Self‑Supervised Temporal Pattern Mining for Circular Manufacturing Supply Chains for Extreme Data Sparsity Scenarios

Описание и применение

В условиях крайне редких и нерегулярных данных о возвратах в циклической цепочке поставок, традиционные методы прогнозирования не работают. Используя самообучение с контрастивным усилением, модель извлекает повторяющиеся временные паттерны из неразмеченных данных, а затем дообучается на небольшом размеченном наборе, улучшая точность прогнозов времени прибытия возвратов.

Задача

Predicting return‑arrival times in a remanufacturing facility with fewer than 400 labeled events, irregular timestamps, and frequent regime shifts caused supervised models to overfit or produce flat predictions.

Результат

The self‑supervised approach successfully extracted reusable temporal patterns from sparse, irregular data, improving prediction quality over traditional supervised methods in a circular manufacturing context.

Рабочий процесс

• Collect unlabeled time‑series data (sensor readings, shipment timestamps, material logs).
• Apply TemporalAugmenter: magnitude scaling, temporal jitter, and random masking to create two views of each window.
• Train a dilated causal 1D convolutional encoder with contrastive loss to pull augmented views together and push apart different windows.
• Fine‑tune the encoder on the small labeled set to predict return‑arrival times.
• Deploy the fine‑tuned model to schedule labor and material flows in the remanufacturing facility.

Доказательства

Traditional supervised forecasting collapsed. XGBoost overfit after a few dozen trees. LSTMs memorized the training set and produced flat, useless predictions on holdouts.

Источник

The masking step was the single most impactful design choice. In sparse regimes, missing timestamps are the norm, not an anomaly, so training the encoder to reconstruct around gaps directly matches the deployment condition.

Источник