Solution Delivered
​
↳ 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. Flink's stateful stream processing was used to compute windowed aggregates continuously — rolling means, variances, exponentially weighted trends, and cross-signal ratios (e.g., gearbox bearing temperature relative to power output and ambient temperature) — so that model-ready features were available within seconds of a sensor reading being emitted, rather than the following morning.

The same pipeline reconciled historical data: five years of raw SCADA archives, alarm logs, and maintenance work orders were backfilled into BigQuery, giving the data science workstream a unified, queryable history of fleet behaviour and failure events for the first time.
​
↳ 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. This produced a labelled dataset of confirmed failure events per component class — gearbox, generator, main bearing, pitch system, and hydraulics — each paired with its full preceding sensor history.

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: lag features across multiple horizons, rolling statistical windows, temperature differentials between paired components, power-curve residuals (deviation between actual and expected output given wind conditions), and operational context flags for curtailment, icing conditions, and grid events.
​
↳ 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).
  As a first screening layer, Isolation Forest models trained per turbine type flag operating states that deviate from learned normal behaviour. PCA-based reconstruction error served as the interpretable statistical baseline, mirroring the approach validated in Plexteq's earlier wind turbine anomaly detection work. Running these models directly in BigQuery ML kept the screening layer close to the data and inexpensive to retrain fleet-wide.
   

• LSTM autoencoders for temporal degradation patterns (TensorFlow)
  The core detection layer consists of LSTM autoencoder networks trained on healthy-operation sequences per component subsystem. The autoencoder learns to reconstruct normal multivariate sensor sequences; rising reconstruction error over time is a robust signal of gradual degradation that no single-threshold rule can capture. Per-turbine calibration layers adapted the shared architecture to unit-specific baselines, solving the multi-OEM generalisation challenge.
  ​

• Supervised failure classification (gradient-boosted trees, PyTorch experimentation on Vertex AI)
  On top of the anomaly signals, gradient-boosted tree classifiers were trained on the labelled failure registry to estimate the probability of a component-class failure within 7-, 14-, and 30-day horizons. SHAP values were computed for every prediction, so each alert reaching an analyst is accompanied by a ranked, human-readable list of the sensor behaviours driving it.
   

• Remaining useful life estimation (survival analysis)
  For components flagged by the upper layers, a survival analysis model (Cox proportional hazards with gradient-boosted extensions) estimates remaining useful life distributions, allowing the planning team to schedule interventions into existing vessel and crew rotations rather than dispatching emergency call-outs.

↳ 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. Retraining runs weekly per turbine type and is triggered automatically when drift metrics exceed configured bounds. Models are served through Vertex AI endpoints consumed by the streaming scoring layer, so every incoming telemetry window is scored against the current production models in near real time.

↳ 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.