Multi Sensors Diagnosis and Automatic Learning

PATTERN RECOGNITION AND AUTOMATIC LEARNING
IMULTISENSORS DATA ANALYSIS : INDUSTRIAL APPLICATIONS
A HIGHTECH FRENCH START-UP : ON LINE MULTISENSORS DIAGNOSIS

PATTERN RECOGNITION AND AUTOMATIC LEARNING
1992-2002

To launch an ambitious scientic program for intensive software emulations of automatic learning tasks in “artificial perception”, I had set up in 1992 a broad scientific strategy, based on a creative mixture of three ingredients : rigorous mathematical theories of learning and pattern recognition, such as Support Vector Machines and Vapnik's theory, computer intensive validations on concrete industrial applications, large scale simulations on massively parallel computers.
This approach required short-term R&D industrial collaborations, and long-term scientific partnerships with well subsidized state research institutions, as well as with the R&D divisions of major high tech french companies.
Adequate massive parallel computing hardware was for instance incredibly costly, and required solid financial support to hire high level specialized computing engineers. This led me in 1992, after six years of theoretical work on formal neural networks and automatic learning, to create MIRIAD Parallel Processing, a self-supported advanced consulting group, gathering on a part-time basis a dozen of young Phds listed in the preceding paragraphs, most of them applied mathematicians then linked to CNRS, ENS Ulm and Cachan, University of Paris 11, Polytechnique. Our initial cash investment of $ 10,000 was put up by myself and our scientific consultants. The MIRIAD consulting group from then on completely self-subsidized its own massive R&D developments until the end of 1999, when it then started to raise Venture Capital. .
As major founder and scientific director of MIRIAD Parallel processing from 1993 to 1999, I organized and led the realization of several dozens of advanced R&D industrial projects in multi-sensors digitized signal analysis by softwares based on formal neural networks and automatic learning algorithmics, and long term industrial collaborations with the R&D divisions of major french industrial partners : DRET, ETCA, CEA-DAM, CESTA, CNES, SEP, St GOBAIN, RENAULT, RHONE-POULENC, LOREAL, LVMH, DIOR, LESAFFRE, IFTH, SAIC,
GDF, MARTELL, Fr.TELECOM, ALCATEL, etc
The analysis of papers by MEAD / VITTOZ on electronic retinas, and our encouraging results on “retina like” multilayer neural arrays with connection architectures invariant by translations and rotations, led me to set up in 1994 a three years collaboration with Center of Atomic Energy to conceive and implement a “realistic” modelization for an actual large size intelligent electronic retina dedicated to real time detection of moving objects on infra-red images. This SIREDIN project was realized by MIRIAD consulting goup under my scientific direction, coordinating a team of PhDs in advanced electronics, massively distributed computing, applied mathematics, artificial vision (P. GARDA, E. BELHAIRE, O. CHERIF, J. LACAILLE, S. GERVAIS, L. YOUNES, C. LABOURDETTE). The project, boosted by active technical collaborations with advanced hardware labs such as LETI in Grenoble, and CESTA in Bordeaux, succeeded in creating three rough but complete interacting simulators : one micro-simulator for the infra-red retina, one macro-simulator of the surrounding animated scene, and one algorithmic simulator of the vision tasks performed by a rotating “camera eye” and by the retina . Actual emulations of the retina real time distributed vision algorithmics were simulated on SYMPHONIE, a powerful massively parallel computer created by CEA. The simulations were highly successful in terms of computing speed and accuracy of moving hot spots detection, but of course, our team never knew if such experimental and quite costly retinas were or would ever actually be constructed in France.
In 1994, MIRIAD undertook another three years seminal resarch project for the french spatial agency CNES [remote multisensors monitoring of the first ten seconds in the actual launchings of the ARIANE rocket] and aerospace industrial partners (SEP/SNECMA). This allowed us to conceive, implement, and validate innovative adaptive techniques dedicated to intelligent remote monitoring of multisensors signals. The actual CNES / SEP/ SNECMA industrial goal was to analyse in real time the dataflow recorded at 1000 Hz by 25 sensors embarked on board ARIANE, such as oxygen and hydrogen flow rates, pressures, temperatures, through a fast succession of 6 phases of launching, in a highly transitory regime, in order to compute online diagnosis of potential anomalies.
MIRIAD scientific strategy, elaborated in tight scientific collaboration with O. CATONI, was based on empirical modelization by robust “neural networks emulation” of multisensors functional relations, was based on an automatic extraction of critical instants (launching phases transitions) and phase by phase off line automatic learning of these multisensors functional relations, by sophisticated formal neural networks.This modelization was implemented for CNES/SEP by using about 70 artificial neural networks (multilayers perceptrons). The optimal dimensioning of these deterministic neural networks and their accelerated learning was achieved by Miriad innovative initialization and fast gradient descent algorithms.
The key viewpoint I had introduced here was to focus automatic learning on empirical approximate modelization of normal behaviour of rocket engine subsystems, since the the effective engine anomalies were much too rare to enable classical supervised learning on a training set of actual anomalies.The multisensors normality diagnosis invented by MIRIAD were generated by efficient on the fly analysis of probabilistic likelyhood for actually recorded multisensors dynamics, equivalent in this case to estimating roughly many smaller dimensional marginals for the probabilistic joint likelyhood of 25 complicated curves with high dimension. This enabled a phase by phase sophisticated computation of probabilities of normality, by evaluating online the likelyhood of mathematical offsets between observed curves and their neural nets estimates. This stochastic approach relied on fast and robust nonparametric modelization of probability distributions for sets of random curves, based on a moderate size base of observed realizations for these random curves. Theoretical statisticians are quite familiar with the intrinsic mathematical complexity of this key problem in empirical stochastic modelization, since one is here confronted with the celebrated statistical “curse of large dimensions”, which casts its doom on attempts at fitting complex statistichal models for very high dimensional stochastic vectors. In our rocket engine study, we thus discovered a fairly sound mathematical philosophy for a pragmatic attack on high dimensional empirical modeling, and implemented efficient algorithms and methodologies to circumvent this basic theoretical obstacle.

IMULTISENSORS DATA ANALYSIS : INDUSTRIAL APPLICATIONS
1992-2002

As scientific director of the MIRIAD consulting group, I could rely on an unusual team of gifted collaborators, and from 1995 on, the oustanding long term contributions of J. LACAILLE and O. CHERIF played critical parts in our algorithmics and softwares production. From 1992 to 1999, MIRIAD thus realized an impressive list of major R&D industrial applications:
Boltzmann machines simulations on massively parallel computers DRET; ETCA 92-94
Boltzmann machines simulations on dedicated electronic devices DRET 93-94
Adaptive intelligent retinas : massively parallel simulations DRET 93-94
Artificial retinas : adaptive algorithmics and low level vision CEA-DAM 94
Modelization/Massive simulations of infra-red intelligent retina CEA; CESTA 94-96
Shape recognition by wavelet analysis and formal neural nets CEA-DAM 94-96
Neural nets diagnosis on bench tests for rocket engines SEP 94
Neural net engine diagnosis for ARIANE rocket launchings CNES; SEP 94-96
Neural nets classifyers and nonparametric statistics RENAULT 94-98
Neural nets diagnosis for detection of car drivers hypovigilance RENAULT 94-97
Neural net emulation of slow regimes for car engines RENAULT 95-96
Neural nets monitoring of production quality for chemical plant RHONE POULENC 94-98
Neural net emulation of gaz compressors dynamics GDF 98
Neural nets classifyers for sismic events discrimination CEA-DAM 98-99
Data mining of tests protocols for evaluation of medical impact LOREAL 97
Multi-sensors discrimination between biological dynamics LOREAL 97-98
On line product quality monitoring for a US chemical plant RHONE POULENC 98
Data mining for high throughput robotized tests screening RHONE POULENC 98
Multi-sensors search for defects root causes in glass production St GOBAIN 94-97
Multi-sensors search for defects root causes in cosmetics plant LVMH; DIOR 97-99
Multi-sensors search for defects root causes in textile plants IFTH; SAIC 97-99
On line anticipation of quality drifts for metal parts press shaping RENAULT 94-95
On line anticipation of critical defects in yeast production INRA; LESAFFRE 97-99
On line anticipation of critical fermentation events GDF; MARTELL 98
On line anticipation of bottlenecks in cellphones traffic Fr TELECOM / ALCATEL 98-00

A HIGHTECH FRENCH START-UP : ON LINE MULTISENSORS DIAGNOSIS
1999-2004

After five years of successful R&D industrial collaborations, I presented ANVAR (french national innovation agency) in 1997 with an innovation plan dedicated to the realization of Miriad Process, an advanced desktop software for intelligent analyzis of multisensors data flow, based on formal neural networks, automatic learning, non parametric statistics, and sophisticated multiscale signal analysis. We obtained an ANVAR grant, and MIRIAD completed this prototype software realization in 1998, thanks to outstanding scientific contributions by J. LACAILLE and O. CHERIF, who headed our software development team. To industrialize our software, major financial support was still needed, and in 1999, I took a five years leave from my job at Ecole Normale Sup., to transform our consulting group into MIRIAD Technologies, a startup company which, after recruiting a professional CEO, and two exceptional CTOs (O. CHERIF and J. LACAILLE), raised between the end of 2000 and 2004, a capital of 16 million dollars through venture funds linked to major french banks. Under my scientific direction, MIRIAD went on first to develop desktop scientific softwares [Md Scan, Md Monitor], and later in 2002-03, the Md Intelligence suite of auto-adaptive software components, dedicated to on line decision support by technical dataflow analysis. The main scientific goal was to enable fast implementation and validation of auto-adaptative algorithms dedicated to multi-sensors “intelligent” dataflow analysis. MIRIAD thus began a long series of industrial applications for telemonitoring of industrial processes, with major industrial customers or partners such as AIR LIQUIDE, RHODIA, AIRBUS, THALES, ALCATEL, ASTRIUM. Our on line process monitoring applications provided automatic decision support for forecasting incidents on gaz and fluids distribution equipments, for anticipating critical events on cellphone networks , for estimating quality risks in chemical production plants, for early warning of defectivity drifts in electronic chips production plants. MIRIAD software solutions provided also on line telediagnosis for industrial equipments, to implement real time detection and classification of complex patterns on live recordings by high rate sensors, such as electric voltages and intensities, acoustic signals, etc. MIRIAD sophisticated mathematical classification techniques involve complex signal pretreatments such as sliding multiscale signal analysis by wavelet decomposition or transitory fast fourier transforms, for online extraction of time indexed high dimensional feature vectors. Self-calibration of internal pattern classification algorithmics by non supervised automatic learning rely on information theory techniques to select the most significant signal features, and on smart non parametric empirical modelization for the joint distribution of significant signal features. The practical goal in most of MIRIAD industrial diagnosis software solutions is to automatically generate pertinent “signatures” for multisensors patterns, through intensive off line “learning” from historicized dataflow, in order to implement on line robust recognition of fuzzy signal shapes or deformable multisensors patterns. For instance MIRIAD developed real time automated diagnosis software solutions for automatic real time identification of undesirable sound sources on board airplanes , for automatic detection of celebrated 27 Chartist Figures on massive stockmarket dataflows, for on line detection of abnormal voltage patterns in electricity distribution to industrial plants, for on line detection of welding incidents for robotized welding

References
Patent 1 : A method for anomaly detection (inventor R. Azencott)
Non supervised algorithmics for automatic learning of critical features and anomaly detection in signals
Patent delivered : Europe N° 00 410 009 5 ; USA N° 09 519 597 2003

Patent 2 : A method to monitor process performance indicators (inventor R. Azencott)
Auto-adaptive computation of global performance scores, by fusion of heterogeneous performance indicators
Patent delivered : Europe N° 00 410 069 9 ; USA N° 09 784 134 2003

Patent 3 : A system for automatic real time identification of specific sounds (inventor R. Azencott)
Live audio-surveillance systems for buses, trains, airplanes, by automated sound identification software,
Patent delivered : Europe N° 03 054 14 2003

Patent 4 : A technique for geometric shapes identification of curves (inventor R. Azencott)
On line image analysis to retrieve specific fuzzy geometric shapes, with offline self calibration learning.
Patent delivered : Europe N° 00 410 036.8 ; USA N° 09 616 055 2003

Patent 5 : A system for content driven navigation in multimedia databases (inventor R. Azencott)
Automatic content indexation algorithms and softwares for large databases of images and texts
Patent delivered : Europe N° 00 410 103.6 ; USA N° 09 783 621 2002