TIMi formed the central foundation of the entire machine learning approach implemented in this project. The objective was to predict the failure states of SAFRAN industrial machines, equipped with a large number of heterogeneous sensors measuring their state at the millisecond level, over several years. This high temporal granularity, combined with the diversity of measurements, naturally led to the creation of very large databases.
Thanks to the TIMi suite, and more specifically Anatella, it was possible to design a complete data transformation pipeline in a simple manner, and then execute it quickly and at low cost, first on a laptop, then on a server. One of the major strengths of TIMi lies in the implementation of true multiprocessing solutions, allowing calculations to be parallelized without saturating memory and fully exploiting the processor's capabilities. In our case, this resulted in considerable performance gains.
An initial attempt based on Python proved to be poorly suited: analyzing a single file among millions required nearly two weeks of computation. In contrast, with TIMi and a properly sized server, the entire pipeline could be applied to the entire database in just one week. Furthermore, the use of the `.gel_anatella` file format significantly reduced storage space, while making the files more convenient to handle.
Finally, the implementation of Modeller was a decisive step. Much more than just software, Modeller is part of an innovative approach to data management, offering tools with high potential for technological disruption to design, under real conditions, machine learning models as efficient as possible. Review collected by and hosted on G2.com.
The software fully met my expectations in all areas and for all use cases I encountered. Review collected by and hosted on G2.com.
