Software-Defined SCADA Anomaly Detection with Mastra AI and Gemini 2.5 Flash
Modern industrial manufacturing plants are undergoing a major transition toward Software-Defined Controls (SDC) and virtualized Programmable Logic Controllers (vPLCs). As field devices emit continuous OPC UA streams of high-frequency pressure, temperature, and vibration telemetry, traditional threshold-based alarm systems often generate excessive false positives or fail to diagnose cross-system root causes. Industrial automation teams require intelligent, agentic telemetry processors that ingest stream windows, detect subtle multi-variate anomalies, and provide root-cause diagnostics with actionable mitigation steps. In this challenge, you will build a TypeScript-based industrial monitoring agent using Mastra AI integrated with Gemini 2.5 Flash. Your agent will analyze real-time time-series telemetry streams from virtualized PLC nodes, evaluate transient sensor deviations against historical operational baselines, and output structured diagnostic reports with safety-compliant control actions. By leveraging Mastra AI's workflow orchestration, tool execution, and stateful memory along with Gemini 2.5 Flash's rapid multimodal/long-context reasoning, you will create a scalable telemetry pipeline capable of preventing unplanned downtime while maintaining plant safety standards.
What you are building
The core problem, expected build, and operating context for this challenge.
Build a software-defined SCADA anomaly detection agent using Mastra AI and Gemini 2.5 Flash to evaluate OPC UA time-series streams and perform root-cause diagnostics.
How work is evaluated
Evaluates the Mastra AI telemetry agent against simulated SCADA time-series windows containing transient faults, assessing anomaly classification precision, root cause accuracy, and execution response time.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
schema_validity
Agent response must strictly conform to the expected diagnostic response schema
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
safety_gate_verification
Critical bearing anomalies must trigger automatic speed limit reduction or trip alert
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
root_cause_accuracy
Accuracy of root cause subsystem classification against ground truth fault labels • target: 0.9 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
detection_latency_ms
Workflow execution latency from telemetry ingestion to output recommendation • target: 800 • range: 0-3000
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Implement a streaming OPC UA telemetry ingestion workflow in Mastra AI
Integrate Gemini 2.5 Flash for high-throughput time-series anomaly evaluation and root-cause analysis
Construct custom Mastra AI agent tools for SCADA register lookups and control action recommendation
Validate telemetry safety thresholds and generate structured JSON alert payloads
Reference links and supporting material
Simulated OPC UA stream dataset generated from a physical digital twin of a 4-stage motor-compressor skid, containing 10,000 continuous readings with simulated thermal breakdown, bearing wear, and line pressure loss events.
Challenge setup
Evaluates the Mastra AI telemetry agent against simulated SCADA time-series windows containing transient faults, assessing anomaly classification precision, root cause accuracy, and execution response time.
Challenge input
JSON payload with windowed sensor telemetry (vibration, bearing_temp, hydraulic_pressure, drive_rpm) and controller metadata
Mastra AI Framework
Provides TypeScript-native agent, tool, and workflow orchestration with stateful memory
Gemini 2.5 Flash
Provides ultra-fast multimodal and long-context reasoning for high-throughput SCADA telemetry analysis
Evaluated output
JSON object with anomaly_detected (boolean), root_cause_subsystem (string), confidence_score (float), and recommended_action (object)
- Agent response must strictly conform to the expected diagnostic response schema
- Critical bearing anomalies must trigger automatic speed limit reduction or trip alert
- Accuracy of root cause subsystem classification against ground truth fault labels • target: 0.9 • range: 0-1
- Benchmark: Industrial AI Maintenance Benchmark (IndAI-Bench)
- Root Cause Accuracy target: 0.9
- 1 public reference case
- Docker execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Google Gemini · Optional
- Google Gemini Code Assist · Optional
- Gemini 2.5 Flash · Required
Evaluation contract
- schema_validity · Weight 1
- safety_gate_verification · Weight 1
- root_cause_accuracy · Weight 1
- detection_latency_ms · Weight 1
Metrics
- root_cause_accuracy · Target 0.9
- detection_latency_ms · Target 800
Benchmarks
- Industrial AI Maintenance Benchmark (IndAI-Bench)
- robotics
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON payload with windowed sensor telemetry (vibration, bearing_temp, hydraulic_pressure, drive_rpm) and controller metadata
Output: JSON object with anomaly_detected (boolean), root_cause_subsystem (string), confidence_score (float), and recommended_action (object)
Execution harness: Docker
Evaluation tools
- Mastra AI Framework · Agent Orchestration Framework
- Gemini 2.5 Flash · LLM Inference Engine
Recipe freeze marker
No recipe freeze time is recorded.