrial production lines, wind turbines, etc., are the real black-boxes.
Explainability is a crucial aspect when it comes to AI solutions. Why is the model saying this? The goal of using AI is to help the situation not to replace one black-box with another one. We would end up with two black-boxes to worry about: the real-machines and the AI models!
AI-based solutions need to accelerate and increase actionality. Operators, when seeing alarms from the AI system, also need to get the right context to, together with their domain knowledge, make data-driven decisions.
At Jungle, we avoided all the above challenges by doing things a little differently. We develop AI models that can understand what the normal operation of complex machines should be! This means that we can leverage all the historical asset data, and no one needs to go over the daunting task of manually identifying periods of (ab)normal operation to train models.
We learn all the dynamics present in the normal operation of electromechanical assets directly from their data, without making assumptions.
Our modeling strategy even goes a step further. Most common AI approaches use the data of several machines and train a single AI model. The resulting model is, however, an average of all the assets that does not have a physical equivalent.
Common sense tells us, however, that even turbines of the same model, for example, do not behave precisely the same way. What is considered a normal temperature for one turbine can be slightly off for another. So even though they share most of their dynamics, each one has its personality, so to say. Our models make use of this. They extract the general dynamics from the available assets but also allow each one to have their normality levels.
In practical terms, our models output expected ranges of operation for all sensors, as shown in the image below. We see that our normality model successfully leaned the non-linear relationship between different wind turbine sensors. The model predicts (indicated by the shaded areas) where the sensor measurements should fall in.
We can also see that our model can track the normal operation of this turbine regardless of the operating conditions: whether we have periods of low or maximum power production we can always know how the machine should be behaving. This can be thought of as a dynamic threshold that is automatically set by the models according to the specific conditions of operation of the turbines.
What about when things do deviate from what they should be? This is when we start reaping all the benefits of these models by providing extra context!
We have devised a powerful alarm mechanism that warns users for relevant deviations. Using these alarms, we have successfully identified dozens of failures of different nature (temperatures, pressures, etc.) in various applications such as heavy industry machinery and wind turbines.
But this will be the main focus of a future blog post! We will cover:
Did we get you curious? Ok, a small sneak peek then!
Below we show our system automatically detecting an oil system leakage in a wind turbine. Our alarms detected the occurrence months in advance while the SCADA alarm system did not provide any relevant information/alarm during the months preceding the maintenance intervention.
By merely looking at Jungle’s alarms, we can already get a clear timeline of the failure! The oil pressure had been dropping slowly over time since September. At a particular moment, one of the gearbox bearing temperatures started going hot due to the cooling system oil inefficiency.
If you identified yourself with our mission and approach to predictive maintenance, feel free to reach out! Let us help you conquer complexity! Send us a message to firstname.lastname@example.org, and we will be in touch shortly after! It’s that easy to get started.
Article also visible on medium.