Manufacturing AI glossary: 25 terms every operations manager needs before signing off on a deployment
This glossary is written for the operations manager who has to sign off on the AI deployment, not for the ML engineer who designs the model. The vocabulary in industrial AI vendor literature crosses two different professions and three different decades of jargon. The definitions below are plant-manager-tier — one or two sentences each, with a link to the implementation detail when the term carries more weight than a definition can hold. Twenty-five terms cover the working surface area of an evaluation across AI vision (quality inspection) and AI safety (worker monitoring) deployments.
The terms are grouped by where they show up in an evaluation. A buyer running through the glossary in order will see the architecture choices in roughly the sequence they appear in a vendor conversation.
Detection-side terms
1. AI vision. Machine-learning-based inspection that classifies images of parts against a learned distribution of acceptable variation, replacing rule-based pixel-threshold logic. The architecture distinction against legacy AOI is covered in detail in the post on what the world's two largest machine vision companies could not solve.
2. Anomaly detection. A model trained on the distribution of "good" parts that flags anything outside the distribution as a candidate for review, without requiring labelled examples of every defect class. The architecture is the right fit on lines where defect examples are rare or where new defect types appear regularly; we covered the workflow in the post on AI vision for glass and flat panel display manufacturing.
3. Confusion matrix. A four-cell table of true positives, false positives, true negatives, and false negatives that an inspection model produces against a labelled test set. The cell values are how the buyer evaluates the model's behaviour on the specific defect distribution, not on the vendor's benchmark.
4. False reject rate (FRR). The fraction of acceptable parts flagged as defective by the inspection system. The operational target on most industrial lines is 0.5 percent; rates above the target produce operator bypass and silent system failure. The diagnostic detail is in the post on how to reduce false positives in AI quality inspection.
5. False positive (FP). A single flagged event where the part was acceptable. Used interchangeably with the unit-level signal that contributes to the FRR.
Quality-architecture terms
6. AQL (Acceptable Quality Level). The maximum percentage defective that a sampling plan is designed to accept at a defined probability. The framework was built for a world where 100 percent inspection was operationally impossible and remains required in regulated categories; the detail is in the post on what AQL is and why AI vision changes the calculation.
7. Cpk and Ppk. Process capability indices that measure how well a process fits inside its specification limits. IATF 16949 and most OEM PPAP submissions require Cpk on critical characteristics; an AI inspection deployment has to produce SPC-compatible outputs that feed these indices, not just pass-or-fail counts.
8. Gage R&R. Gauge repeatability and reproducibility — a measurement-system-analysis test that quantifies the variation a measurement introduces. AI inspection systems are measurement systems and have to pass Gage R&R against the same standards as any other inspection method to survive an IATF audit.
9. Predictive quality. The use of inspection data to identify upstream process issues before they generate the next batch of defects, by reading the same dataset for spatial clustering, morphological drift, and SPC-density signals. The architectural detail is in the post on predictive quality and how AI vision detects process drift before it becomes defects.
10. Autonomous quality control. A closed-loop architecture where the AI infers the corrective adjustment to an upstream process variable and the PLC executes the adjustment inside pre-validated bounds. The architectural distinction between Level 4 (predictive — humans act) and Level 5 (autonomous — PLC executes) is in the post on what autonomous quality control is.
Safety-architecture terms
11. AI safety monitoring. Real-time worker monitoring using CCTV-fed vision models and biometric wearables to detect precursors to incidents in the seconds before harm occurs. The architecture is covered in the post on what HyperQ AI Safety is and the moment-before window the system is built for.
12. Active detection. Monitoring that identifies the precursor state to an incident in real time and fires an alert before the incident occurs, as distinct from passive recording that captures the incident after the fact. The regulatory shift from passive to active across APAC jurisdictions is the subject of the APAC safety compliance checklist for Singapore, Malaysia, and South Korea.
13. Reasonably practicable. The legal standard under Singapore's Workplace Safety and Health Act and equivalent APAC statutes for the controls a dutyholder is required to take. The 2024 MOM enforcement update reinterpreted the test in operational terms, covered in the post on the camera that records is not the camera that complies.
14. Dutyholder. A named individual or organisation with statutory responsibility for safety controls. Under Malaysia's CDM 2024, five dutyholder roles carry distinct accountabilities; the framework is in the post on CDM 2024 and the audit-trail requirement that follows.
15. Audit trail. A timestamped record of the safety controls operating at each moment, with inspection result, classification confidence, model version, and worker response logged per event. The regulator's documentary base; the cross-site discipline is in the post on the multi-site EHS blind spot most directors live with.
Deployment-architecture terms
16. Edge inference. Running the AI model on a device next to the camera on the production line, rather than over a cloud round-trip. The latency, reliability, and data-sovereignty arguments for edge in manufacturing AI are in the foundational post on what edge inference is and why it matters for manufacturing AI.
17. ONVIF. A standard protocol for IP-based CCTV cameras that lets a safety or vision system pick up existing camera installations through auto-recognition rather than requiring camera replacement. The capability is what reduces the deployment time to one hour on most existing camera infrastructures.
18. Visual Language Model (VLM) with PEFT. A model architecture that combines visual and language reasoning, fine-tuned with Parameter-Efficient Fine-Tuning on industrial safety-specific data. The architecture underlying HyperQ AI Safety's detection across the four primary categories (fall, fire, intrusion, PPE non-compliance).
19. Concept drift. The gradual movement of production conditions outside the model's training distribution, producing a slow degradation in inspection accuracy that is distinct from hardware-side issues like thermal throttling. The diagnostic and the retraining response are in the post on continuous learning for edge-deployed AI vision.
20. Thermal throttling. The performance degradation that occurs when an edge inference accelerator runs sustained load and its thermal management starts limiting clock frequency to stay inside its temperature envelope. A hardware problem, not a model problem; the mitigations are in the foundational edge inference post.
Integration and operations terms
21. MES integration. The data path from the inspection or safety layer into the Manufacturing Execution System, where each event is linked to the production lot, the product change, the model version, and the conditions at the moment of inspection. The practical patterns are in the post on how to integrate AI vision with MES or ERP systems.
22. InspectWindow. The PLC-side timer that starts at the camera trigger and runs for the duration the line allows before the rejector decides on the part. Inference latency above the InspectWindow auto-rejects good parts as scrap; the mechanism is in the edge inference post.
23. OSAT inspection. Outsourced Semiconductor Assembly and Test — back-end packaging, wire bond, die attach, BGA, underfill, and surface inspection that sits downstream of wafer fabrication. The defect taxonomy and the high-product-mix problem differ from wafer-fab AOI; the analysis is in the post on AI quality inspection for Malaysia semiconductor manufacturing.
24. Hazardous-zone classification. The categorisation of areas in a manufacturing facility by the risk of fire, explosion, or other hazard, used to specify the safety controls (PPE, monitoring, access) required for each zone. European facilities use ATEX zone classifications; APAC facilities operate under local equivalents such as IECEx and national hazardous-area standards. The detection-side requirement is the same regardless of the labelling scheme.
25. Asymmetric commitment. The buyer-side discipline that produces a defensible deployment decision: send a representative sample of data, receive a written analysis with confusion matrices and a deployment plan in two weeks, with no contract obligation until the system has been measured against the customer's actual operating conditions. The framing is consistent across the AI vision and AI safety deployments we have written about; the buyer's-guide-level treatment is in the post on how to choose an AI vision system for manufacturing operations.
How to use this glossary in an evaluation
The glossary is sequenced to match the order in which the terms appear in a typical evaluation. A buyer in an awareness-stage conversation will be working through terms 1 to 5. A buyer at the architecture-decision stage will be working through 6 to 15. A buyer at the deployment-planning stage will be working through 16 to 25. The links to the implementation posts behind each term let the buyer dig deeper on any term where the definition is not sufficient.
If a term in this glossary describes a capability that the buyer is evaluating, the linked post has the implementation detail and the buyer-side discipline that decides whether the deployment makes sense. If the buyer is past the research stage and into selection, the next step is the asymmetric commitment described in term 25.