Data processing: artificial intelligence – Neural network – Learning task
Reexamination Certificate
2000-05-01
2002-12-17
Black, Thomas (Department: 2121)
Data processing: artificial intelligence
Neural network
Learning task
C702S022000, C702S027000, C700S048000
Reexamination Certificate
active
06496813
ABSTRACT:
FIELD OF THE INVENTION
The present invention concerns pattern classification, notably in the field of the characterisation or the identification of odours.
BACKGROUND OF THE INVENTION
In odour characterisation or identification apparatus (so-called “electronic noses”) a sample of an unknown product is identified by classifying it in one of several classes which the apparatus recognises thanks to a preliminary learning operation. Apparatus of this type comprises two main parts, a data acquisition portion and a data processing portion. In general, the data acquisition portion comprises a sampling head enabling the sampling of volatile components which are given off by the tested product, and a battery of sensors to which the sampled components are supplied and which produce output signals which, taken together, are characteristic of the volatile components given off by the product and, thus, are characteristic of the product. In general, the data processing portion comprises automatic data processing means which record the sensors' output signals and perform analyses thereon. By detecting or analysing the patterns inherent in the different sets of data relating to particular products, the data processing portion detects the class of the product (which can correspond to the nature of the product, for example, “a wine”, “a whisky”, etc., or to a quality of a product, for example, “fresh milk”, “fermented milk”, etc.).
The goals pursued in acquiring and in processing the data are different depending upon whether the apparatus is in learning mode or in identification mode.
In the learning mode of the apparatus, the products from which the various samples are taken are known and are analysed several times. The processing methods which are used at this time are those which enable, on the one hand, the detection and characterisation of the different products and, on the other hand, the construction of a robust identification model. The methods used are generally statistical methods (Principal Component Analysis (PCA), Discriminant Factor Analysis (DFA), . . . ) and neural networks (see, for example, J. W. Gardner and P. N. Bartlett, “Sensors and sensory systems for an Electronic Nose”, Nato Asi, Series E, Vol. 212, pp.161-180 (1992)). In the identification phase, the origin of the samples is unknown and the goal is to identify this origin by using the data, and the model which has already been constructed.
The statistical methods used in processing the data recorded by an electronic nose are linear methods. The variables calculated according to such methods are so-called “synthetic variables” and convey two items of information, quantitative information, indicating the percentage of inertia in the initial cloud of points that is explained by this variable, and qualitative information regarding the structure and discrimination of the cloud which it discloses. By way of contrast, neural networks are among non-linear methods and are capable of learning the behaviour of the different sensors. The two types of methods directly employ the recorded raw data. In general, they are used separately.
Now, it is recalled that the expression “synthetic variables” refers to combinations (whether or not linear) of the measured variables. In general they are obtained by optimisation of a precisely defined criterion. The environment of this optimisation process can be conditioned by additional knowledge or variables. For example, in PCA a criterion is optimised without any additional information whereas, in DFA, there is available knowledge regarding each sample's affiliation to a group, and in PLS (“partial least square”) one or several additional variables are measured for each sample (for example, the concentration of a particular chemical element).
The document WO96/26492 describes a particular neural network structure capable, amongst other things, of being used in electronic noses. This neural network comprises, in the output layer thereof, fuzzy logic neurones that enable it to be indicated that the tested product belongs to an unknown class. In this document, it is also proposed to increase the number of neurones in the input layer of the neural network so as to input not only the sensor output signals but also signals representing variables calculated in a Principal Component Analysis of the original signals, for example, the two or three first components of the PCA.
However, by adding wholesale, to the measured signals representing the data, signals resulting from a Principal Component Analysis of the data, in order to constitute the input signals to the neural network, WO96/26492 follows the approach of conventional techniques which superpose statistical methods and neural networks. Moreover, even though the variables obtained by a principal component analysis convey substantial quantitative information, their contribution in terms of qualitative information is not necessarily great, particularly when a large number of distinct products is being handled. Thus, the systematic use of a certain number of predetermined variables derived from a statistical analysis does not automatically ensure the discrimination of the different products. Furthermore, these variables are correlated with certain of the initial variables (channels) and, consequently, their use as input signals to the neural network leads to redundancy.
SUMMARY OF THE INVENTION
The present invention relates to a pattern classification apparatus notably for use in the field of odour recognition, in which the statistical methods and the neural techniques are intertwined so as to exploit to the greatest possible extent the synergy between them, and to a method for optimising such an apparatus during the learning phase thereof.
More particularly, the present invention provides a classification apparatus, notably for use in the recognition or the characterisation of odours, comprising: means for acquiring, during a learning phase and by a plurality of channels, raw data representing a plurality of instances of a plurality of different classes; a processing unit for processing said raw data so as to determine, for each class (j), characteristics common to different instances of the class; and a neural network for outputting a classification, into one of the classes, of an instance presented to said classification apparatus, said neural network comprising an input layer, an output layer and at least one intermediate layer disposed between the input and output layers, the output layer comprising a plurality of neurones each being adapted to recognise instances of a respective class from the plurality of classes: characterised in that the processing unit determines for each class (j) a sub-space (SEj) in which the instances of this class (j), for which raw data has been obtained, are optimally separated from the instances of all the other classes, said sub-space (SEj) being defined by a group (VDj) of one or a plurality of synthetic variables (V), and determining the discrimination sub-space (SED) which comprises the sub-spaces of all of the classes, said sub-space (SED) being defined by the plurality of variables (VD) comprised in the sets (VDj) of synthetic variables of the plurality of classes: and that the input layer of the neural network comprises a plurality of neurones each corresponding to a respective variable (V) of the plurality of synthetic variables which define the discrimination sub-space (SED).
The present invention also provides a learning method of a classification apparatus comprising means for acquiring, by a plurality of channels, raw data representing instances of a plurality of different classes; a unit processing said raw data and a neural network comprising an input layer, an output layer and at least one intermediate layer disposed between the input and output layers, the output layer comprising a plurality of neurones each being adapted to recognise instances of a respective class from the plurality of classes, this apparatus being notably for use in the recognition or the characterisation of odours, th
Amine Hicham
Labreche Saïd
Loubet François
Tan Tze Tsung
Alpha M.O.S.
Black Thomas
Holmes Michael B.
Weingarten Schurgin, Gagnebin & Lebovici LLP
LandOfFree
Classifying apparatus using a combination of statistical... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Classifying apparatus using a combination of statistical..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Classifying apparatus using a combination of statistical... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2995807