| 摘要 |
PROBLEM TO BE SOLVED: To extract such featured values that input/output relations non-linearly change by reflecting the past history on a sandglass type neural network for processing multi-dimensional time series data. SOLUTION: A featured value extraction part 2 comprises a sandglass type neural network 20 and a feedback unit 30 forming a feedback structure from the output layer to input layer of the sandglass type neural network 20. The feedback unit 30 is provided with a Z converter 31; a low pass filter 32; a high pass filter 33; a band pass filter 34; and a hysteresis generator 35, and an output from an output layer L5 is data-processed in a time region, a frequency region, and a space region, and feeded back to an input layer L1 so that it is possible to achieve the reflection of the past history, the band adjustment of an output frequency, and the adjustment of input/output characteristics corresponding to the state. Thus, it is possible to extract such featured values that input/output relations non-linearly change. COPYRIGHT: (C)2009,JPO&INPIT
|
