Towards Early-Stage Corrosion Prediction Using UHF RFID Measurements: A Machine Learning Feasibility Study

There is some confusion about the data. The methodology is written as if the authors collected real-time RFID measurements, but the paper later states that the study used a previously published public dataset. The wording should therefore be clearer about what was newly collected and what was reused.

The biggest concern is the validation strategy. The 1,089 records are repeated measurements, not 1,089 independent corrosion specimens. Multiple records come from the same underlying physical samples. This means that very high results such as (R^2 = 1.00) should be interpreted cautiously, because the model is not being tested on completely new physical specimens.

The so-called LOSO validation also does not leave out a physical specimen. Instead, the authors group data by the RFID “readcount” variable. So the training and test sets can still contain measurements from the same corrosion specimens. This limits how strongly the paper can claim generalization.

There is also a clear inconsistency in the LOSO description. Table 6 shows that both Fold 5 and Fold 6 contain records from all four corrosion stages, but the text says Fold 6 is the only fold containing all four stages. This should be corrected.

The paper also describes a “90:10 split with 15-fold cross-validation,” but it is not fully clear how these two procedures were combined. A more detailed explanation is needed so that another researcher could reproduce the experiment.

Another concern is the claim of continuous corrosion-thickness prediction. In reality, the model uses only four fixed target values: 0, 43, 77, and 108 µm. So the model is mainly learning four nominal corrosion states rather than demonstrating general continuous thickness prediction.