Method for configuring a sensor detection device, and corresponding computer program and adaptive device

Patent metadata

Field	Value
Publication number	US-9524467-B2
Application number	US-201214342205-A
Country	US
Kind code	B2
Filing date	Aug 29, 2012
Priority date	Aug 29, 2011
Publication date	Dec 20, 2016
Grant date	Dec 20, 2016

Abstract

A method for configuring a device for detecting a situation from a set of situations where a physical system comprises the following steps: reception of a learning sequence corresponding to a given situation of the physical system; determination of parameters of a hidden-state Markov statistical model recorded in the detection device and relating to the given situation, on the basis of a prior initialization of these parameters. The prior initialization comprises the following steps: determination of a plurality of probability distributions from the learning sequence; distribution of the probability distributions between the various hidden states of the statistical model in question by global optimization of a function of adaptation of these probability distributions to the various hidden states and to impossible transition constraints; and initialization of the parameters of the statistical model in question using given representative probability distributions for each hidden state of the statistical model in question.

First claim

The invention claimed is:

1. A method, implemented by processing circuitry of a detection device, for configuring the detection device for detecting a situation from a set of situations wherein a physical trait of an object is observed by at least one sensor, comprising the following steps:
reception, by the at least one sensor which is physically attached to or proximal to the object, of a sequence of observation data for the object, referred to as a learning sequence, and corresponding to a given situation of the object,
determination, by the processing circuitry of the detection device, from the learning sequence, of parameters of a hidden-state Markov statistical model recorded in storage means of the detection device and relating to the given situation, by prior initialisation of these parameters, and then updating of these initialised parameters,
wherein the parameters of the hidden-state Markov statistical model relating to the given situation include a matrix (a i,j ) of transition probabilities of each hidden state i towards each other hidden state j of this hidden-state Markov statistical model,
wherein the prior initialisation comprises the following steps:
the statistical model in question comprising a given number Cn of hidden states, determination, by the processing circuitry, of a plurality of Ln probability distributions from the learning sequence, by dividing the sequence into Ln sub-sequences and allocating to each sub-sequence a probability distribution that models it statistically, the number Ln of determined probability distributions being greater than the number Cn of hidden states of the statistical model in question,
distribution, by the processing circuitry, of the Ln determined probability distributions determined between the Cn various hidden states of the statistical model in question,
determination, by the processing circuitry, for each hidden state of the statistical model in question and from the probability distributions allocated to this hidden state, of a single probability distribution representing this hidden state, and
initialization, by the processing circuitry, of the parameters of the statistical model in question from the determined representative probability distributions,
wherein, the statistical model in question further comprises impossible transition constraints, which correspond to coefficients of the matrix (a i,j ) of transition probabilities set to zero, between certain hidden states,
wherein the distribution of the Ln determined probability distributions between the Cn various hidden states of the statistical model in question is done by global optimisation of a function of adaptation of these Ln probability distributions to the Cn various hidden states and to the impossible transition constraints, said function of adaptation including a term relating to the probabilities of transition from one state to another for each of the Ln sub-sequences with respect to a next one of the Ln sub-sequences, and
wherein the method further comprises a step of configuring the detection device so that the statistical model in question includes the parameters determined by said prior initialisation and then said updating.

2. The method for configuring as claimed in claim 1 , wherein the step of distribution comprises the execution of an iterative K-Means algorithm on a number of classes equal to the number Cn of hidden states of the statistical model in question, this iterative algorithm comprising, at each iteration:
for each of the Ln probability distributions determined from the learning sequence, the association of this probability distribution with one of the classes, this association using the Kullback Leibler divergence and the impossible transition constraints, and
the calculation, for each class, of a probability distribution representing its centre.

3. The method for configuring as claimed in claim 2 , wherein the step of distribution comprises an initialisation of the iterative K-Means algorithm consisting of:
sorting the Ln probability distributions in the sequential order of the sub-sequences with which they are associated in the learning sequence,
distributing the Ln probability distributions sorted in the classes in this sequential order, from the first to the last class,
for each class thus initialised, determining a probability distribution representing its centre.

4. The method for configuring as claimed in claim 2 , wherein the function of adaptation of the Ln probability distributions to the Cn various hidden states and to the impossible transition constraints of the statistical model in question is, for a given distribution of the Ln probability distributions determined between the Cn various hidden states of the statistical model in question, this distribution being in accordance with the impossible transition constraints, a sum of Kullback Leibler distances between each Ln probability distribution determined and each probability distribution representing the centre of the hidden state associated in this distribution.

5. The method for configuring as claimed in claim 2 , wherein:
the function of adaptation of the Ln probability distributions to the Cn various hidden states and to the impossible transition constraints of the statistical model in question is, for each distribution “a” of the Ln probability distributions determined from the learning sequence between the Cn various hidden states of the statistical model in question, a product between a function taking into account the Kullback Leibler divergence between each Ln probability distribution determined from the learning sequence and each probability distribution representing the centre of the hidden state that is associated with it in this distribution “a”, and probabilities that each Ln probability distribution determined from the learning sequence is associated with the hidden state defined by the distribution “a”, knowing the hidden state associated by the distribution “a” with the probability distribution preceding it in a given order of the Ln probability distributions issuing from the learning sequence,
the global optimisation of this adaptation function is achieved by execution of the Viterbi algorithm for the selection of a distribution that maximises it.

6. The method for configuring as claimed in claim 2 , wherein, each Ln probability distribution being a normal distribution, the probability distribution representing the centre of a class Ki is a normal distribution determined by the calculation of its expectation μ i and its variance E i according to the expectations μ i,j and the variances E i,j of all the probability distributions of this class Ki, as follows:

μ
i

=

1

Card
⁡

(
Ki
)

⁢

∑

j
∈
Ki

⁢

μ

i
,
j

⁢

⁢
and

Assignees

• Jallon Pierre
• Gris Florence
• Commissariat L Energie Atomique Et Aux Energies Alternatives
• Movea

Inventors

• Jallon Pierre
• Gris Florence

Classifications

• G06N7/00 Primary

  Computing arrangements based on specific mathematical models · CPC title

• G06N20/00 Primary

  Machine learning · CPC title

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• G06F2218/00

  Aspects of pattern recognition specially adapted for signal processing · CPC title

• G06F18/295

  Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models · CPC title