Industrial AI
Industrial AI that knows your plant
General-purpose AI has read the internet. It has never seen your pump station, and it cannot read the manual in your filing cabinet. TDengine closes that gap with an asset model, an indexed knowledge base built from your own documents, and analysis that runs where the data already lives.
How does TDengine give AI context about your plant?
An AI assistant can only answer useful questions about a plant if something tells it what the plant is. In TDengine that context comes from TDengine IDMP, which holds an asset model of your site and lets you attach the documents that describe each asset — equipment manuals, standard operating procedures, P&ID drawings, calibration records and past failure reports. IDMP indexes those documents so they can be searched by asset, and associates what it finds with the assets themselves — links we verify during the load rather than take on trust.
The analysis is then split between two engines. TDgpt is TDengine’s time-series AI engine — installed as a module and invoked from SQL for forecasting, anomaly detection and imputation, with that work staying in the database. It does not need a language model. A language model of your choosing, which can run entirely inside your own network, handles chat, panel generation, multi-step reasoning and the written report, including root cause analysis.
Logic Control & Telemetry (Pty) Ltd is a TDengine System Integrator (SI). The platform capability above is TDengine’s; the engineering that makes it work on a specific South African plant — modelling the assets, encoding how equipment affects equipment, and getting your document set in — is ours.
The stack
Four layers, and the order matters
Most disappointing industrial AI pilots have the first and last layers and nothing in between. Context and knowledge are what separate an assistant that reasons from one that guesses.
1. Data
TDengine TSDB
Stores the signals
High-ingest time-series storage for plant and telemetry tags, with long retention and SQL access. Nothing here understands what a tag means — it stores values against timestamps.
- High-throughput tag ingest from PLC, SCADA and remote telemetry
- Long retention without punitive storage growth
- Standard SQL for engineering and application teams
2. Context
TDengine IDMP — asset model
Explains what the signals mean
An asset tree plus a relationship network. The tree tells an AI agent what a thing is and where it sits. The relationship network tells it what that thing feeds, controls and affects — which is what makes causal reasoning possible instead of correlation guessing.
- Asset hierarchy from site down to measurement point
- Engineering units, limits, target values and classifications per attribute
- Typed relationships across the tree, not just parent-child
- One asset can appear in several trees without duplicating data
3. Knowledge
TDengine IDMP — knowledge base
Supplies the plant-specific experience
Manuals, SOPs, P&IDs, calibration records and past failure reports attached directly to the asset they describe. IDMP parses and indexes them so AI Chat and investigations can search your manuals, SOPs and drawings by asset. Extracted terms are associated with the model where they match; mismatches are corrected during load, not assumed correct on upload.
- Documents attach to individual elements or to element templates
- Word, PDF, Markdown, drawings and site photographs are parsed on upload
- Equipment models, fault types and operating steps are extracted and linked
- Plant knowledge is searched first; general model knowledge is the fallback
4. Intelligence
TDgpt + your chosen LLM
Does the analysis and the reasoning
TDgpt is TDengine’s time-series AI engine. It is installed as a separate module alongside TSDB and invoked from SQL so forecasting, anomaly detection and imputation run against data that stays in the database. A separate large language model handles language, multi-step reasoning and report writing — including root cause analysis.
- TDgpt: forecasting, anomaly detection and imputation, invoked as SQL — no LLM required
- LLM: natural-language questions, panel generation and written reports
- The LLM is external and swappable — including a local OpenAI-compatible endpoint
- TDgpt analysis stays in the database; chat, RCA and panel generation need the LLM connection
Division of labour
TDgpt does the maths, not the reading
TDgpt analyses signals. It never reads your manuals — the language model does that, through the knowledge base. Which engine handles what determines the groundwork you need in place before either is useful.
Forecasting
Projects a tag forward over a chosen number of rows and returns a confidence band alongside the prediction. Useful for demand, level and consumption planning.
SELECT _frowts, FORECAST(discharge_pressure, "algo=holtwinters,rows=48") FROM pump_101;Anomaly detection
Opens and closes a window when the algorithm judges behaviour abnormal, rather than when a threshold you guessed is crossed. This tracks a shifting baseline as operating conditions change.
SELECT COUNT(*) FROM pump_101 ANOMALY_WINDOW(vibration, "algo=iqr");Gap filling
Fills missing values in a series using TDengine’s IMPUTATION function (Moment model). Relevant where remote sites drop out and reporting needs a continuous record. The time-series foundation model must be deployed on the TDgpt module before IMPUTATION can use it.
SELECT _improwts, IMPUTATION(level, "algo=moment,freq=s") FROM reservoir_12;TDgpt ships statistical algorithms, machine learning and deep learning libraries, and adapters for time-series foundation models. It runs as a separate module alongside TSDB. Custom algorithms can be added in Python and then called from SQL without changing application code.
Knowledge
What you load onto an asset
This is the part most operators already own and have never been able to use. It is sitting on shared drives, in filing cabinets, and in the heads of people approaching retirement.
Equipment manuals
Vendor manuals and engineering handbooks. IDMP indexes them for AI Chat, so questions about expected operating range, service intervals and alarm setpoints come from the document rather than guessed.
Standard operating procedures
The procedure for that equipment, attached to the element or its template. It lives on the asset (and can sit on the operator panel) instead of a departmental shared drive.
P&IDs and drawings
Process and instrumentation diagrams attached at the right level of the hierarchy. They are indexed with the asset, so engineers and AI Chat use the same drawing.
Calibration and maintenance records
Calibration reports attach as related documents. Notes on the element can hold maintenance and incident history — often the first place to look when a measurement starts to drift.
Failure case history
Incident reports and post-mortems attach like any other document, so they are searchable with the asset. Turning them into a structured case library — symptom, cause, remedy — is extra work TDengine supports; it is not automatic on upload.
Site photographs
Site photos upload with Word, PDF, Markdown and process drawings. The system parses and indexes them with the rest of the knowledge base.
Data sovereignty
Your knowledge does not become someone else’s training data
For most South African mines and utilities this is the question that decides the project. The boundary holds in four places.
No vendor lock to a model
TDengine does not bundle a language model. You choose a hosted provider, a privately deployed model, or one running entirely inside the plant. Any OpenAI-compatible endpoint works, including a local Ollama or vLLM service with no API key.
Knowledge stays yours
Your asset model, documents, rules and skills are your assets, held in your system. With a local model, plant knowledge never crosses your network boundary — which is what makes this viable for operators who cannot send process data to a public API.
The agent inherits permissions
Access control applies to AI features as well as people. Elements a user cannot access are hidden in the asset tree — including from the agent acting for them — so an assistant is not a way around authorisation.
Every change can be audited
IDMP’s audit trail (off until an administrator enables it) logs who changed which object, when, and from where, in a tamper-proof record. TDengine also describes tracing skill calls and AI operations; we confirm the exact log scope on the build you run.
Go deeper
More detail on the parts that decide whether this works.
AI root cause analysis
The eight-step investigation workflow, what must be in place first, and how to validate the output before trusting it.
Learn moreAsset context and knowledge
Building the asset model, encoding equipment relationships, and loading manuals and SOPs so an agent can use them.
Learn moreTDengine Historian
How TSDB and IDMP fit together in an OT architecture, and where LCT plugs into your control systems.
Learn moreFrequently asked questions
Can AI find the root cause of a plant fault?
It can produce a ranked, evidence-tested set of hypotheses and the report behind them, provided it has context. On its own a language model has never seen your plant and cannot do this. TDengine IDMP supplies the missing context through an asset model, typed relationships between equipment, and your own manuals, SOPs and failure history attached to the assets they describe. The engineer still makes the final judgement.
Can LCT load equipment manuals and SOPs into TDengine?
Yes. TDengine IDMP allows documents — manuals, SOPs, P&ID drawings, calibration reports, incident reports and site photographs — to be attached directly to an asset or to an element template. IDMP parses and indexes them for search by asset. Equipment names and fault types found in a document are associated with matching objects in the asset model during load — we verify those links, they are not guaranteed correct without review. The AI then answers questions about that asset from your documents first.
What is the difference between TDgpt and the LLM?
TDgpt is TDengine’s time-series AI engine. It is installed as a module and does the numerical work — forecasting, anomaly detection and imputation — invoked from SQL, with that analysis staying in the database. It does not read documents and it does not require an LLM. A separate large language model, connected over an OpenAI-compatible interface, handles natural-language questions, panel generation, multi-step reasoning and report writing. Root cause analysis uses the deep-thinking LLM, not TDgpt.
Does our process data have to leave site for AI to work?
No, provided the language model is hosted on your side. TDgpt analysis never leaves the database — it runs inside TDengine and does not need an LLM at all. Chat, panel generation and root cause analysis do send retrieved time-series windows and document excerpts to whichever OpenAI-compatible endpoint is configured. A local service keeps that traffic on your own network; a hosted provider means those excerpts leave site under your agreement with that provider.
How much of this is TDengine and how much is LCT?
The platform capability is TDengine’s. The work that makes it useful on a specific plant is ours: modelling the asset hierarchy, encoding process and fault-impact relationships, gathering and attaching the document set, configuring the model connection and permissions, writing skills for your procedures, and validating results against incidents your team already understands.
Can TDengine replace our asset model and operator screens?
IDMP is TDengine’s asset and visualisation layer — hierarchy, templates, attributes and operator panels in the same product as the historian. On top of that you attach manuals, SOPs and P&IDs so an assistant can answer from your documents, with a language model that can run on-premises. That covers a first historian, a SCADA archive that was never built to hold long history, and replacing PI Asset Framework plus PI Vision. Where PI is involved, the asset model is rebuilt from your plant rather than imported from AF — a considered model of the equipment that matters, not a dump of every PI interface.
Talk about what an AI-ready asset model looks like on your plant
A scoping conversation covers your tag list, what documentation exists, and whether a bounded proof of concept makes sense before committing to a full model.