Predictive Maintenance System with Sensor Drift Detection

Overview

Industrial predictive maintenance systems often assume that sensor data is always reliable. In real-world environments, sensors drift over time due to aging, calibration loss, or environmental factors, leading to false alarms or missed failures.

This project builds a drift-aware predictive maintenance system that:

• Detects sensor drift using unsupervised machine learning

• Corrects drifted sensor readings

• Predicts imminent machine failure using only trusted data

The system explicitly separates sensor health modeling from machine health prediction, making it more realistic and robust than standard approaches.

---

Problem Statement

Traditional predictive maintenance models:

• Treat all sensor readings as trustworthy

• Fail when sensors drift gradually

• Mix sensor faults with actual machine degradation

---

Goal of this project:

• Detect sensor drift first, then perform predictive maintenance using drift-corrected sensor signals.

---

Key Ideas

• Sensor drift ≠ machine failure

• Drift detection should be unsupervised

• Maintenance prediction should use validated sensor data

• Explainability is critical for real-world adoption

---

Dataset

• NASA C-MAPSS Turbofan Engine Degradation Dataset

• Multivariate time-series sensor data

• Multiple engines with full life cycles

• Widely used benchmark in predictive maintenance research

---

Methodology

1. Exploratory Data Analysis

• Visualized sensor behavior across engine life cycles

• Identified degrading, noisy, and low-variance sensors

• Validated observations using variance and trend analysis

---

2. Sensor Drift Simulation

• Injected gradual sensor drift into selected sensors

• Drift applied to only a subset of engines

• Drift starts mid-life to simulate realistic conditions

---

3. Sensor Drift Detection

• Unsupervised Isolation Forest

• Features used:

  • Sensor value

  • Temporal delta (rate of change)

• Outputs:

  • Drift score

  • Drift flag (healthy vs drifted)

---

4. Drift-Aware Cleaning

• Drifted sensor readings are not trusted

• Applied last-known healthy value (forward fill)

• Preserved temporal continuity without dropping data

---

5. Feature Engineering

• Rolling window features:

  • Mean

  • Standard deviation

  • Trend

• Captured degradation behavior over time

---

6. Predictive Maintenance Modeling

• Binary classification task:

• Will the engine fail in the next 30 cycles?

• Model used:

  • Random Forest Classifier

  • Trained only on drift-corrected features

---

7. Evaluation & Comparison

• Compared drift-aware system vs baseline (no drift handling)

• Drift-aware approach:

  • Improved failure recall

  • Reduced false maintenance alerts

---

8. Explainability

• Feature importance analysis

• Temporal trends identified as dominant predictors

• Predictions explained using feature behavior

---

System Architecture

This diagram illustrates the end-to-end workflow of the drift-aware predictive maintenance system, showing how sensor drift detection is decoupled from machine failure prediction.

---

Key Results

• Successfully detected gradual sensor drift

• Low false positives on clean engines

• Improved failure recall using drift-aware preprocessing

• Demonstrated clear benefit over naive baseline models

---

Why This Project Stands Out

• Goes beyond standard predictive maintenance

• Explicitly models sensor reliability

• Uses a two-stage ML system:

• Sensor health model

• Machine health model

• Mimics real industrial ML pipelines

• Strong focus on explainability and system design

---