Model Drift

A change in the statistical distribution of the input sensor data (features) fed into the model.

• Example: A vibration sensor gradually decalibrates, reporting consistently lower amplitudes than the actual physical state.
• Example: A new raw material supplier introduces a different viscosity, shifting the operating torque baseline.
• Impact: The model sees data patterns it has never encountered, leading to nonsensical or inaccurate predictions even if the fundamental physics of failure haven't changed.

A change in the fundamental relationship between the input data and the target prediction (e.g., failure). The same sensor reading now means something different.

• Example: A software update to the motor controller changes the RPM response to a given current draw. The old relationship 'high current = imminent bearing failure' is no longer valid.
• Example: A new maintenance policy replaces parts proactively, meaning a 'high temperature' reading no longer correlates with a run-to-failure event.
• Impact: This is the most dangerous form of drift because the model's core logic becomes invalid, requiring retraining on new historical data.

A shift in the distribution of the model's output predictions over time, often monitored without immediate ground-truth labels.

• Example: A Remaining Useful Life (RUL) model suddenly starts predicting failure in 10 days for 90% of the fleet, whereas historically it predicted 30-90 days.
• Detection: Monitored via statistical tests (e.g., Population Stability Index) on the prediction scores themselves.
• Impact: Serves as a leading indicator that either Data or Concept Drift has already occurred upstream, triggering an investigation before a false alert causes a costly shutdown.

A specific subtype of drift caused by the natural passage of time, including seasonal patterns or tool wear that were not captured in the training window.

• Example: A model trained in winter fails to account for increased ambient humidity in summer, which affects cooling rates and temperature sensor readings.
• Example: Gradual tool wear over 12 months changes the acoustic signature of a CNC machine, drifting far from the 'healthy' baseline established at installation.
• Mitigation: Requires continuous learning pipelines or scheduled retraining cycles that incorporate recent seasonal data to adapt to cyclical patterns.

A divergence between the proxy metric used to label data for training and the actual physical state of the asset.

• Example: A model is trained to predict failure based on a 'maintenance work order' label. If the maintenance team changes their logging behavior (e.g., logging pre-emptive repairs as 'failures'), the target definition drifts.
• Example: A quality inspection model's 'ground truth' is a human inspector. If a new inspector with different standards is hired, the labels shift.
• Impact: The model's accuracy appears to drop, but the real issue is the ground truth definition has changed, requiring a label reconciliation process.

A self-inflicted drift where the model's own predictions influence the future data it receives, creating a runaway feedback cycle.

• Example: A prescriptive maintenance model recommends reducing the speed of a motor to extend its life. The subsequent sensor data reflects this 'de-tuned' state, and the model never sees high-stress failure patterns again, becoming blind to them.
• Example: A model predicts a high risk of failure, prompting an immediate repair. The historical record now shows 'prediction = high risk, outcome = no failure', corrupting future training data.
• Mitigation: Requires counterfactual logging and bandit testing to understand what would have happened without the intervention.

A predictive maintenance model trained on vibration signatures from a specific steel alloy suddenly underperforms. The procurement team switched to a cheaper supplier with a slightly different grain structure. The statistical properties of the vibration data shift abruptly, causing the model to misclassify normal operation as anomalous. This is a classic case of covariate shift where the input distribution P(X) changes but the failure definition remains constant.

Over six months, a CNC machine's spindle bearings degrade naturally. The baseline vibration amplitude slowly increases. A model trained on the initial 'healthy' state begins to flag every operation as a warning. This gradual concept drift occurs because the relationship between the sensor reading and the 'healthy' label has evolved. Without online retraining or a dynamic Health Index, the model becomes useless, generating alert fatigue for maintenance engineers.

A foundry's predictive model for hydraulic pump failure exhibits high accuracy in winter but fails every summer. The increased ambient temperature alters the fluid viscosity, shifting the baseline operating pressure. This cyclical or seasonal drift requires the model to incorporate contextual features like ambient temperature. A static model without this feature will suffer from recurring performance degradation annually.

To meet a surge in demand, a production line manager increases the operating speed of a conveyor motor by 15%. The predictive model, trained on data at the original speed, now sees entirely new frequency patterns via Fast Fourier Transform (FFT). The statistical relationship between the sensor input and the failure output is broken, a phenomenon known as concept drift. The model requires transfer learning to adapt to the new regime.

An accelerometer monitoring a critical turbine slowly loses calibration over two years. The raw amplitude readings drift downward by 5%. The predictive model, interpreting this as a 'healthier' state, fails to predict a catastrophic bearing failure. This is data drift at the source level. It highlights the necessity of data observability and automated change point detection to distinguish between sensor decay and genuine asset health improvement.

A Digital Twin of a robotic arm is used to simulate degradation for Remaining Useful Life (RUL) prediction. After a physical repair replaces a joint with a non-OEM part, the physics-based simulation no longer matches reality. The twin's predictions drift from the physical asset's behavior. This virtual concept drift requires continuous sensor fusion and model adaptation to resynchronize the simulation with the physical world.

A specific type of model decay where the statistical relationship between the input data and the target variable changes. In predictive maintenance, this occurs when the same vibration signature that once indicated normal operation now precedes an imminent failure due to a new operating regime. Unlike data drift, concept drift invalidates the fundamental P(y|X) mapping, requiring model retraining rather than simple input recalibration.

A change in the distribution of the input features P(X) without a corresponding change in the relationship to the target. For example, a seasonal temperature shift alters the baseline sensor readings across a factory floor, but the correlation between high temperature and bearing failure remains constant. Detected via statistical tests like Kolmogorov-Smirnov or Population Stability Index (PSI).

The algorithmic identification of abrupt shifts in the statistical properties of a streaming sensor time-series. Unlike gradual drift, change points mark instantaneous transitions—such as a step-change in vibration amplitude after a tool crash. Algorithms like PELT (Pruned Exact Linear Time) and Bayesian Online Change Point Detection segment data into homogeneous regimes for root cause analysis.

The mathematical representation of how a system's health deteriorates over time, forming the basis for Remaining Useful Life (RUL) prediction. Degradation models—such as exponential, Paris-Erdogan crack growth, or Wiener processes—explicitly parameterize the wear trajectory. When the physical degradation mechanism changes, the model's assumptions drift, necessitating a structural update to the underlying equations.

A specific form of data drift where the distribution of input variables changes between training and inference, but the conditional distribution P(y|X) remains stable. In manufacturing, this manifests when a model trained on a specific machine type is deployed on a different asset class with distinct operational envelopes. Addressed through importance weighting or domain adaptation techniques.

The continuous observability layer that tracks prediction distributions, feature drift, and performance metrics in production. Effective monitoring for predictive maintenance requires calculating the Population Stability Index (PSI) on input features, tracking the Kullback-Leibler divergence of prediction outputs, and logging ground-truth failure events to compute precision and recall degradation over rolling windows.