L.A. Stankevich, V.A. Efremov

Sequential pattern recognition (SPR) is one of the known and important problems in the field of recognition. Goal is to identify or classify complex patterns consisting of ordering over time sequence of more simple patterns. Such problem arises in many applications where recognition of temporal sequence patterns is needed. In this work SPR problem has following formalization. Some data sequences is given, in which some data group can be selected that defines a situation. At this each element of the data group can be considered as pattern, this is set of parameters with fixed links. It is requested to recognize of situation corresponding to certain sequence of pattern. To solve this problem, neurological networks built on modules with grid and cluster activator models are proposed. For practical usage, the network with modules on grid activator model is selected. Proposed system for sequential pattern recognition consists of two neurological modules: one for pattern processing and one for pattern sequence processing. For each of sequential pattern vectors of pattern features are calculated that are came to enter of module for pattern processing and form element of vector of sequence pattern features. Obtained vector of sequence pattern features comes to enter of module for pattern sequence processing an then string definition of sequence content is formed. For learning system for sequential pattern recognition two algorithms are developed: one for module for pattern processing and one for module for sequence pattern processing. These algorithms realize process of tuning weight of internal element links of modules at associative saving pattern examples. Process of recognition is realized by third algorithm that activates of learned the modules when sequence of patterns is given and defines classes of pattern sequences. Experiment on neurological network learning and recognition of situations represented by groups of sequential scenes showing goal-directed actions of operator during preparation of eat is conducted. As result of the experiment, tuning parameters of the neurological network were founded that best recognition results of situation recognition allow for obtaining.