ML
Learn patterns and predict
- Input
- Consumption, sensor and equipment history
A model is trained on equipment data, then analyses future trends or unusual behaviour.
- Output
- Consumption forecasts, anomaly detection and health indicators
Part 8 of 12
AI & offline maps
A Persian-speaking assistant that knows the project’s data and works on the user’s behalf, the ARG AI backend for models and prediction, and the offline map server for sites without internet.
ML · RAG · LLM · AGENT
Industrial AI deployment guideAI alongside the digital twin’s data and tools
The project-connected assistant and the independent ARG_AI backend serve complementary roles: one helps users understand the situation and work inside the software; the other provides model training, analysis, prediction and result monitoring.
ML
Learn from data and measure results
Prepare history → train models → predict and detect → evaluate and monitor
LLM + RAG + Tools
Understand the project, retrieve knowledge and act
User intent → relevant knowledge and project tools → analysis or proposed change → user confirmation
RAG
Retrieve project knowledge
- Input
- Project documents, catalogs and device templates
Knowledge retrieval finds passages relevant to the question and equipment and supplies them to the assistant.
- Output
- Answers grounded in documents and the project context
LLM
Explain the findings clearly
- Input
- User question, project data and tool results
The assistant explains findings, summarises information and uses system tools to investigate the question.
- Output
- Engineering dialogue, reports and reviewable recommendations
Agent
Work inside the system
- Input
- User intent and the current page or project context
Tools read equipment state, alarms and history; proposed creation or changes enter the user confirmation workflow.
- Output
- Multi-step investigation and proposed dashboards or logic
The LLM and RAG assistant: from project knowledge to system design
RAG supplies relevant knowledge, the LLM interprets the request and results, and agent tools move the work forward inside the system.
Understand equipment and project knowledge
The assistant draws on device catalogs, SNMP and Modbus templates, OIDs and uploaded project documents. RAG combines keyword and semantic retrieval to answer in the context of the relevant equipment.
Create structures, dashboards and 3D scenes
A user request can become a proposal for buildings, floors, spaces, devices and points, or for dashboard creation and cloning, 2D elements, scenes and 3D objects. System tools and user confirmation carry out the workflow.
Help design logic and scenarios
The assistant reads and checks existing logic and proposes creating or editing Rules, Limits, Scenarios, Scripts and State Machines. Command-producing logic starts disabled and requires a separate activation confirmation.
Maintain context as the work continues
Conversation history, summaries and user or project memory help continue the work. Equipment aliases, preferences and routines can be retained, pinned or forgotten.
IN THE PROJECT WORKSPACE
From an alarm question to equipment investigation
Ask the assistant to inspect active alarms, offline equipment or a point’s history. The request carries project context and tools read the required information from the system. Proposed changes are presented for user confirmation.
IN THE INDEPENDENT ARG_AI BACKEND
From equipment history to a model you can evaluate
For energy forecasting, historical data trains the model; results are checked against test data and recorded with error metrics. Anomaly detection learns normal behaviour to identify observations that need investigation.
The model lifecycle: training through evaluation
- 1
Prepare the data
Equipment data becomes training features. Energy forecasting uses time windows and hour/day patterns.
- 2
Train and register
Select a model type and device; training jobs, execution state and model versions can be tracked in ML management.
- 3
Evaluate and predict
Energy forecasting separates training and test data and measures RMSE and MAE before the model is used for inference.
- 4
Monitor and review
Review model health, latency, errors and drift; operations evaluates the result alongside actual equipment conditions.
Deploy against each site’s real data
For a refinery, petrochemical plant or other infrastructure, project engineering defines data sources, equipment mapping, history quality, permitted documents and acceptance criteria. The deployment guide describes OT/DMZ integration and local LLM or provider API options. In this industrial architecture, AI informs engineers and operators while process control remains in the existing control system.
13
Smart Assistant
A Persian-language assistant built on a large language model, with access to project data, that learns from the system itself and can act on the user's behalf — but never without confirmation. It is available on every page, and when the dashboard designer is open, it can see the canvas. It works with a primary language-model provider and a backup, and never accepts a tool's result from the model itself.
Figure 5 · The assistant loop. Indexed knowledge sources: the device catalog, SNMP and Modbus templates, common OIDs, the knowledge folder, and documents uploaded per project. Each row's vector embedding is stored so that switching models allows re-indexing.
Query and Diagnosis (41 tools)
- Project summary, equipment, points and instantaneous values, active alarms, and alarm statistics
- Detection of offline equipment, stale points, and anomalies; equipment health; overdue maintenance
- History of a point, energy consumption, meters, tickets, events, notifications
- Search across structure, dashboards, people and roles, and assets; reading and testing rules, limits, scenarios, scripts, and state machines; search of the knowledge base
Taking Action (44 tools, with a confirmation card)
- Creating a full structure, organization, building, floor, space, floor duplication, equipment (single or bulk), and points
- Creating/cloning/deleting dashboards, adding or designing 2D elements for a space, creating 3D scenes and objects
- Creating and editing rules, limits, scenarios, scripts, and state machines; enable/disable, execute, send event, reset; creating and assigning roles
- Every suggestion becomes a card, confirmable within 30 minutes; command-issuing logic is created "disabled" and needs a second confirmation to activate; writing to a point that is itself a condition is refused
Memory
Conversations are kept on the server and compressed into a rolling summary; the assistant remembers facts, aliases, preferences, and routines for the user or project (pinning, forgetting).
Knowledge Base (RAG)
The device catalog, SNMP/Modbus templates, common OIDs, the knowledge folder, and documents uploaded per project are indexed into 1,200-character chunks; hybrid keyword + semantic-vector search, with answers still available even when the vector service is unreachable.
14
ARG AI Backend
An independent machine-learning and analytics service for buildings and data centers that ingests twin data and produces forecasts, detection, optimization, and intelligent reports. It works with multiple language-model providers and can run entirely locally (offline, with a local language model). A full API with 145 documented operations.
33ready-made machine-learning models
145documented API operations
3compliance frameworks: ISO 27001 · SOC 2 · ASHRAE
LSTM · DQNdeep models via the Python service
Analytical Models
| Domain | Models and Outputs |
|---|---|
| Energy | Consumption forecasting (Random Forest, LSTM), demand-peak forecasting and load shifting, demand response (tariff analysis, load scheduling, cost optimization), consumption-pattern clustering (K-Means) and classification of new patterns, weather-energy model and pre-cooling/pre-heating driven by weather forecasts, carbon footprint |
| Detection and Diagnostics | Single-sensor anomaly detection (Isolation Forest + Z-Score) and multivariate (cross-sensor correlation), fault detection and diagnosis (multi-class), predictive maintenance (failure forecasting), equipment health curve and remaining useful life, data-drift monitoring |
| Comfort and Operation | Occupancy prediction for HVAC/lighting, thermal comfort (PMV/PPD), equipment schedule optimization (Q-Learning reinforcement learning + genetic algorithms, DQN), integrated cost/comfort optimization across systems |
| Building Systems | Lighting management (daylight, occupancy, circadian rhythm), elevators (traffic prediction, wait time), water (leaks, consumption forecasting, pressure anomalies), fire safety (risk, smoke/CO₂ patterns), access control (entry/exit patterns, intrusion detection) |
Training and Execution
- Training each model on a device's data; synchronous and asynchronous training jobs with status tracking and cancellation
- Parallel batch analysis, prediction cache with hit statistics, rate limiting per category
- Model health and automatic retraining; per-model prediction metrics and a monitoring dashboard
- Data import from CSV/JSON/bulk records with import history; sensor-data simulation for testing
Intelligent Services
- Chat assistant with multiple providers (automatic selection, connection testing, rate and buffer management), streaming responses, sessions
- AI-generated intelligent reports with PDF/Excel export; prompt templates with variables
- Alarm engine with rules, acknowledgment, and history; rule-based auto-remediation with an execution log; webhooks
- Multi-site support with a per-site dashboard and health overview; subsystem health map; BMS and data-center dashboards
- Digital twin simulation and "what-if" analysis; ISO 27001 / SOC 2 / ASHRAE compliance checks with history
- Mobile API (quick status, health summary, alarm feed, quick analysis); JWT token authentication
15
ARG Offline Map Server
Many industrial, military, and infrastructure sites have no internet access, or must not have any. The ARG Offline Map server keeps map tiles inside the customer's network and serves the system's map dashboards, site views, asset tracking, and fleet tracking — with no external dependency.
Capabilities
- Map region management: defining a region by selecting an area on the map or from presets (country, province, city)
- Downloading tiles for a region and zoom range with a tile-count estimate; download jobs that can be tracked and canceled
- Uploading ready-made MBTiles packages; displaying the geographic bounds and details of each region
- Standard tile serving (z/x/y) compatible with Leaflet and MapLibre; OSM, dark, and satellite base layers
- 3D globe view with terrain relief, atmosphere, and region boundaries; table and map views of regions
- Cache with hit-rate statistics, storage and free-space statistics, cache clearing, and health checks
- Documented API (18 operations: regions, downloads, tile service) with sample connection code from Leaflet
Integration with the System
- The offline map server's address is entered in the system settings, and maps are automatically fed from it
- 8 tile sources for online mode (OSM, Carto light/dark, satellite, topographic...) and an "offline" source for closed networks
- "Map"-type dashboards, MiniMap, drawing tools, and marking of equipment and sites on the map
- Suited to data centers, power substations, geographically dispersed sites, and fleets
See the Digital Twin live
The live ArgBMS demo with real 3D twins, dashboards and logic.