5. Results and discussion5.1. Orthogonal experiments and MEA-BPNNThe levels of experimental factors are shown in Table 6. Commonlyused three-level orthogonal tables include L9(q4), L18(q7),and L27(q13). Based on the selection principle for orthogonal tables[38], L18(q7) is adopted in this paper. Each experiment in theorthogonal table is simulated by Fluent and the results correspondingto qV are given in Table 7.From the orthogonal experiment results, 15 randomly selecteddata points are used as training data set and 3 other data pointsare used as test data. The Levenberg-Marquardt algorithm is usedto train the network in the training function. The process of similartaxisand dissimilation is shown in Fig. 17. After several similartaxisoperation cycles, in Fig. 17(a) and (b), scores of the initialsuperior and temporary subgroups become constant. This meansthat each of these subgroups is mature. Comparing Fig. 17(a) and(b), it is found that the scores of initial temporary subgroups subgroup1, subgroup 2, and subgroup 3 are higher those of the initialsuperior subgroups subgroup 2, subgroup 4, and subgroup 5.Hence, the dissimilation operation is conducted thrice and 3 newsubgroups are added to the temporary subgroups. In Fig. 17(c), itcan be observed that the score of each subgroup remains constantin the similartaxis process, which means that these subgroups aremature. Comparing Fig. 17(c) and (d), it can be noted that thescores of temporary subgroups are lower than those of superiorsubgroups and therefore the dissimilar operation is no longer executed.In this instance, superior subgroups have already beenadded to the global billboard.Moreover, the generalised ability of neural networks isimproved after weight and threshold optimisation by MEA. Further,the prediction error of MEA-BPNN is 4.53 105, which islower than that of 1.1 103 of BPNN without optimisation.5.2. Optimisation resultsThe process of similartaxis and dissimilation for structuralparameter optimisation is given in Fig. 18. It can be found thatthe scores of subgroup 1, subgroup 2, and subgroup 4 in initialtemporary subgroups are higher than those of subgroup 1, subgroup3, and subgroup 5 in initial superior subgroups (Fig. 18(a)and (b)). Therefore, the dissimilation operation is conducted thriceand 3 new subgroups are added to the temporary subgroups. InFig. 18(c), it can be noticed that the score of each subgroup remainsconstant in the similartaxis process, which means that these subgroupsare mature. Comparing Fig. 18(c) and (d), it can be notedthat the scores of temporary subgroups are lower than those ofsuperior subgroups and hence dissimilation is not executed. Theoptimal structure parameters obtained after decoding are givenin Table 8.5.3. Separation efficiencyThe size distribution of COF particulate matter from Li et al.’sexperiments [39] fits the Rosin Rammler Equation [40]. The completesize range is divided into 10 intervals, each represented bya mean diameter for which trajectory calculations were performed.The density of particles is 960 kg/m3. The cumulative particle sizedistribution and parameters, such as the minimum, maximum,mean, and spread diameter (n) are shown in Table 9.