Bearing Spectral Forecasting: Reproducible Code and Numerical Benchmarks
Abstract
Software and compact numerical benchmarks for multihorizon vibration-spectrum forecasting and latent-state consistency analysis using run-to-failure bearing data. The package contains Python source code, experimental configurations, bearing-group splits, pinned dependencies, automated tests, SHA-256 integrity records, and aggregate numerical results. It supports verification and reconstruction of distributed numerical results without external datasets, and protocol-level retraining after obtaining the LDM and XJTU-SY datasets from their original providers. See README.md for the scope of each workflow. The implementations cover recurrent, convolutional, attention-based, linear, and controlled-differential-equation forecasting models, heterogeneous ensembles, statistical comparisons, uncertainty evaluation, numerical-invariant checks, and sensitivity analyses. Version 1.1.0 integrates bearing-level results and statistical reconstruction for two restricted-input controls. It also corrects synthetic probe aggregation by selecting the GRU recurrent state and the H-CDE slow states, excluding pooled rows, and averaging runs within assets before averaging assets. The synthetic table generator and numerical reference are updated together. The underlying synthetic probe records and forecasting MSE results are unchanged. The recorded training uses float32. New training runs need not reproduce the exact historical fitted weights, because the primary run identifiers do not reconstruct their initial parameters. Included alignment records document previous checks; they are not a fresh validation of excluded per-window predictions. Raw vibration signals, prepared arrays, model checkpoints, per-window predictions, training histories, and figures are not included. Dataset access and licensing remain with the original providers. See README.md, DATASETS.md, and THIRD_PARTY_NOTICES.md for instructions and attribution. The code is distributed under the MIT License. Archive integrity File: bearing_spectral_forecasting_reproducibility_v1.1.0.zipSize: 1,874,001 bytesSHA-256: 92865fd6302a54cc3554a65b99ff66256128a4e1c247044d4bf6956210456a8d
