LIBRARY ı6ıul 0) - Cranfield University

" digital hardware, which is basically composed of a digital computer and
analog-to-digital (A/D) converters.
" signal processing component, provided by dedicated software algorithms.
2.6.3 Signal processing and interpretation
Signals in general can be classified into two groups, namely: deterministic and
stochastic signals. Stochastic signals are those whose behaviour is highly
unpredictable, that is, they occur randomly [ref. 155]. On the other hand,
deterministic signals are those that have known characteristics and that can be
explicitly described by mathematical and physical models [refs. 155,158].
Stochastic signals are affected by random noise but if the noise content is
negligible, the process may be regarded as deterministic. It should be noted that the
welding current and voltage signals in their unsmoothed states are stochastic. [ref. 51]
A deterministic signal can be formed from a stochastic signal provided the
amplitude or time classes of the signal are formed over a sufficiently long period [ref.
158]. Theoretically, for a precise information to be extracted from a stochastic signal
an infinite record length is necessary and the information based on finite length
records must always be qualified by statistical statements referring to the probability
of the information being correct within a certain percentage [ref. 159]. The period for
which data is collected (i. e. sampling time) should be sufficiently long such that the
mean value of a definite portion of the signal is equal to the overall average of the
total signal [refs. 156,158] and/or the Fourier transform of data collected over a
longer period should not differ significantly from the Fourier transform of the data
collected over the initially chosen sampling time [ref. 159]. This enables statistical
analysis to be performed on the signal.
For monitoring gas metal arc welding, sampling times ranging from 100
milliseconds to 1 second and sampling frequencies21
ranging from 200 Hz to 10 kHz
have been reported in the literature [refs. 38,161,162].
The minimum sampling frequency necessary for a sampled data to represent
the continuous time signal without aliasingu is set theoretically by the Shannon
sampling theorem as twice the maximum frequency-component of the signal [ref.
163]. Sampling frequencies equal to or greater than 8 times the maximum signal
frequency are generally used [ref. 163].
Large amounts of data are usually collected during monitoring, most of this
information is not useful for process control. The data need to be reduced to make
analysis simpler, faster and to save on storage capacity [ref. 51].
The most common data processing approach is to break the sampled transient
data into its basic statistical features such as mean, minimum, maximum, standard
deviation, etc. This is called feature extraction and significantly reduces the data
without losing important information, filtering out irrelevant information [ref. 160].
21 Sampling frequency is the frequency at which the Analog-to-Digital converter acquire and
converts the analog signal to a discrete-time signal (series of consecutively sampled data). The data
is normally acquired during the sampling time at a fixed sampling frequency.
22 Aliasing is a term used in control theory to define the distorsion that occurs in a signal when it is
reconstructed from a digitized signal which was sampled with a frequency not high enough to fully
represent the original analogue signal.
42