Frequently Asked Questions
1. Wind park experts complain that in most cases data ownership remains with producers and
they are likely not willing to share data. What is OST’s experience from the POCs and what
was the relation of the operator and producer?
OST Response: OEMs dont disclose their information for tactical reasons as they aim to get Full Service
Agreements. The data however still belongs to the owner of the equipment.
We know two types of FSA:
a) Energy performance based
b) Availability based
In the case of a windpark covered by a FSA, the Predictive Maintenance approach makes less sense and therefore, the sharing of data is less relevant. Nevertheless, we believe that it might be important for a windfarm operator to know the real production performance potential of the asset, typically in the case of:
a) post-COD, at least during the 2 first years of operation, to assess the real performance of the windfarm and consequently the profitability of the investment.
b) acquisition of a new windfarm as part of the due diligence process. Our experience has shown that symptoms of incipient failures are often manifested through power inefficiencies
c) repowering as part of the decision process.
For these cases, our solution provides a comparison between the real power curve and the expected one calculated by a machine learning algorithm. The data necessary to provide this analysis is generally available. Therefore, for a given wind condition, our solution calculates the brutto (total available energy production potential) and netto (in operation only) expected energy output, and compares it with the real production output of an individual wind turbine. This approach enables us to make a quantitative comparison between the real performance of wind turbine and its digital model calculated during a specific period of comparison.
2. What makes OST’s product better in comparison to similar solutions provided by other majors.
a) We are independent and our only objective is to provide the best independent information for our client.
b) Our unsupervised modeling approach demonstrates its effectiveness on real PoCs performed on wind turbines, with or without gearboxes
c) We can deliver a solution where strict data and digital knowledge ownership is guaranteed to our customers with full transparency. If required, our platform can be installed on a private cloud and the model configuration performed by our clients, even if they don’t have a specialist knowledge of data science.
d) We provide an end-to-end solution including unique state of the art wireless sensors/systems technology specifically designed for predictive analytics. This allows critical failure modes to be detected without significant additional hardware investment.
e) We can customise our solution and integrate it into existing IT architectures and provide additional mobile apps.
3. What are the expected energy inefficiency savings during the first and second year of operation of a new wind turbine?
Based on our PoCs,,approximately 4-8% of energy inefficiencies are detected (3.5GWh in 14 months on 4x 2.1MW). Our analysis has demonstrated that 50% of these inefficiencies could easily be recovered as they are linked to suboptimal tuning of the pitch and yaw control systems.
4. Does the OST system use supervised or unsupervised learning? It seems there is a lot of logged data to justify supervised learning. But also, there are new vendors/models where you may need unsupervised learning.
We believe that supervised learning is not optimal in the predictive analysis in maintenance and energy optimisation. By definition, the events are unpredictable and consequently difficult to train. Technical assets are often different and the data is limited or difficult to access. Therefore, we have decided to focus our approach strictly on unsupervised learning. We are moving one step further and can integrate our own sensor technology (produced by a major sensor manufacturer) to master the input data betterl. We are working on different cases, with high scalability, where we are developing a predictive analytics standard “package”. For example, we are currently developing a package with an elevator manufacturer. The package consists of 4 sensors, one gateway and our predictive platform. This package can be installed on 70% of existing elevators in less than 2h and is able to detect proactively 80% of the main failure modes. The expected results is 15-20% OPEX reduction in service. Other similar projects are under discussion (equipment in chemistry/pharma and power distribution).
5. What probabilistic model is used for training? Is there anything customised in the model, like a custom architecture, or it is an off-the-shelf model.
Our “maths” engine is based on a Bayesian Network. This “maths” engine is the same of all our models. Our software platform enables configuration of the models by application engineers, even if they have limited knowledge in data science. The accuracy of the model depends on the available parameters for a given system. The models are validated with our customers during the PoCs.
6. Is data fed in raw format or there are specific features that are manually crafted to improve precision / recall of the model?
We feed our models with structured data. We prefer a combination of several models designed with limited parameters instead of one single model designed with a large number of parameters. This approach provides better results from more robust models.
7. How much data is available to train the models?
Normally, we need sufficient data to cover the full operating range of a system. We are developing models consisting of a set of sub-models with limited parameters. Each sub-model is dedicated to detect specific failure modes or critical events. As the variability of the model environment is smaller, the training period is reduced and the robustness of the models is stronger.
8. How specific are the models in predicting anomalies in a specific vendor/model? If they are specific, how are new vendors on-boarded if the model does not generalise.
We focus our modelling by application where scalability is available. As already mentioned, the “maths” engine and the model configuration approach is the same for all our applications. Once an application is incorporated into our model library, we adapt the real models to specific pieces of equipment. For example, we have “standard” models for gearless wind turbines, wind turbines with gearboxes, reciprocating compressors, pumps etc. The model development is based on real world PoCs. The models can then be adapted easily for each specific equipment.
9. In the business model, one of the important advantages would be to log data and train a model that no competitor can outperform (due to lack of data). Is this part of the business model?
We already have this functionality. We made a comparative analysis on a case with a leading Hydro Power company in Switzerland and were able to detect an event 15 days before shutdown, whereas all other competitors discovered the event only 9 days before shutdown. The main reasons for our outperformance, is our unsupervised training approach combined with application knowledge expertise.
We think that our major advantages in the future will be the capability to collect the data from our wireless technologies cost effectively. This approach will provide us a key competitive advantage. Considering the elevator case described above, there is no business case (except maybe for the major lift manufacturers) if the solution doesn’t include cost effective data collection capability. For the elevator case, the cost of the hardware is not expected to exceed circa 600EUR. It should be noted that our wireless solution is unique. It offers the possibility to collect a large amount of data in a few seconds synchronously. This synchronisation allows us to easily correlate the data and therefore to identify anomalies easily and accurately.
10. Generally: OST has already quite a lot of POCs. Is there a strategy to build a blockbuster product for one specific application area first or does OST want to be active in as many as possible industries?
The reason of the large amount of PoCs was to understand the potential and robustness of the technology on real cases. It was also important to understand the market dynamic and the customer expectations. From our PoCs, we understood for example how to “package” our algorithms into a suitable product which meets our client’s requirements in terms of data and digital knowledge ownership.
We also understood the complexity of getting the right data. That’s why we have decided to create our own wireless system.
Today, we have the tools in place, which have been successfully tested in the field. Our next steps is to develop specific scalable “packages”; wind, refineries and specific equipment will be our primary focus.
Optimised Systems Technologies AG
6330 Cham, Switzerland.
Boulevard de Pérolles 7 , 3rd Floor,
1700, Fribourg, Switzerland.
+41 41 743 0728
+41 41 743 0729
Optimised System Technologies (Portugal) Lda.
Rua Dr. Antonio Candido 10-2.andar,
1050-076 Lisboa, Portugal
Gian Maria Philipp