Divide et impera: how disentangling common and

distinctive variability in multi-set data analysis can

aid industrial process troubleshooting and
understanding
Supporting Material
Figure SM.1: Industrial batch process data - Reacting unit 1: detection
of the number of common components. The red vertical dashed lines de-
note the observed canonical correlations, while the blue bars represent their
respective empirical null -distributions estimated through permutations. a)
p−value < 0.01, b) p−value > 0.01

2
Figure SM.2: Industrial batch process data - Reacting unit 1: scores plot
related to the first distinctive component of the second time period batch
data block (see Figure 2). The batch ID reflects the manufacturing campaign
time. PC stands for Principal Component

3
Figure SM.3: Industrial batch process data - Reacting unit 2: a) Squared
Prediction Error (SPE) control chart resulting from a 3-principal compo-
nent model trained on Z1 . Each graphical symbol corresponds to an entire
process run, while the red dashed line denote the SPE 95% confidence limit.
b) Average SPE contribution plot related to all the Z2 batches signalled as
faulty in a). c) Time profiles of the loadings of the first distinctive com-
ponent of the second time period batch data block for the 21 measured
variables (found to be statistically significant after executing the effective
rank estimation algorithm described in Appendix B - Z2 -deflated variance
explained: 21.10%). The vertical dashed lines separate the 6 stages of the
industrial process. As not all the variables were active in these stages, part
of such profiles is missing. The corresponding scores plot is represented in
d). The batch ID reflects the manufacturing campaign time. PC stands for
Principal Component

4
Figure SM.4: Industrial batch process data - Reacting unit 1 and reacting
unit 2 combined: detection of the number of common within-distinctive
components. The red vertical dashed lines denote the observed canonical
correlations, while the blue bars represent their respective empirical null -
distributions estimated through permutations. a) p−value < 0.01, b) and
c) p−value > 0.01

5
Figure SM.5: Industrial batch process data - Reacting unit 1 and reacting
unit 2 combined: scores plots related to the first common within-distinctive
component. a) relates to Figure 5a), while b) to Figure 5b). The batch ID
reflects the manufacturing campaign time. PC stands for Principal Compo-
nent 6
Figure SM.6: Industrial batch process data - Reacting unit 1 and react-
ing unit 2 combined: scores plots related to the first distinctive between-
distinctive component. a) relates to Figure 7a), while b) to Figure 7b). The
batch ID reflects the manufacturing campaign time. PC stands for Principal
Component 7