min read

Solving predictive maintenance challenges with Time-Machine AI models

Today we will go on a journey to discover some of the key advantages our Time-Machine AI models bring to predictive maintenance pipelines. They helped us solve some of the biggest challenges we have encountered on our way, so hold on to your seats!

Time-Machine AI models

We have developed an AI model that we call Time-Machine (TM) models. They allow us to overcome different challenges in the space where jungle operates: predictive maintenance solutions for electromechanical industrial assets. By the end of the post, their name will make a lot more sense ;)

Differently from most common standard AI models that learn a mix/average of normal behaviours contained in historical asset data, TM models learn all the individual behaviours that each sensor of each asset, e.g. a wind turbine, has had over its entire lifetime (see Figure 1).

Figure 1 - Jungle’s Time-Machine Unified model learn all historical normal behaviours for multiple assets and their sensors.

Time-Machine models can be thought of as archives of normal behaviours for electromechanical assets. We can ask the TM model to mimic how an asset would have behaved if it would operate similarly when compared to a certain period.

This simple but fundamentally better approach of TM models opens up a new entire world of exciting opportunities for us. Next, we will go over some of the major challenges they solve by design.

Solved challenges

Learning from all historical asset data

The normal behaviour of a wind turbine may change multiple times during its lifetime. Replaced and/or serviced components sometimes do not behave the same way as before. And this represents a huge logistical and AI challenge. A wind turbine has different “personalities” over time as shown in the figure below.

Figure 2 - A wind turbine may change its normal behaviour after interventions such as maintenance events (on the left) and software updates (on the right).

Figure 3 shows a couple of real examples to demonstrate the impact of such (more common than desired) events:

  • Sensor recalibration - at times sensors drift and require recalibration (top plot of Figure 3). This leads to AI models no longer following the new normal sensor behaviour.
  • Control changes - at times, wind turbines receive software updates leading to changes in their control parameters. For example, the rotor speed control of the turbine shown in the middle plot of Figure 3, was changed and it started operating with a higher maximum rotor speed. This led to a standard AI model to no longer represent the turbine for that regime of operation.
  • Maintenance interventions - on the bottom plot we see a maintenance event that led to oil pressures starting to operate at higher values. This would immediately trigger alarms since the sensor and model predictions no longer match.

Figure 3 - Three examples of events that lead to normal behaviour changes that were no longer captured by AI models trained with historical data. From top to bottom: sensor calibration, control changes and maintenance interventions.

For most times it is a good thing that AI models detect deviations from the normal operation - it’s their main purpose. But in the cases described above, since the new behaviour is the new normal the model should therefore be expected to adjust to it.  However, the model would need to be retrained in order to learn and follow the new normal behaviours! This would mean that the data before the maintenance intervention could not be used anymore. And this is a huge bottleneck and disadvantage of standard AI models.

Standard AI models can only use historical data that represents the most recent normal behaviour of each asset. Times may lead to wasting more than 90% of historical data.

It is cumbersome if at all feasible, to understand when the last normal behaviour started. In fact, this would require technical teams that understand the assets at a deeper level to waste countless hours labelling all historical data from a wind farm. This is simply not a viable nor a scalable solution. Our TM models solve this challenge by design. Since they learn all the historical normal behaviour automatically, we can simply leverage all years of stored data to extract the most from it.

Reduction of false positive/negative alarms

In case there are different historical normal behaviours, standard AI models will not learn them individually but they will learn a mix (a weighted average) of all of them as shown on the bottom left side of Figure 4. This will lead to large confidence bands of the AI model which will result in false-negative and late detections.

Since new normal behaviours can also significantly differ from past ones, models that learned an averaged normal past behaviour may also lead to false-positive alarms simply because they cannot follow the new dynamics.

This is a simple and self-explanatory example that goes against a common belief in the AI space: “The more data, the better!”. This is only true if you know what you are doing!


Usually, when a standard AI model is trained, users need to define a training period. And once trained, the AI model will make predictions based on what it learned from it. But it will act as a black box since it will not give any insights on why it’s making such predictions or let know what examples or periods are mostly being used to make such predictions.

Differently, users can specify exact periods that TM models should be used to make predictions (as shown in Figure 4). This will give them insight regarding the model predictions and allow for powerful queries as discussed next in the What-If scenarios section.

Our Time-machine models allow for the opening of the black-box letting our users know the exact periods the model is basing its predictions.

Figure 4 - Averaged historical behaviours lead to low and wide confidence bands and lead to false-negative alarms and late failure detections. Our TM models allow users to see predictions based on specific periods.

What-If scenarios

Since users can specify periods to make predictions, this allows for a powerful application of TM models: What-If scenarios. For example, our users can ask questions to TM models such as:

  • What would be the power production for this period, if turbine X would have the same normal behaviour from a year ago before a specific maintenance event/upgrade?
  • How would turbine X respond as if it would be turbine Y?
  • How would turbine X behave if it would have the same control upgrades that turbine Z had?

TM models can be thought of as advanced query databases that allow users to make sophisticated What-If scenarios.

These are extraordinarily power querying capabilities that allow our users to understand their assets in a much deeper and more powerful way! We will do a deeper dive into What-If scenarios in a future post, so stay tuned!

Continual learning

Usually, standard AI models are static after training and are not able to learn new dynamics from sensor data. Our models solve this big hurdle of AI models since they are able to expand their knowledge about the assets for an indefinitely amount of time.

TM models can keep on learning with new sensor data even after they have been trained with a large historical dataset.

Another big advantage of TM models is that they leverage the dynamics learned with the historical data. This way, our models are not learning from scratch the new normal behaviours, they are doing whilst using their previous knowledge. This leads to faster and robust learning of new normal behaviours.

Figure 5 - TM models have the powerful capability of learning future asset normal behaviours from new data.

Time consistency

Time consistency is another big advantage of TM models. Since it’s always the same model used for both historical data and future data, predictions for the original training data range will not be changed in any way (as shown in Figure 6). This means that there will be always a single ground truth of normality behaviour, differently from standard approaches where models with distinct training periods need to be stitched together - we called this the Frankenstein approach.

TM models allow our users to trust the predictions of our models and leverage the ever-growing “asset archive” that TM models represent.

Figure 6 - Our TM model has inherent time consistency, i.e. past dynamics and normal behaviours are not altered when new data is added to it.

Logistic nightmare

The capability that TM models have to perform continual learning allows us to also overcome a major logistic problem. Since they are capable of learning future new behaviours, we don’t need to train and maintain new models. Our TM model keeps on expanding its normal behaviour library, keeping our count of AI models to one. Pretty sweet huh?

Figure 7 - Our approach allows us to reduce tens of thousands of single-variate sensor AI models to a single and large Unified Model.


In this post we went of several revolutionary applications that Time-Machine models make possible:

  • Leverage all historical data of electrical assets;
  • Lower the amount of false-positive and negative alarms;
  • Possibility of creating What-If scenarios with interpretable predictions; and
  • Continual learning while preserving time consistency.

And all these features allow our users to understand their assets in a profound and actionable way.

In a future post, we will introduce how leveraging the entire historical data from a wind farm allows us to create an ever-growing library of failures. With it, we can provide our users with better and informed case detections. And with automatically created case descriptions, historical similar failures and actions to resolve them.

This is one of the key technologies that will enable our way towards prescriptive maintenance solutions.

Silvio Rodrigues

CIO & Co-founder

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.

Back to blogs