AI-Powered Maintenance Event Prediction for a Wind Turbine Services Company
Challenge
Plexteq partnered with a wind turbine maintenance and asset management company in the Netherlands to rebuild the predictive maintenance core of their monitoring platform - replacing threshold-based alerting with a real-time, ML-driven failure prediction engine that increased the prediction rate by 37% and cut telemetry ingestion latency from hours to seconds.
Key Challenges
- Alarm Noise Over Signal Threshold- and rule-based alerting produced excessive false positives during normal operational variance while missing gradual, multi-signal degradation patterns - eroding analyst trust and burying real failure precursors in alarm noise.
- Stale, Fragmented Telemetry SCADA telemetry from hundreds of turbines - thousands of sensor channels covering gearbox, generator, main bearing, pitch, and hydraulic systems at 10-minute and 1-second resolutions - arrived through heterogeneous feeds and was processed only in nightly batches, making real-time fleet assessment impossible.
- No Reliable Failure Labels - Failure events are rare and inconsistently labelled: maintenance work orders, SCADA alarm logs, and component replacement records lived in separate systems with no reliable linkage between a recorded failure and the sensor history that preceded it.
- Multi-OEM Generalisation The fleet spanned multiple turbine OEMs, models, and age profiles, so a single global model underperformed - the solution needed to generalise across turbine types while remaining sensitive to unit-specific normal behaviour.
Plexteq partnered with a wind turbine maintenance and asset management company in the Netherlands to rebuild the predictive maintenance core of their monitoring platform - replacing threshold-based alerting with a real-time, ML-driven failure prediction engine that increased the prediction rate by 37% and cut telemetry ingestion latency from hours to seconds.
Key Challenges
- Alarm Noise Over Signal Threshold- and rule-based alerting produced excessive false positives during normal operational variance while missing gradual, multi-signal degradation patterns - eroding analyst trust and burying real failure precursors in alarm noise.
- Stale, Fragmented Telemetry SCADA telemetry from hundreds of turbines - thousands of sensor channels covering gearbox, generator, main bearing, pitch, and hydraulic systems at 10-minute and 1-second resolutions - arrived through heterogeneous feeds and was processed only in nightly batches, making real-time fleet assessment impossible.
- No Reliable Failure Labels - Failure events are rare and inconsistently labelled: maintenance work orders, SCADA alarm logs, and component replacement records lived in separate systems with no reliable linkage between a recorded failure and the sensor history that preceded it.
- Multi-OEM Generalisation The fleet spanned multiple turbine OEMs, models, and age profiles, so a single global model underperformed - the solution needed to generalise across turbine types while remaining sensitive to unit-specific normal behaviour.
Solution
Real-Time Telemetry Ingestion with Apache Flink
The foundation of the engagement was replacing the nightly batch pipeline with a streaming ingestion layer. Plexteq designed and deployed an Apache Flink pipeline consuming SCADA telemetry from the fleet through Pub/Sub, performing in-flight validation, deduplication, sensor-level quality scoring, and unit normalisation before landing the data in BigQuery.
The same pipeline reconciled historical data: five years of raw SCADA archives, alarm logs, and maintenance work orders were backfilled into BigQuery.
Failure Labelling and Feature Engineering
Before any modelling could begin, Plexteq's data engineers built a failure event registry by programmatically linking component replacement records and maintenance work orders to SCADA alarm sequences and turbine downtime windows, resolving conflicts through a set of validation rules co-designed with the client's senior maintenance engineers.
Feature engineering was performed in BigQuery SQL for large-scale historical windows and in pandas for experimentation, ultimately producing over 120 engineered features per turbine.
Layered Anomaly Detection and Failure Prediction Models
Rather than a single model, Plexteq delivered a layered architecture in which each layer answers a progressively more specific question:
- Unsupervised anomaly detection (Isolation Forest + PCA baseline, BigQuery ML).
- LSTM autoencoders for temporal degradation patterns (TensorFlow)
- Supervised failure classification (gradient-boosted trees, PyTorch experimentation on Vertex AI) On top 4. Remaining useful life estimation (survival analysis)
MLOps on Vertex AI
The full model lifecycle was industrialised on Vertex AI: training pipelines defined as Vertex AI Pipelines, experiment tracking and model registry for every model generation, automated evaluation gates comparing challenger models against production on held-out failure events, and continuous monitoring for feature drift as fleet composition and seasonal conditions change.
Analyst Workflow Integration
Predictions were integrated into the client's existing monitoring platform rather than delivered as a standalone tool. Each alert carries the component-level failure probability per horizon, the SHAP-based explanation, the RUL estimate, and links to the underlying sensor traces - turning alarm triage from a manual investigation into a review-and-schedule workflow.
Real-Time Telemetry Ingestion with Apache Flink
The foundation of the engagement was replacing the nightly batch pipeline with a streaming ingestion layer. Plexteq designed and deployed an Apache Flink pipeline consuming SCADA telemetry from the fleet through Pub/Sub, performing in-flight validation, deduplication, sensor-level quality scoring, and unit normalisation before landing the data in BigQuery.
The same pipeline reconciled historical data: five years of raw SCADA archives, alarm logs, and maintenance work orders were backfilled into BigQuery.
Failure Labelling and Feature Engineering
Before any modelling could begin, Plexteq's data engineers built a failure event registry by programmatically linking component replacement records and maintenance work orders to SCADA alarm sequences and turbine downtime windows, resolving conflicts through a set of validation rules co-designed with the client's senior maintenance engineers.
Feature engineering was performed in BigQuery SQL for large-scale historical windows and in pandas for experimentation, ultimately producing over 120 engineered features per turbine.
Layered Anomaly Detection and Failure Prediction Models
Rather than a single model, Plexteq delivered a layered architecture in which each layer answers a progressively more specific question:
- Unsupervised anomaly detection (Isolation Forest + PCA baseline, BigQuery ML).
- LSTM autoencoders for temporal degradation patterns (TensorFlow)
- Supervised failure classification (gradient-boosted trees, PyTorch experimentation on Vertex AI) On top 4. Remaining useful life estimation (survival analysis)
MLOps on Vertex AI
The full model lifecycle was industrialised on Vertex AI: training pipelines defined as Vertex AI Pipelines, experiment tracking and model registry for every model generation, automated evaluation gates comparing challenger models against production on held-out failure events, and continuous monitoring for feature drift as fleet composition and seasonal conditions change.
Analyst Workflow Integration
Predictions were integrated into the client's existing monitoring platform rather than delivered as a standalone tool. Each alert carries the component-level failure probability per horizon, the SHAP-based explanation, the RUL estimate, and links to the underlying sensor traces - turning alarm triage from a manual investigation into a review-and-schedule workflow.
Results
Measurable Impact
- 37% increase in failure prediction rate versus the legacy rule-based system, validated over a six-month shadow-run on live fleet data
- Telemetry ingestion and scoring latency reduced from up to 24 hours to under 15 seconds end-to-end
- Average early-warning lead time of 21 days for gearbox, generator, and main bearing failures
- 52% reduction in false-positive alarms, cutting analyst triage workload by more than half
- Historical analysis over five years of fleet data reduced from multi-day manual exports to interactive BigQuery queries
- Fully automated weekly retraining with drift-triggered updates - no manual model maintenance required from the client's team
Measurable Impact
- 37% increase in failure prediction rate versus the legacy rule-based system, validated over a six-month shadow-run on live fleet data
- Telemetry ingestion and scoring latency reduced from up to 24 hours to under 15 seconds end-to-end
- Average early-warning lead time of 21 days for gearbox, generator, and main bearing failures
- 52% reduction in false-positive alarms, cutting analyst triage workload by more than half
- Historical analysis over five years of fleet data reduced from multi-day manual exports to interactive BigQuery queries
- Fully automated weekly retraining with drift-triggered updates - no manual model maintenance required from the client's team