Can you present yourself and your background ?
I am Director of Research at CNRS, the French National Research Centre. My research area is The optimal control of dynamic systems. I graduated from the Grenoble-INP university and from Sup’Aéro. With the endorsement of the CNRS, I became Amiral Technologies’ scientific advisor while keeping my position as a Director of Research in my lab (Gipsa-lab, a research unit that belongs to both the CNRS and the University of Grenoble-Alpes).
My work covers both theoretical aspects such as proof of stability, convergences of algorithms, dynamic inverse problems, nonlinear predictive control; as well as applied aspects in various domains such as micro-grid management, intelligent districts, cryogenics, wind and hydraulic turbines, combined cancer treatment, etc… Most of my work is undertaken with industrial partners (Schneider Electric, CEA, IFPEN, GE, SEB, INSERM).
I think you are starting to get the picture: I have a dream job with the lowest drudgery and a maximum level of satisfaction. I am aware of my privilege and remember it every morning.
Tell us about Amiral Technologies and what makes it so special?
Amiral Technologies has been granted an exclusive licence to exploit some of our CNRS research-based algorithms, to develop a market solution for industrial failure prediction.
More specifically, these algorithms consist in an automatic generation of mathematical properties that describe time-dependent data series originating from an equipment which behaviour is governed by laws, and that are captured by physical sensors. These properties are extremely rich to the point that they significantly increase the odds of finding solutions to problems such as the detection of signs of failure, ageing or nearing end-of-life for an equipment. The automatic generation of properties (or features) is similar to having a large number of photographs of the time series. These photographs, if obtained on a sufficiently representative number of healthy equipment data, allow to shape a robust and rich view of what normality looks like. This allows detection of anomalies with quite a high precision and a low rate of false alarms. To make things simple, we can say that these photographs represent the fingerprints of normality. It is the abundance of discriminant properties the makes the identification so reliable.
This is what allows Amiral Technologies to develop a failure detection solution without the need of historical data of failure. The industry (luckily) does not have so many occurrences of failures to learn from.
Why have you chosen to work with machine learning?
To say the truth, I have not specifically chosen machine learning. Machine learning is just a tool among others. These tools help exploring possible relations between variables to reveal the internal rules governing the dynamic system.
We however need to use these tools with caution, keeping a critical mind. Some find in Machine Learning the solution to all problems, others reject it in principle to avoid feeding a dogmatic view that they believe is just a passing and misleading trend. I believe that we should avoid these two extreme stands. These tools, when successful, use data to give an answer to a question. We still need to know which question we should ask and which data to involve in order to address it successfully. Generally, raw data coming out of physical sensors is not the one to present to these tools. This is where using a good feature generator such are Amiral Technologies’ becomes essential.
DiagFit, what is it?
DiagFit is the name of the software developed by Amiral Technologies that incorporates the above-mentioned innovation in an optimised and generic processing chain.
This processing chain starts with sensor generated data, goes through the search of a predictive model and ends by delivering prognostics on the aspects mentioned above (defects, anomalies, ageing, end-of-life). A permanent iterative adjustment phase (as new data kicks-in) closes the process and loops back to the prognostics phase.
Do you believe that scientific research is really making progress in this domain?
There is a lot of noise in this area, a lot of announcements and false promises. This is due to the economic stakes involved and to the competitive landscape, both resulting in an appetite for pushing solutions that scale to the extreme. This makes things difficult when one wants to analyse existing solutions objectively and to uncover the underlying scientific content. Things get even more complicated by the fact that most solutions address other parts of the value chain such as data acquisition, transmission and processing (computational architecture, networks, IoT, etc.) and not necessarily the efficiency of data processing and the predictive algorithms.
So, to answer the question: yes, scientific research is in constant progress, but the conditions to take it to market goes through a link with the whole of the value chain and through tight collaborations with visionary pilot customers.
What is your vision for the future of Amiral Technologies?
Most predictive maintenance start-ups announce algorithmic miracles before even confronting their technology to real and diverse use cases.
Amiral Technologies has accumulated a significant experience on concrete industrial cases. It is eventually the market feedback that will determine the algorithmic, architectural and strategic options that will prevail for data-based industrial maintenance.
Amiral Technologies’ offer as designed at the time the company was created had already pretty much evolved from the first idea that we had with my friend Katia Hilal, three year earlier, when we first started to question the industry players. It is very likely, and almost certain, that in three years’ time, it will have a shape and a value proposition that will be even more different. This is why I believe that the key words for success will be: Technological watch, tight collaboration with customers, constant inventiveness and systematic deep analysis of the accumulated case studies in order to unveil hidden genericities in an iterative process.
My wish is to see Amiral Technologies recognized for its values of seriousness, integrity and credibility. I hope that Amiral Technologies’ name will be associated with a fine and delicate mix between technical inventiveness and scientific maturity which does not oversee nature’s laws but makes the most of them by extracting the maximum of intuitions and incorporating the results in its roadmap.
In a few years we will make a point that Amiral Technologies has never promised to build a reliable digital twin by only exploiting a few data and without even an hour discussion with specialists.
I have no doubt that Amiral Technologies will become leader in reliable diagnosis systems, with a business model that is tightly linked with the value that it truly brings to its customers.
I even take the bet that one day, customers will invest time and resources to help us bettering our algorithms, thresholds and adaptation rules, to better fit with their industrial specificities.
I am sure that that day will come because a reliable predictive maintenance is not an option that an industrial can discard. The hesitation we see today over investing in this field is only the consequence of past false promises, which have set the bar quite low.
Fortunately, the unavoidable increase in the number of attempts to put different solutions and offers in place and the critical view that we will have on their real efficiency will definitely clarify the landscape. Amiral Technologies will certainly not be the one to complain!!