Code and data for "Leakage-Aware Evaluation and Real-Time Feasibility of Machine Learning for Condition Monitoring of Servo-Driven Machinery"

Running it reproduces every number, table and figure in the paper from the raw data. Contents - code/ — the full pipeline: 23-dimensional feature extraction; five partitioning protocols (random segment split, random segment cross-validation, record-grouped cross-validation, leave-one-load-out, leave-one-fault-size-out); leakage quantification in two window-overlap configurations; the leakage-free benchmark with confidence intervals; hyper-parameter tuning performed strictly inside the training folds; a capacity sweep over three deep architectures; per-segment latency measurement; cross-domain transfer; and paired significance tests (Nadeau-Bengio corrected resampled paired t-test with a repeat-level Wilcoxon check, Holm-corrected within two families). - code/figures/make_figures.py — regenerates all ten figures from the result files.

It never retrains a model, so the figures cannot drift from the tables. - data/simulink/ — the simulated servo drive-current dataset: 120 independent randomised simulation runs across four conditions. servo_runs_meta.csv records the random seed and the full parameter vector of every run, so the dataset is exactly reproducible. - tests/ — four self-checks: the experimental design, numeric traceability of the manuscript, citation consistency, and figure/table ordering. - refs/, requirements.txt, run_all.py, README.md. The CWRU vibration data are not redistributed here; they are publicly available from the Case Western Reserve University Bearing Data Center, and code/data/download_cwru.py fetches the mirror used in the paper. Two details, documented in the README, change results materially: the grouping key must be the recording identifier rather than the segment, and grouped folds must be class-stratified at the group level. scikit-learn's GroupKFold does not stratify and, on CWRU, produces a fold containing no inner-race recordings at all.