Concept Drift

Concept drift occurs when P(Y|X) — the conditional probability of failure given sensor readings — changes over time. Unlike data drift, which affects P(X), concept drift alters the fundamental mapping between inputs and outputs. For example, a vibration threshold of 4.5 mm/s that previously indicated bearing wear may no longer correlate with failure after a lubricant change, invalidating the original decision boundary.

Drift manifests in distinct temporal patterns:

• Sudden drift: Abrupt change triggered by events like equipment retrofits, software updates, or process changes
• Gradual drift: Slow evolution as machinery ages, materials degrade, or environmental conditions shift incrementally
• Incremental drift: Continuous adaptation where the relationship morphs through intermediate states
• Recurring drift: Cyclical changes tied to seasonal factors like ambient temperature or production schedules

Common catalysts for concept drift in manufacturing environments:

• Hardware modifications: Replacement of components with different specifications or materials
• Process parameter changes: Adjustments to feed rates, pressures, or operating speeds
• Environmental shifts: Changes in ambient temperature, humidity, or vibration from nearby equipment
• Sensor recalibration: Altered measurement baselines that shift feature distributions
• Maintenance interventions: Repairs that change equipment behavior signatures

Monitoring for concept drift requires continuous statistical vigilance:

• ADWIN (Adaptive Windowing): Dynamically adjusts window size to detect distribution changes in prediction error rates
• DDM (Drift Detection Method): Tracks online error rate and signals drift when error exceeds a calculated threshold
• Page-Hinkley Test: Sequential analysis technique detecting abrupt changes in model performance metrics
• Model-based detection: Comparing predictions from a static reference model against a continuously retrained model

Unaddressed concept drift directly undermines maintenance strategy:

• False negatives increase: Failures occur without warning as the model no longer recognizes degradation patterns
• False positives escalate: Unnecessary maintenance is triggered, wasting resources and causing downtime
• RUL estimates diverge: Remaining Useful Life predictions become progressively inaccurate
• Trust erosion: Operators lose confidence in the system, reverting to reactive maintenance approaches

Effective countermeasures against concept drift:

• Online learning: Incrementally update model parameters as new labeled data arrives
• Ensemble methods: Maintain a diverse set of models and weight predictions based on recent performance
• Sliding window retraining: Periodically retrain on the most recent data window while discarding obsolete samples
• Trigger-based adaptation: Automatically initiate model updates when drift detectors signal a significant change
• Feature engineering robustness: Design features that are invariant to known environmental or operational shifts

The overarching degradation of a model's predictive performance over time due to changes in the production environment. While concept drift is a specific cause, model drift can also result from data drift (changes in input distributions) or label drift (changes in target distributions). Monitoring for model drift requires tracking statistical properties of predictions against a baseline established during training.

A change in the distribution of the input features P(X) between training and production. Unlike concept drift, the relationship P(y|X) remains stable, but the model encounters data from regions it was never trained on. Common causes include:

• Sensor recalibration shifting value ranges
• New raw material suppliers altering vibration signatures
• Seasonal temperature changes affecting thermal readings Detected using Kolmogorov-Smirnov tests or Population Stability Index (PSI).

A specific type of data drift where the input distribution changes but the conditional distribution of the output given the input remains constant. In manufacturing, this occurs when a machine operates in a new regime—such as higher RPMs—that was underrepresented in training data. The failure patterns at those RPMs are the same, but the model has never seen them. Addressed through importance weighting or retraining on stratified samples.

A change in the distribution of the target variable P(y) without any change in the input-output relationship. In predictive maintenance, this manifests as a sudden increase in the base rate of failures due to a bad batch of raw materials. The model's conditional predictions remain correct, but its overall precision-recall balance shifts. Corrected by adjusting decision thresholds or recalibrating output probabilities.

The algorithmic identification of abrupt shifts in the statistical properties of a sensor stream, signaling a transition to a new degradation phase. Unlike gradual concept drift, change points represent instantaneous regime changes—such as a bearing seizing or a tool breaking. Algorithms like PELT (Pruned Exact Linear Time) and Bayesian Online Change Point Detection are critical for triggering immediate model retraining or fallback logic.

The strategy governing how frequently a predictive model is updated in response to detected drift. Approaches include:

• Periodic retraining: Scheduled updates regardless of drift signals
• Trigger-based retraining: Automated pipelines initiated when drift metrics exceed thresholds
• Online learning: Continuous weight updates from streaming data Each approach balances computational cost against the risk of stale predictions causing missed failure detections.