Optimisation of Correlation Matrix Memory Prognostic and Diagnostic Systems

Condition monitoring systems for prognostics and diagnostics can enable large and
complex systems to be operated more safely, at a lower cost and have a longer lifetime
than is possible without them. AURA Alert is a condition monitoring system that uses a
fast approximate k-Nearest Neighbour (kNN) search of a timeseries database containing
known system states to identify anomalous system behaviour. This search algorithm,
AURA kNN, uses a type of binary associative neural network called a Correlation Matrix
Memory (CMM) to facilitate the search of the historical database. AURA kNN is evaluated
with respect to the state of the art Locality Sensitive Hashing (LSH) approximate kNN
algorithm and shown to be orders of magnitude slower to search large historical databases.
As a result, it is determined that the standard AURA kNN scales poorly for large historical
databases.
A novel method for generating CMM input tokens called Weighted Overlap Code Construction
is presented and combined with Baum Coded output tokens to reduce the query
time of the CMM. These modifications are shown to improve the ability of AURA kNN to
scale with large databases, but this comes at the cost of accuracy. In the best case an AURA
kNN search is 3.1 times faster than LSH with an accuracy penalty of 4% on databases with
1000 features and fewer than 100,000 samples. However the modified AURA kNN is still
slower than LSH with databases with fewer features or more samples.
These results suggest that it may be possible for AURA kNN to be improved so that
it is competitive with the state of the art LSH algorithm.