We just published a new white paper in our document library « DiagSign Automatic Feature Generation and the State of the Art ». This paper summarizes the state of the art in terms of academical methods to solving the problem of feature generation from time series. These can be classified in two categories : Structured models approaches, and Dimensionality reduction approaches. DiagSign is a protected invention that was born in the labs of the French National Research Centre (CNRS). It uses a purely mathematical method to automatically generate features from time series that are relevant, highly discriminant, and very adapted to transitional signals. DiagSign can generate as many features as needed, possibly thousands, in record time and in a massively parallelizable way. In addition, it is agnostic to the type of signal nor does it care what equipment is diagnosed. The white paper explains why Amiral Techologies’ DiagSign surpasses these academical methods and avoids their drawbacks : It does not assume an a priori structured model; It keeps all the information in the signal; It offers tremendous and low cost scalability. Read the paper here
What is unsupervised learning? Unsupervised Learning (see wikipedia) is used when no historical data is available that involves past faulty behavior and/or aging progression. Therefore, Unsupervised Learning is mandatory to design data-based alarm system and aging monitoring algorithms in the majority of industrial situations. This is at least true in the current state of data availability and labelling. Typical solutions to this problem need a so called normality space to be defined. This is the space of configuration of features that characterize the healthy behavior of the equipment. Roughly speaking, when the features leave the normality space (in some sens), an alarm can be raised. As it is always the case in Machine Learning (ML)-based learning, the main questions are: What are the features to use in the previously described process? What is the impact of this choice of features on the quality of the resulting alarm system in terms of coverage rate and false alarm ? An obvious partial answer to these questions is that if the set of features being used does not intersect with the set of features impacted by a specific potential default, then this default will not fire the alarm system. Saying it differently, this fault will not be covered. Consequently, the more discriminant features one involves in the definition of the normality space, the more chance one gets to cover the multiple possible failures in the equipment. Why is Amiral Technologies’ solution for unsupervised learning so powerful and innovative? This is precisely why the Unsupervised Learning solutions of Amiral Technologies outperform alternative solutions. Indeed, Amiral Technologies’ Automatic Feature Generation enables very rich and highly discriminant sets of features to be generated. This leaves small chances for faulty behaviors (even unseen in the learning data) to remain undetected. Moreover, the availability of such a high number of features enables the definition of voting systems that reduce the risk of false alarms as these can rapidly render the solution unacceptable by the practitioners and the maintenance operators.
A l’occasion de la sortie du rapport Villani sur l’Intelligence Artificielle, j’avais donné mes premières impressions en tant que fondatrice de startup. Vous trouverez l’intégralité de cette reflextion ici