The designed algorithm of autoassociative neural network has the active and adaptive dynamics such as the learned neural network function performs an orthogonal projection of the presented corrupted image of the energy device parameters into the space of a finite number of uncorrupted images of possible parameter compositions. A single-layer recurrent linear neural network with fully interconnected neurons was used. It operates as an auto-associative memory to store the patterns contained in the training data during the training phase (teacherless learning). These patterns (or combinations of patterns), if partially corrupted and resubmitted to the network in the equipping phase, can be repaired by the network.
Spuštění funkcionality autoasociativní sítě v adaptivním a aktivním režimu
The autoassociative neural network algorithm was tested according to the following description. During the network adaptaion, 4 vectors of size 9 were stored into the autoassociative memory. The noise was added to 6th, 7th and 8th component of the vector and then, this modified vector was passed to the network during the network activation – with and without the noise filtering. Results show that without fixing components without noise, the corruption of noisy components is lower, but components with noise are also corrupted. If the components with noise are fixed, corruption of noisy components is almost zero.
Downloads:
- User guide (in Czech)
- License agreement and trial version
The software has been developed with a financial support of the Technology Agency of the Czech Republic under the research project No. TK04020003.
