Syntactic techniques for pattern recognition on sampled data signals

The established statistical pattern recognition techniques perform measurements either directly or on preprocessed sets of data samples and attempt to extract a multidimensional vectorial representation from which pattern types can be classified using clustering techniques. This permits the identification, for example, of specific fault conditions within the input data. The new syntactic technique analyses the input pattern data using a descriptive grammar to see whether the samples fall within an expected shape or envelope. A new type of grammar is described which allows fast rejection of noise samples which deviate significantly from the expected pattern. The construction of the necessary grammars is discussed, and a comparison provided between the statistical and syntactic techniques when attempting to recognise simulated pattern data. This clearly shows superior noise tolerance for the syntactic technique plus an ability, within a single grammar description, to accommodate pattern data with various scale factors in signal amplitude and arrival time. The new syntactic parsing technique shows particular promise for analysis of fault pattern data for condition monitoring of plant and machinery.