Traditional Problem Definition - Computational Intelligence in Time Series Forecasting - page 66

of complex time series with cyclic components and for extraction of underlying
sine and cosine functions of different frequencies, frequency analysis has been
employed, supported by building and analysis of corresponding periodograms for
interpretation of the data. Finally, for prediction of crude oil properties the Fourier
transformation has been used as a nonlinear, parametric model that can forecast
future values by processing the past values.
2.10.3 Production Monitoring and Failure Diagnosis
Production monitoring and failure diagnosis are the major objectives in on-line
observation of overall performance of a production plant. In manufacturing, the
major attention is paid to the monitoring and diagnosis of numerical machines and
of machine tools. In both cases - apart from modern approaches relying on
intelligent technology - statistical methods, based on time series analysis, are still
used. The main reason is that, for monitoring purposes, an abundant number of
observation data are collected on-line to be processed statistically.
Damiano et al. (1999) reported on the use of nonlinear time series to form a
one-step prediction map for machine monitoring and failure diagnosis in which the
sequence of previously collected observation data helped in the estimation of the
next time series data point. The map built in this way models efficiently the
dynamics of the system generated by a time series. Applying nonlinear time series
analysis, the optimum time delay is determined to be used for reconstruction of the
attractor , required for creation of the map that approximates the attractor. For
reconstruction of a multidimensional attractor , the method of delays was used,
where the vector components were created from the given time series using time
series values mutually separated by the delay time.
The one-step prediction is now applied to the machinery diagnosis. A baseline
time series is built out of data collected from the machine under normal operating
conditions and the nonlinear time series analysis used to build the corresponding
one-step prediction map. Using the map, the average map error is calculated for the
baseline time series. The calculated error and the map built are then employed for
machine monitoring by calculating the average absolute map error using the
current time series. The calculation results are then compared with the map error
for baseline time series and the difference between the two types of map error is
finally used to detect the possible changes in the machine being monitored.
2.10.4 Tool Wear Monitoring
In the following example, the most significant problem in flexible manufacturing
systems, the problem of monitoring of tool wear during the cutting and drilling
process, is assessed. This monitoring task is needed to maintain constant quality of
products and to avoid damage to the workpiece. To achieve this, a set of versatile
nondestructive sensing elements have to be installed for on-line tracing of the
status of tool wear during normal operation. The objective is to detect and replace
the tool when worn beyond the tolerable limits. In practice, acoustic emission
sensors are regularly used instead of power- or force-based sensors, because of