Results and Discussion

The overall prediction results of the MLP model and the corresponding confusion matrices for both laboratory and plant trials are in Tables 1 and 2. From Tables 1 and 2, it is clear that the performance of the MLP model is extremely accurate. The MLP model is able to distinguish almost perfectly the boundaries between all the errors for both laboratory and production trials and only has some difficulty with recognition of the boundaries between the normal condition and the errors. Even these boundaries are recognized with a very high accuracy between 93% and 97%.

In Table 1, NC represents normal condition, Err1 represents no coating, Err2 represents zinc phosphate layer only, Err3 represents peening, and Err4 represents scratching. Confusion matrices here are 5 by 5 matrices showing how many samples belonging to a particular condition were classified accurately and how many samples were misclassified as other conditions, for example 13,407 nonlubricated samples were classified accurately as NC, 109 as Err1, 68 as Err2, 226 as Err3, and 510 as Err4.

Most importantly, while the extra layer of calcium-stearate coating applied to lubricated samples makes the defects visually indistinguishable from the normal condition except for Error 3 (peening) in the laboratory trial and Error 4 (scratching) in production trials, an MLP model is able to recognize them almost as accurately as for nonlubricated samples where only the Error 4 (scratching) for the laboratory trial and Error 3 (peening) in production trials are visually indistinguishable from the normal condition.

In Table 2, NC represents normal condition, Err2 represents zinc-phosphate layer only, Err3 represents peening, and Err4 represents scratching; no Err1 representing no coating was introduced, due to possible damage to tooling and sensors in the plant. Confusion matrices here are 4 by 4 matrices, showing how many samples belonging to a particular condition were classified accurately and how many samples were misclassified as other conditions. For example, 10,277 nonlubricated samples were classified accurately as NC, 74 as Err2, 169 as Err3, and 36 as Err4.

0 0

Post a comment