Heavy electric vehicle fleets depend on accurate battery state estimation for safe operation, maintenance planning, and economic optimisation. Battery management systems embedded in individual vehicles face well-documented limitations in accuracy and robustness, while academic research has produced advanced estimation algorithms evaluated in isolation rather than as deployable systems. To the best of the author’s knowledge, no prior work has combined ISO 23247 digital twin standard alignment, explicit machine learning infrastructure design, and an open-source proof of concept targeting heavy electric vehicle fleets. This thesis designs and validates a digital twin architecture for battery health monitoring in cloud-connected heavy electric vehicle fleets. The architecture is organized into five layers: ingestion, data, ML infrastructure, service, and application, mapped to the ISO 23247 framework with documented extensions for battery-specific requirements. Three prediction services are specified: real-time State of Charge estimation using a CNN-LSTM model, daily State of Health monitoring using Gaussian Process Regression, and on-demand energy consumption forecasting. The first two services are implemented as a proof of concept using PyBaMM physics-based simulation for training data generation. Empirical evaluation shows that SoC warm cache p95 latency of 21 ms satisfies the 1-second target with a 48× margin, and that a single SoC service instance supports approximately 177 vehicles at 1 Hz. A Redis prediction cache reduces SoH inference overhead by a factor of 581, making the 8.9-second GPR cold inference viable as a daily batch process. Partial validation against a forklift LFP degradation dataset identifies two requirements for production deployment: an ageing component in the training data for long-term degradation tracking and capacity-normalised features to generalise across cell formats. The architecture provides documented extension points for federated learning, edge computing, and additional prediction services, and is designed to generalize beyond heavy electric vehicles to other domains requiring real-time state estimation from sensor telemetry.