Deploying Trustworthy AI for Industrial Maintenance

The article “ [IA 3 020] ” lays the groundwork. It presents three AI approaches for industrial maintenance (data-driven, document-driven, hybrid), a hyperspecialized agent-based architecture, and business-oriented Key Performance Indicators (KPIs): source fidelity, recall of the first k results (R@k), 95th percentile latency (P95), and mean time to repair (MTTR). It explains why to specialize, what to measure, and when to refrain from doing so.

That leaves the most practical question: how do you implement this AI, from the initial scoping phase through to maintaining it in operational conditions?

That is the focus of this article. It describes a methodology structured into five operational steps, marked by decision gates, and grounded in recognized frameworks, particularly the Body of Knowledge (BoK) of the Confiance.ai program and the requirements of European Regulation 2024/1689 on artificial intelligence (AI Act). These frameworks are used as references for consistency and evidence requirements, not as restrictive constraints: the Omundu methodology retains its practical focus and its primary goal—a return on investment (ROI) that is measurable on the shop floor.

The proposed method applies to AI used to assist with industrial maintenance, positioned on a separate chain from any Safety Instrumented Function (SIF) and any associated Safety Integrity Level (SIL). It is based on four assumptions:

– an existing body of documentation (manuals, drawings, procedures, frequently asked questions (FAQs)), even if imperfect;

– an identified maintenance team with diverse skill sets;

– read-only access to the industrial network via standard protocols: OPC-UA (Open Platform Communications-Unified Architecture), SCADA (Supervisory Control and Data Acquisition), or MES (Manufacturing Execution System);

– on-premises or edge hosting, with no data flowing outside the network.

To ground each step in operational reality, this article draws on an illustrative scenario representative of a typical deployment on an assembly line, “Line A”:

– site: manufacturing SME, 120 employees, Auvergne-Rhône-Alpes region;

– scope: multi-equipment assembly line, with three hydraulic presses, two welding robots, a chain...