PREPROCESS

For the data to be pre-processed in the correct way the arrangement of the data should be as shown in Figure 1.

Figure 1. Illustration of a data arrangement of data from an electronic nose/tongue. NB - to use SENSABLE the time mode must always be the second mode, whereas the sample and sensor modes can be interchanged.

Notice that samples should be in either the first or the third mode. The signals (time mode) as shown in Figure 2 must be placed in the second mode to be able to use both pre-processing and models correctly.

Base-correction

Takes place in the second mode (= time mode) and uses the first time as S₀ (baseline signal) in the base-correction formulas.

Figure 2. Sensor signal (S) for the eleven sensors with the baseline signal (S₀) subtracted.

The data shown in Figure 2 is baseline independent meaning that the sample independent baseline (signal of a carrier gas) can be subtracted giving the same starting value for every combination of sample and sensor.

If the sensor based data is baseline dependent the first measurement of the headspace (sample) will be used as the baseline signal. If the size of the first measurement changes for each sensor some cautions must be taken as to make sure all samples are treated equal using the selected base-correction.

Transformation

Takes place in the second mode (= time mode). For logarithmic transformations be aware that it is not possible to take the logarithm to a negative number. This means that it only makes sense to use the logarithmic transformation in combination with the relative base-correction (S/S₀).

Feature extraction

Takes place in the second mode (= time mode) and extracts specific time related features. Note the feature extraction step can reduce the data to a two-way matrix and that the time mode is exchanged with the chosen feature (e.g. the maximum value).

Centering and Scaling

Centering takes place across the first mode and scaling within the third mode. This means that if the samples are to be centered and the sensors to be scaled then the data should be arranged as illustrated in Figure 1. NB - it is not possible to centre and scale more modes at a time. To do this, the pre-processed data must be saved and permuted in MATLAB if centering or scaling of a different mode is required.