Industrial AI Architecture: The Engineering Playbook for Edge-Cloud AI in Manufacturing, Energy and Heavy Industry

The industrial AI opportunity is enormous. Research consistently projects that AI-driven automation, quality intelligence, and predictive maintenance will unlock trillions of dollars in value across manufacturing, energy, and heavy industry over the next decade. The technology is mature. The use cases are proven. The ROI data is compelling.

And yet Gartner estimates that approximately 73% of industrial AI proof-of-concept projects never reach production deployment. The majority do not fail because the AI models are wrong. They fail because the architecture is wrong.

The architecture decision is where the long-term success or failure of an industrial AI programme is determined. A model that achieves 94% accuracy in a cloud-connected test environment but cannot run inference during a network outage in an underground mining facility is not a deployed AI system - it is an expensive experiment. A computer vision system that produces high-quality defect classifications but cannot push its findings to the MES for automated non-conformance record creation has not solved a business problem - it has created a new data island.

This white paper presents Ombrulla's Industrial AI Reference Architecture - a battle-tested, four-layer framework that we have developed and refined across deployments in discrete manufacturing, process industries, oil and gas, power generation, and infrastructure asset management. It covers every layer of the stack: from edge compute and sensor integration through data pipelines and MLOps to enterprise application integration and business intelligence.

Whether you are designing your first industrial AI deployment or rearchitecting a programme that has stalled at pilot stage, this white paper gives you the engineering blueprint, the decision frameworks, and the economic models to move from architecture design to production impact.

Industrial AI architecture is the holistic system design that defines how artificial intelligence capabilities are distributed, connected, managed, and governed across a physical industrial environment - from the sensor on the machine to the dashboard on the executive's screen.

It is fundamentally different from enterprise IT architecture and from consumer AI application design in three critical ways:

• -
  Latency constraints are physical, not just technicalA predictive quality model that takes 800ms to return a classification result is too slow to intervene on a production line running at 120 parts per minute. The architecture must bring intelligence to where the latency budget exists - often that means the edge, not the cloud.
• -
  Data is multi-modal and messyIndustrial environments generate time-series sensor data, machine event logs, maintenance records, image feeds, acoustic signals, and vibration spectra - often simultaneously and from systems with no common data schema. The architecture must handle this heterogeneity from day one.
• -
  Operational continuity is non-negotiableA factory cannot pause production while a cloud AI model is being retrained. A pipeline inspection cannot wait for 4G connectivity to be restored. The architecture must guarantee operational continuity independent of network availability.

The Ombrulla Industrial AI Reference Architecture organises the full technology stack into four distinct but tightly integrated layers. Each layer has a specific responsibility domain, defined interfaces with adjacent layers, and clear engineering requirements that must be met for the overall system to function in a production industrial environment.

The edge compute layer is where industrial AI meets physical reality. It is the layer responsible for capturing raw operational data from machines, processes, and environments; running AI inference locally to generate actionable intelligence; and doing both with the latency, reliability, and connectivity independence that industrial operations require.

Getting this layer right is the single most critical architecture decision in industrial AI. A poorly specified edge layer will constrain the entire system above it - limiting the types of AI models that can run, the inference speeds achievable, and the operational continuity of the system.

Edge Compute Hardware Selection: IoT Hardware selection at the edge must balance AI inference capability with operational environment requirements. Ombrulla's hardware selection framework evaluates four axes:

On-Device AI Model Deployment: AI models deployed to edge hardware must be optimised for the available compute envelope. Ombrulla's standard edge model preparation pipeline includes four stages:

• -
  Model compression: Pruning and knowledge distillation reduce model parameter count by 40–70% with minimal accuracy loss on industrial defect detection tasks.
• -
  Quantisation: Converting model weights from FP32 to INT8 or FP16 precision reduces memory footprint by 4x and accelerates inference by 2–4x on integer-optimised NPUs.
• -
  Framework conversion: Converting from PyTorch or TensorFlow training format to TensorRT, ONNX Runtime, or TF Lite for edge-optimised inference execution.
• -
  Benchmarking and validation: Each compressed model is benchmarked against the production inference latency budget and validated against the held-out test set to confirm accuracy is maintained within acceptance thresholds.

Industrial environments are rich with data that is almost entirely inaccessible to AI systems in its raw form. PLCs speak proprietary protocols. Historians store data in formats designed for human review, not machine learning feature extraction. Sensors generate unlabelled time-series streams with no contextual metadata. The connectivity and data pipeline layer is the engineering discipline that transforms this raw operational data wealth into structured, contextualised, ML-ready datasets.

Industrial Protocol Integration

The connectivity layer must speak the native language of industrial control systems before it can aggregate and transform data. Ombrulla's standard protocol integration stack covers:

• -
  OPC-UA (IEC 62541): The preferred industrial standard for machine-to-machine communication, supporting both data access and alarm/event handling with a security model appropriate for OT environments.
• -
  MQTT over TLS: Lightweight publish-subscribe protocol preferred for IIoT sensor networks, remote assets, and constrained-bandwidth environments such as wireless sensor networks and satellite-connected field assets.
• -
  Modbus TCP/RTU: Ubiquitous legacy PLC protocol that remains the most widely deployed industrial communication standard globally; must be bridged to modern data pipelines via OPC-UA or MQTT gateway devices.
• -
  PROFINET / EtherNet/IP: Real-time industrial Ethernet protocols prevalent in discrete manufacturing; require certified gateway hardware for integration into data pipeline layers.
• -
  GigE Vision / USB3 Vision: Camera interface standards for industrial machine vision; provide high-bandwidth image data to edge AI inference engines.

Time-Series Data Architecture

The majority of industrial AI feature data is time-series in nature - vibration spectra, temperature trends, pressure waveforms, flow rates, and energy consumption profiles. The data pipeline layer must handle time-series data at scale with:

• -
  Time-series database (TSDB): Purpose-built storage for high-frequency sensor data. Ombrulla standardises on InfluxDB and TimescaleDB for on-premises deployments, and Azure Time Series Insights or AWS Timestream for cloud-native deployments.
• -
  Stream processing: Apache Kafka and Apache Flink provide the backbone for high-throughput, low-latency event stream processing - enabling real-time feature computation, anomaly detection triggers, and cross-sensor correlation at industrial data volumes.
• -
  Data contextualisation: Raw sensor readings acquire manufacturing context through enrichment with asset hierarchy, production order, shift, batch, and operator metadata - transforming a time-series value into a machine-learning-ready training sample.

The cloud intelligence layer is the industrial AI programme's brain. It is where models are trained on aggregated operational data, evaluated against business performance metrics, versioned and governed through a model registry, and deployed to the edge and application layers through automated CI/CD pipelines.

Organisations that treat this layer as an afterthought - manually managing models in Jupyter notebooks, deploying model files via FTP to edge servers, or running retraining jobs on an ad hoc basis - consistently find that their industrial AI programmes degrade over time as production data drifts from the training distribution. Industrial AI at production scale demands a mature MLOps practice.

MLOps Platform Architecture for Industrial AI

Ombrulla's MLOps architecture for industrial deployments is built on three platform principles:

• -
  Experiment reproducibility: Every training run is tracked with full lineage - dataset version, hyperparameters, environment, and evaluation metrics - using MLflow or Weights & Biases. Any model can be reproduced exactly from its tracking record.
• -
  Automated model evaluation pipelines: New model versions must pass automated quality gates - minimum accuracy thresholds per defect class, latency benchmarks on target edge hardware, and regression tests against a curated validation dataset - before they are eligible for production deployment.
• -
  Model registry governance: The model registry maintains the approved production model for each use case, the challenger model under evaluation, and a full version history with deployment records. Model promotion from challenger to production requires explicit approval workflow with defined business owner sign-off.

Model Retraining Triggers

In industrial environments, model performance degrades over time due to three primary causes - all of which require different retraining strategies:

An industrial AI system that produces high-quality predictions but cannot act on them automatically is a reporting tool, not an operational system. The application and integration layer is where AI intelligence becomes operational consequence - triggering maintenance work orders, updating quality dashboards, adjusting process parameters, and feeding digital twins with real-time asset condition data.

Enterprise System Integration Patterns

• -
  ERP Integration (SAP, Oracle, Microsoft D365): AI-generated quality findings and maintenance alerts push directly to ERP via certified API connectors, creating non-conformance records, initiating procurement requests for spare parts, and updating cost centre reporting with AI-generated quality metrics.
• -
  MES Integration: AI quality gate decisions are embedded directly into production order flows - a Critical severity defect detection automatically parks the production order and requires quality engineer review before the line restarts. Pass results update OEE calculations in real time.
• -
  SCADA / Process Control Integration (via OPC-UA): AI model outputs feed back to process control systems as soft sensors - adjusting setpoints, triggering alarms, or initiating safe shutdowns based on predicted process deviations.
• -
  CMMS Integration (IBM Maximo, Infor EAM, SAP PM): Predictive maintenance AI generates CMMS work orders with pre-populated asset ID, fault code, recommended action, and estimated time-to-failure data - eliminating manual maintenance planning overhead.
• -
  Digital Twin Platforms (Bentley iTwin, Siemens Industrial Metaverse): Real-time defect findings and asset condition scores update the digital twin model, enabling 3D spatial visualisation of asset health across a facility or asset network.

Computer vision AI represents one of the highest-value application categories within the Ombrulla Industrial AI Architecture. The ability to give machines the capacity to see - and to make quality, safety, and condition decisions from what they observe - addresses inspection and monitoring challenges that neither traditional sensor systems nor rule-based machine vision can solve economically at scale.

Camera Infrastructure Architecture

Embedding computer vision into an industrial AI architecture requires camera infrastructure decisions that go well beyond selecting a lens and sensor. The key architectural considerations are:

• -
  Lighting design: Computer vision model accuracy is highly sensitive to lighting consistency. Industrial lighting architectures must provide controlled, reproducible illumination - typically structured light, ring light, or coaxial illumination for reflective surfaces; overdiffuse backlight for silhouette-based inspection. Lighting control should be synchronised to camera trigger signals.
• -
  Camera-to-AI pipeline latency: The full camera capture-to-inference-to-output pipeline must be profiled end-to-end. In high-speed inspection scenarios, the camera interface, image pre-processing, inference, and output steps must collectively complete within the inspection window defined by line speed.
• -
  Multi-camera synchronisation: Assembly and structural inspection use cases often require simultaneous capture from multiple camera angles. Hardware synchronisation via trigger cables or PTP (IEEE 1588) network synchronisation ensures frames from multiple cameras capture the same moment in the production cycle.
• -
  Image data management: A 10 Megapixel camera capturing at 10 fps generates approximately 3.6 TB/day of raw image data. The architecture must define image retention policies - raw images, compressed thumbnails, and only defect findings - appropriate to storage budget and regulatory requirements.

Vision AI Model Types by Use Case

For a detailed breakdown of defect taxonomy by industry and annotation guidelines for training vision AI models, see Ombrulla's companion article: AI Visual Inspection Defect Taxonomy Library.

One of the most common questions Ombrulla's architecture team addresses in industrial AI engagement conversations is: 'Should we run this in the cloud or at the edge?' The honest answer is almost always: 'It depends on the use case' - and the decision matrix below provides the framework for making that determination systematically rather than based on vendor preference or existing infrastructure bias.

The field of MLOps - the engineering discipline of managing machine learning models in production - is still maturing in the industrial context. Most industrial organisations that have deployed AI models have done so with significant manual intervention at every stage of the model lifecycle. That approach does not scale.

Ombrulla's Industrial MLOps practice is built on four foundational capabilities that distinguish production-grade AI from research-grade deployments:

• -
  1. Continuous Training Pipelines: Production industrial AI models must be retrained on a cadence that matches the rate of change of the underlying process. A continuous training pipeline automates: data ingestion from the feature store; dataset versioning and splitting; distributed training job submission; automated evaluation against defined quality gates; and model registration upon successful evaluation. The entire pipeline runs on a defined schedule and can also be triggered by drift detection alerts.
• -
  2. OTA Model Deployment to Edge: Updating AI models on edge hardware distributed across multiple facilities requires an over-the-air (OTA) deployment pipeline that manages: model package creation and signing; staged rollout with canary deployment (deploy to 5% of edge nodes first, validate, then roll out to 100%); automatic rollback on performance degradation; and full audit trail of model versions deployed to each physical edge node. This is analogous to the OTA firmware update pipelines used in automotive and consumer electronics, adapted for industrial AI inference workloads.
• -
  3. Production Monitoring and Drift Detection: The Ombrulla monitoring framework collects three categories of production signals from deployed industrial AI models: data distribution statistics (PSI, KL divergence) computed on incoming feature data vs training baseline; model output statistics (class distribution, confidence score distributions) compared to historical baselines; and business outcome metrics (defect escape rate, false positive rate, inspection cycle time) from downstream business systems. Deviations beyond defined thresholds trigger automated alerts and, for critical degradation, automatic rollback to the previous production model.
• -
  4. Human-in-the-Loop Feedback: Industrial AI models improve fastest when they can learn from human expert corrections in production. Ombrulla implements human-in-the-loop (HITL) feedback loops that capture: inspector overrides of AI classifications (recorded as soft labels for review); quality engineer confirmations of AI-flagged defects (recorded as hard labels for retraining); and novel finding flags that route to the AI team for new class labelling. This continuous feedback mechanism creates a flywheel effect where model accuracy improves over time rather than degrading.

The business case for industrial AI architecture investment must be built on a rigorous Total Cost of Ownership model that captures the full cost of building, running, and evolving an AI programme - not just the software licence fee. At the same time, the value case must be grounded in operational data rather than vendor projections.

TCO Components for Industrial AI Architecture

Value Driver Framework

Industrial AI programmes deliver value through five distinct mechanisms, each of which should be quantified separately in the business case:

• -
  Defect escape prevention: The value of AI-detected defects caught before they reach customers or downstream processes. Quantify using historical warranty cost, scrap rate, and production rework data.
• -
  Unplanned downtime prevention: The value of AI-predicted failures caught before asset failure. Quantify using historical MTTR, average downtime cost (production loss + maintenance emergency premium), and OEE baseline.
• -
  Inspection labour productivity: The reduction in inspector time per inspection event multiplied by the fully loaded hourly cost and annual inspection volume.
• -
  Throughput improvement: The increase in production throughput enabled by faster, more accurate quality gates that eliminate conservative human over-rejection of borderline parts.
• -
  Data monetisation and compliance: The value of structured AI inspection records for regulatory compliance demonstration, insurance premium reduction, and asset resale value documentation.

Ombrulla's industrial AI implementation methodology is structured into five phases, each with defined deliverables, quality gates, and success criteria. The phased approach ensures that organisational capability is built alongside technical capability - the leading cause of single-use-case AI deployments that never scale is the absence of the MLOps and data engineering capability required to operationalise a second and third model.