Challenge¶
The client, a leading Czech automotive manufacturer, operates one of the most modern automotive production lines in Central Europe in Central Bohemia. With daily production exceeding 2,500 vehicles and thousands of robotic workstations, the reliability of the production line is directly proportional to revenue — every hour of unplanned downtime represents a loss of approximately CZK 4.2 million.
In 2024, the manufacturer faced several fundamental problems in maintenance and production management. Maintenance was conducted primarily reactively or according to fixed time intervals — robots and CNC machines were serviced every X thousand operating hours regardless of the actual condition of components. This led either to the premature decommissioning of functional parts (unnecessary costs) or to unexpected failures between planned service windows.
The existing SCADA systems did collect operational data, but it was locked in silos of individual production segments (body shop, paint shop, assembly). There was no unified digital model of the entire production line that would allow simulating the impact of changes, predicting failures across segments, or optimising the production flow as a whole.
The manufacturer’s IT management, in cooperation with the production engineering department, issued a project for the creation of a digital twin of the production line with integrated predictive maintenance. CORE SYSTEMS was selected as the main implementation partner based on prior experience with industrial IoT and the Azure platform.
Solution¶
CORE SYSTEMS designed a solution called TwinLine — a digital twin platform connecting the physical production line with its virtual model in real time. The platform covers three main areas:
Real-time Digital Twin: Creation of a dynamic 3D model of the production line in Azure Digital Twins, reflecting the current state of every machine, robot, and conveyor system. The model encompasses 847 key assets and more than 15,000 sensor points. Each asset has its digital counterpart with current operating parameters, maintenance history, and predicted component lifespan.
IoT Sensor Mesh: Installation and integration of 3,200 additional IoT sensors (vibration, temperature, acoustic emission, current consumption) on critical assets where the existing SCADA instrumentation did not provide sufficient data granularity for predictive analytics. The sensors communicate via an industrial MQTT broker to Azure IoT Hub.
Predictive Maintenance Engine: Development of an ML pipeline for failure prediction based on degradation models — the system analyses trends in vibrations, temperature profiles, and other parameters and predicts the remaining useful life (RUL) of critical components with a horizon of 2–6 weeks.
Implementation took place in close cooperation with the manufacturer’s maintenance teams, who provided domain expertise for labelling historical data and validating predictive models.
Architecture¶
The TwinLine architecture is hybrid — edge computing in the manufacturing plant processes data with low latency, while the cloud layer in Azure handles analytics, ML training, and long-term storage.
The edge layer runs on Dell EMC PowerEdge industrial servers located directly in the production halls. Each segment (body shop, paint shop, final assembly, engine assembly) has its own edge cluster with Kubernetes (K3s). The edge layer receives data from MQTT brokers, performs filtering, aggregation, and the first level of anomaly detection using lightweight ML models (ONNX runtime). Critical alarms are generated locally with a latency below 500 ms.
The communication layer uses Azure IoT Hub as the central ingestion point for all sensor data. Data flows through IoT Hub into Event Hubs (partitioned Kafka-compatible stream) and further into two parallel branches — a hot path for real-time processing and a cold path for batch analytics.
The hot path is processed by Azure Stream Analytics in combination with custom Flink jobs on Databricks for complex correlation across production segments. For example, detecting a chain effect — slowdown of a robot in the body shop affecting the paint shop’s cycle time.
The cold path stores raw data in Azure Data Lake Storage Gen2 in Delta Lake format. The Databricks Lakehouse platform serves as the analytics engine for training ML models, historical analyses, and reporting. The feature store retains pre-computed features for predictive models.
Azure Digital Twins maintains a graph of relationships between assets, spaces, and processes defined in DTDL (Digital Twins Definition Language). The model is synchronised with the real state of the line at 1-second intervals. The platform’s API enables what-if scenario simulations — for example, the impact of a specific robot’s failure on the overall production flow.
The presentation layer combines Grafana dashboards for maintenance teams (asset status, failure predictions, service window planning) and a custom web application (React + Three.js) for 3D visualisation of the digital twin intended for production management.
All infrastructure is defined as code (Terraform) and deployed via a CI/CD pipeline in Azure DevOps.
Results¶
The TwinLine platform was deployed to pilot operation in September 2024 on the final assembly line and to full operation across all segments in January 2025. Results after the first six months of full operation:
Predictive maintenance: The system successfully predicted 87% of critical component failures with a lead time of at least 14 days. This enabled the scheduling of maintenance into planned service windows (weekends, night shifts) and the elimination of most unplanned downtime during production hours.
Downtime reduction: Unplanned production line downtime decreased by 45% — from an average of 23 hours per month to 12.6 hours. The remaining unplanned downtime is primarily caused by external factors (material deliveries, energy supply).
OEE (Overall Equipment Effectiveness): The overall effectiveness of the production line rose from 78% to 91%, placing the plant at the level of the best plants within the automotive group. The improvement comes from higher availability (fewer outages), better performance (cycle time optimisation), and lower scrap rates.
Financial impact: Annual savings of CZK 32 million in direct maintenance costs — the shift from time-based to condition-based maintenance reduced spare parts consumption by 28% and optimised the use of maintenance capacity. Indirect savings from higher production line productivity are estimated at an additional CZK 85 million per year.
Production flow optimisation: Simulations on the digital twin identified bottlenecks in the production process — adjusting the sequencing at two robotic workstations in the body shop increased throughput by 7% without any hardware investment.
Data culture: The project catalysed a change in attitude towards data in production. Maintenance teams actively use dashboards and predictions when planning work. Production engineers simulate process changes on the digital twin before implementation on the physical line.
Technology¶
The TwinLine technology stack is optimised for the industrial environment with an emphasis on reliability and low latency:
- Azure IoT Hub — central ingestion point, device management, device-to-cloud messaging
- Azure Digital Twins — asset relationship graph, DTDL models, real-time synchronisation
- Databricks (Azure) — ML training, feature store, Delta Lake, batch analytics
- Apache Kafka (Event Hubs) — sensor data streaming, event-driven architecture
- K3s (Edge Kubernetes) — orchestration of edge workloads in production halls
- MQTT (Mosquitto) — industrial messaging protocol for IoT sensors
- Grafana — operational dashboards for maintenance and production
- Python (scikit-learn, PyTorch) — predictive models, degradation analyses
- Terraform — Infrastructure as Code for the Azure environment
- Azure DevOps — CI/CD pipeline, Git repositories, project management
The collaboration between CORE SYSTEMS and the manufacturer continues with the expansion of the platform to the Kvasiny and Vrchlabí plants and integration with the MES system for automatic production rescheduling based on predictions.