Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper proposes a cumulative degradation time approach specifically for ATE driver boards. Herein, a machine-learning (ML)-based anomaly detection method is integrated with accelerated life testing (ALT). This invokes ML models to spot the abnormal patterns of voltage-temperature relationships during their high-stress operation, while ALT presents a sound method to project the driver board performance at elevated thermal levels. This paper, therefore, would be of special interest to professionals concerned with the aspects of ATE reliability and maintenance in the semiconductor industry.
However, I have several comments that should be addressed.
1. In the introduction (lines 52 and 53), the authors wrote, "This study focuses on applying PdM to one of the most crucial components within the ATE system: the test head." Applying PdM to degrading equipment components requires addressing the following challenges: (1) developing a physically meaningful degradation model; (2) explicitly modeling uncertainty and imperfect inspections; (3) performing probabilistic remaining useful life (RUL) estimation; and (4) establishing a risk-aware decision-making framework. However, this article only addresses the first problem, which must also be addressed within the framework of preventive maintenance and condition-based maintenance modeling. Therefore, in this section and elsewhere where the authors claim to be solving a predictive maintenance problem, the text needs to be adjusted (e.g., abstract).
2. In reference [11] (Nelson, Wayne, “Graphical analysis of accelerated life test data with the inverse power law model,” IEEE Transactions on Reliability, Vol. 21, No. 1, pp. 2–11), the correct publication year is 1972, not 2009.
3. The form of equation (1) is problematic because η appears on both sides of the equation.
4. Overall, the article is descriptive in nature. The authors must provide a level of detail sufficient to enable interested readers to reproduce the reported experimental results. Appendices containing the necessary explanations may be added.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have presented "Thermal Degradation Diagnosis of ATE Driver Boards Using ALT-Derived Cumulative Degradation Time". The comments are;

1. Please make the Arrhenius→CDT conversion fully explicit and consistent: Equation (2) defines AF(T) while Equation (3) accumulates tequiv(i) = tequiv(i-1) + Δt/AF(T); earlier text says “AF-weighted” time, which sounds like Δt·AF. Pick one phrasing and show the worked substitution with the chosen activation energy (you list Ea = 0.5–0.7 eV ≈ 48.2–67.5 kJ/mol). Also avoid overloading η (as exponent and as a function argument in Eq. (1)) and confirm that the IPL form is T_fail(V) = 1/(K·V^η). Finally, justify the anomaly threshold: why 4.2 V and CDT = 0.999597 from Table 1—was this a 95th-percentile rule, a ROC-optimal cut, or a process spec? A one-paragraph rationale plus a small schematic around Fig. 6/Table 1–2 would settle this.
2. Results show 1.000 accuracy for decision tree and random forest (Fig. 7; Table 4). To ensure the effect is not from temporal or run leakage, please (i) split by run/board (not random samples), (ii) report a forward-in-time test, and (iii) give counts per class, ROC/AUC, and calibration. Add basic metrology details for Fig. 5: thermal-camera calibration, temperature uniformity on the board, voltage measurement accuracy, sampling rate, and ALT duration. If possible, include a second driver board or a repeat ALT to demonstrate repeatability.
3. Table 3 lists models, but the exact dials are still terse. Please provide feature scaling, windowing (if any), train/validation/test splits, random seeds, and the full hyperparameters actually used for the best models (tree depth, min samples, RF trees, SVM C, etc.). A short pseudocode box (“temperature, voltage → compute CDT → features → model fit → decision”) plus the precise train/test sample counts will let others rebuild Table 4 and Fig. 7.
4. Add units to all axes (e.g., °C, V, seconds or CDT units) and mark event instants on Fig. 6. In Tables 1–2, correct the header label to t_equiv(i) consistently (it appears as tEquip in places), include column units, and add a caption note that Table 1 is ALT at 180 °C and Table 2 is normal at 60 °C. In Fig. 7, annotate class totals in each confusion-matrix quadrant and note the test-set size in the caption. A small “parameters at a glance” box (Ea used, Tuse, Tstress, AF formula, threshold CDT and voltage) would make the method easy to adopt.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper presents a PdM framework for an automatic test equipment driver board. The paper addresses the reliability challenge in semiconductor manufacturing. The core contribution is the introduction and application of cumulative degradation time, derived from accelerated life testing principles using the Arrhenius equation. The paper shows that various AI models can achieve exceptionally high accuracy in diagnosing the degradation state of driver boards. However, there are some major comments that need to be addressed:

1. I suggest that the authors discuss the potential for overfitting. It would be convincing to show results from a rigorous evaluation method, such as a hold-out test set from a separate, unseen experimental run. Also, the authors can consider introducing slight Gaussian noise to the voltage/temperature signals during training to demonstrate model robustness. The near perfect scores suggest an overly simplistic separation between normal and abnormal states.
2. The paper is missing an experimental setup description. I suggest that the authors provide details such as the experimental protocols, dataset description, characteristics, and train/test/Val split ratios. Table 3 mentions some configurations, but was any hyperparameter selection/optimization performed? If default parameters were used, the authors should state this clearly.
3. I suggest that the authors rename section 4: AI models as “Anomaly detection models” or ML-based Anomaly detection”, etc
4. The study is conducted on a single driver board model under a single stress type (steady-state high temperature). However, real world ATE environments are more complex. I suggest that the authors acknowledge this limitation and the limitations of the paper.
5. The authors should consider discussing how to select and calibrate the anomaly threshold.
6. Please clarify the exact input features for the models. Was tEquiv(i) used alone or in combination with the raw signals? What is the dimension of the input feature?
7. Did the features underdo preprocessing, scaling, or normalization? It is suggested to include the details in the experimentation section.
8. It is suggested that the authors remove Figure 8 and its description from the conclusion section. This information belongs to the methodology section.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors addressed all my comments.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have responded to all the queries raised by the reviewer in the previous version as well as modified the manuscript accordingly. No more changes needed.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have addressed all my comments