Optimizing method of learning data set for signal discrimination apparatus and signal discrimination apparatus capable of optimizing learning data set by using a neural network
View Patent ↗A method of the present invention is processed by a selector. The selector selects each member constituting a learning data set from a data set source. Each member of the source is feature data extracted through a transducer and assigned to any one of categories in advance. The selector calculates each member's divergence degree of the source to obtain an average divergence degree. If an output neuron of the output layer of a neural network is related to different categories of all the categories represented by the output layer, the selector includes every member of the source corresponding to the category of the minimum average divergence degree in the selection from the source to the learning data set. The selector also excludes, from the selection, every member of the source corresponding to every remaining category of the different categories.
1. A selection method of learning data set for signal discrimination apparatus, said method being used for the signal discrimination apparatus comprising:
a transducer for extracting feature data including parameters from a measurement signal; and
a competitive learning neural network, including an input layer with input neurons and an output layer with output neurons, said input neurons corresponding one-on-one to the parameters of feature data extracted through the transducer, each of said output neurons being, based on a learning data set, coupled to all input neurons of the input layer through weight vectors to be related to any of categories, said network being configured to relate feature data extracted through the transducer to any output neuron of the output layer to classify into any of the categories;
wherein the method is processed by a selector, further included in the apparatus, for selecting each member constituting said learning data set from a data set source of which each member is feature data extracted through the transducer and is assigned to any one of the categories, in advance,
the method comprising a step (A) performed after a preprocess of sequentially entering every member of the source into the network to try to relate each member of the source to any output neuron of the output layer, said step (A) being repeated until each output neuron of the output layer is related to a single category of the categories,
the step (A) comprising steps of
(a) judging whether an output neuron of the output layer is related to different categories in all categories represented by the output layer;
(b) calculating each member's divergence degree of the source corresponding to the different categories with respect to the output neuron in question if related to the different categories;
(c) calculating each average divergence degree of the different categories based on said each member's divergence degree of the source;
(d) including every member of the source corresponding to the category of the minimum average divergence degree in the selection from the source to the learning data set; and
(e) excluding every member of the source corresponding to every remaining category of the different categories from the selection.
2. The selection method of learning data set for signal discrimination apparatus of claim 1 , wherein each member's divergence degree of the source at the step (b) is magnitude of difference vector between the weight vector on the output neuron in question and the member in question.