A Data-Driven Framework for Unsupervised Monitoring of Transmission Systems Using End-of-Line Testing Data: A Case Study at Ford Motor Company

1 Introduction

• Drivetrain defects are consequential because they contribute substantially to automotive warranty expenditures and raise long-term reliability concerns.Large-scale product defects can also trigger recalls with economic and reputational penalties, and mechanical failures can have safety consequences under some conditions.Source: §1 · pp. 1–2
• Conventional EoL procedures use calibration routines, diagnostic trouble code checks, and threshold-based pass/fail criteria, while modern transmissions generate high-dimensional time-series data.These data include hydraulic pressures, clutch currents, speeds, torque, and vibration measurements collected at high sampling rates.Source: §1 · p. 2
• High dimensionality, temporal dynamics, misalignment, repeated cycles, and physical signal coupling complicate feature extraction and can make simple static thresholds inadequate.Collected recordings may include earlier repeated test cycles even though only the last cycle matters; coupled signals can also create multicollinearity and redundancy.Source: §1 · p. 2, Figure 1 · p. 3
• The proposed framework removes recycled segments, aligns signals with fast Dynamic Time Warping, models input–output relationships with regression residuals, and monitors KPCA representations using Hotelling’s T 2.The method targets nonlinear dynamics and deviations from the learned normal profile.Source: §1 · p. 2

2 Literature review

• Transmission EoL quality checks commonly compare vibration and acoustic measurements with calibrated limits or regression criteria (Khond et al. (2023); Liang et al. (2025); Su et al. (2024)).Related work also includes statistical noise inspection and practical NVH-based anomaly-detection prototypes (Horv´ath (2025); Blickensdorff et al. (2023)).Source: §2 · p. 3
• Rotating-machinery diagnostics span time- and frequency-domain features (Jorani et al. (2023); Peng et al. (2019); Chin et al. (2021)) and learned spatiotemporal models (Wang et al. (2025); Miao et al. (2023); Wei et al. (2022)).Automotive predictive-maintenance work also uses ensemble learning and multimodal sensor fusion (Arena et al. (2021); Akhtar et al. (2026)).Source: §2 · p. 4
• Temporal misalignment, variable durations, repeated cycles, and coupled signals complicate monitoring, while linear latent-variable models may inadequately represent nonlinear dynamics (Peng et al. (2019); Chin et al. (2021); Huangfu et al. (2023); Wang et al. (2025)).The review also identifies scarce labeled failures and operational variability as challenges in EoL settings.Source: §2 · p. 4
• The study combines data cleaning, cycle removal, smoothing, temporal alignment, regression residuals, KPCA, and Hotelling’s T 2 monitoring with weighted phase-wise anomaly ranking, validated on production data.The contributions are framed as addressing EoL data inconsistencies, signal dependencies, high dimensionality, and realistic operating conditions.Source: §2 · p. 5

3.1 Data Preprocessing

• Preprocessing removes rejected and incomplete tests, retains only each phase’s final valid cycle, and smooths signals with non-overlapping moving averages.The moving average reduces high-frequency noise and data dimensionality while preserving the overall signal trend.Source: §3.1 · pp. 5, 7

3.2 Signal Alignment

• FastDTW (Salvador and Chan (2007)) aligns signals to a reference trajectory, approximating the optimal warping path with linear time and space complexity.The alignment addresses temporal differences caused by variable testing conditions.Source: §3.2 · pp. 7–8

3.3 Input–Output Relationship Modeling

• Regression residuals remove dominant linear coupling between paired signals, reducing spurious variability and producing features approximately invariant to shared operating-condition changes.The residual represents output variation not explained by the input–output dependency and is used for subsequent feature extraction and monitoring.Source: §3.3 · pp. 8–9

3.4 Low-Dimensional Representation via Kernel PCA

• KPCA maps preprocessed features into nonlinear scores that capture dominant structure while filtering noise and redundancy; dimensionality reduction is applied independently to each phase.The study uses a polynomial kernel to capture nonlinear interactions while retaining a structured similarity measure based on inner products.Source: Figure 2 · p. 6, §3.4 · p. 9

3.5 Monitoring Statistics and Anomaly Detection

• The framework applies Phase I Hotelling’s T-squared monitoring to KPCA scores, flagging observations that exceed the upper control limit.Monitoring is conducted separately for each phase.Source: §3.5 · p. 10

3.6 Anomaly Score Computation

• The anomaly score combines phase-specific deviations as threshold-normalized ratios, preventing phases with larger thresholds from dominating the ranking.Other one-class classification methods can also be used, with scores measured relative to a threshold at the (1 − α) quartile.Source: §3.6 · p. 11

4 Real Case Study Data Description

• 502 transmission units were tested over three days, with multiple sensor signals recorded during standardized end-of-line procedures.The collection spanned September 30, October 1, and October 2, 2024.Source: §4 · p. 12
• Each transmission is represented by high-dimensional time series, including torque, speed, and other process measurements across operational phases.The recorded signals vary in sequence length and temporal progression across transmissions.Source: §4 · p. 12
• Phases use numerical identifiers and steps use chronological identifiers, but their specific meanings are confidential.The phase–step definitions were established by company domain experts.Source: §4 · p. 12
• Measurement noise, operational fluctuations, and dependencies among signals contribute to variability in the dataset.Source: §4 · p. 12
• The data contain both in-control and out-of-control transmissions, with out-of-control status identified from quality inspection outcomes.The paper uses this structure to evaluate anomaly detection under manufacturing conditions.Source: §4 · p. 12

5.1 Dropping Rejected Test

• Rejected tests and tests lacking required phase–step pairs are removed to estimate the in-control transmission distribution.The required pairs correspond to critical segments identified by company domain experts.Source: §5.1 · p. 16

5.2 Removing Cycles

• When a phase contains repeated noncontiguous blocks of its minimum step ID, the procedure keeps only the last cycle.Figure 5 illustrates line-pressure and step traces before and after cycle removal.Source: §5.2 · p. 16, Figure 5 · p. 17

5.3 Incomplete Test Detection

• 472 of 502 transmissions were retained after four became incomplete following cycle removal and were excluded.Retained transmissions contribute only phase–step pairs specified as important in Table 2.Source: §5.3 · p. 17

5.4 Non-Overlapping Moving Average

• A non-overlapping moving average with window size 3 reduces the effective signal length by a factor of 3.Source: §5.4 · p. 17

5.5 Signal Alignments

• Fast Dynamic Time Warping aligns each signal to a reference transmission, producing signals of length 1208.The paper uses DTW after moving-average smoothing to address differences attributed to unknown conditions during testing.Source: §5.5 · p. 17
• 85 of 472 transmissions with approximately constant ConvOutPressur signals were removed, leaving 387 transmissions.The authors reported the distinct behavior to the company for diagnosis as inconsistent with expectations.Source: §5.5 · p. 18

5.6 Input Output Relationship - Signal De-correlation

• Table 5 summarizes sensor input–output relationships that motivate the signal-decorrelation step.Source: Table 5 · p. 18, §5.6 · p. 18
• Regression residuals replace Euclidean-distance features for correlated sensor pairs, capturing unexplained information while decorrelating signals and avoiding sensitivity to input or output scales.The paper notes that applying regression after DTW means residual signals need no further alignment.Source: §5.6 · p. 18

5.7 Qualitative Results

• Each phase’s KPCA feature dimensions are reported in Table 6.The supplied table caption identifies the table as feature dimensions for each phase.Source: §5.7 · p. 18, Table 6 · p. 18
• The per-phase Hotelling’s T-squared charts use α ≈ 0.00054 to target an overall Type I error rate of 0.0027.Figure 11 displays the phase statistics, control limits, and highlighted anomalies.Source: §5.7 · p. 18
• The top five percent of anomalies detected by Alg. 2 were visually different signals, which the authors report their algorithm detected successfully.Figures 12 and 13 show examples across phases combined.Source: §5.7 · p. 18
• Figures 6–8 compare signal traces before and after DTW alignment.The supplied captions identify three parts of the before-versus-after signal traces.Source: Figure 6 · p. 19, Figure 7 · p. 20, Figure 8 · p. 21
• Figure 9 shows phase 100 signals by independent variable, dependent variable, and residual; Figure 10 shows residuals after regression adjustment.Figure 10 includes OutputTorque given InputTorque, OutputSpeed given InputSpeed, and OutputSpeed given TSSSpeed.Source: Figure 9 · p. 22, Figure 10 · p. 22

5.8 Quantitative Results on Phase 90 Dataset

• Evaluation was constrained by immature warranty follow-up, no established expert protocol for the studied phases, and expert feedback limited to transmissions flagged by the method.The authors state that insufficient time had passed for enough warranty information to become available.Source: §5.8 · p. 23
• The three Phase 90 subsets contained 3,174 samples with four confirmed anomalies, 1,545 with none confirmed, and 2,131 with three confirmed.The subsets were used to assess applicability and generalization across transmission types.Source: Table 7 · p. 23, §5.8 · p. 23
• 0.769 F1 score versus 0.429 for the company benchmark; the proposed model also outperforms it in accuracy and recall.Table 9 reports the comparison for the first subset.Source: Table 9 · p. 28, §5.8 · p. 28
• For the second subset, accuracy was 0.60, precision was 0.60, recall was 1.00, and F1 was 0.75.The model flagged 0.3236% of the dataset as anomalous.Source: §5.8 · pp. 28–29, Table 11 · p. 29
• For the third subset, the method detected all three confirmed anomalies and flagged 0.0056 of the dataset as anomalous.No company expert feedback was available for this subset.Source: Table 12 · p. 29, §5.8 · p. 30
• Increasing the proportion flagged as anomalous tends to reduce false negatives at the expense of more false positives; the acceptable balance depends on operational needs and business rules.Practitioners can adjust the flagged proportion through the Type I error or related design parameters.Source: §5.8 · p. 30

5.9 Online Weight Update for Alg.2

• Before ground-truth feedback is available, practitioners can select top anomalies using historical warranty rates, whose value the authors withhold for confidentiality.The authors describe receiving warranty feedback later in a streaming manner.Source: §5.9 · pp. 30–31
• The online update turns the anomaly score into a probability with a sigmoid classifier and learns phase weights from warranty feedback.The method uses binary cross-entropy and gradient descent with a fixed learning rate.Source: §5.9 · p. 31
• Feedback is incorporated for all deployed transmissions, including those not flagged as anomalous in any phase.This allows the update to account for transmissions that may be anomalous despite not being detected by a phase.Source: §5.9 · p. 31
• After each update, the phase weights are divided by their total so they sum to one.The algorithm describes this normalization after weight updates.Source: §5.9 · pp. 31–32

6 Conclusion

• The framework combines data cleaning, cycle removal, smoothing, temporal alignment, regression residuals, KPCA, Hotelling’s T-squared monitoring, and weighted anomaly prioritization.The conclusion presents this as a framework for high-dimensional end-of-line transmission testing data.Source: §6 · p. 32
• The real-world case study reports identification of anomalous transmissions while accounting for temporal variability, signal coupling, and high dimensionality.The conclusion describes the combined nonlinear feature extraction and statistical process monitoring approach as practical and interpretable for manufacturing quality control.Source: §6 · pp. 32–33
• Future work could tune the model with reference anomalies in a semi-supervised setting to decrease false positives or replace its two stages with an end-to-end model.The current implementation extracts low-dimensional features first and monitors in a second stage.Source: §6 · p. 33

• Generative AI assisted with code implementation and grammar checking, polishing, and improving the paper's text.Source: Gen AI Use · p. 33
• The authors developed all ideas presented in the work.Source: Gen AI Use · p. 33
• The authors thoroughly reviewed all outputs produced by generative AI.Source: Gen AI Use · p. 33